EP1797552B1 - Method and device for the extraction of a melody on which an audio signal is based - Google Patents

Abstract

The finding of the present invention is that the melody extraction or automatic transcription may be implemented clearly more stable and if applicable even less expensive when the assumption is considered sufficiently that the main melody is the portion of a piece of music which man perceives the loudest and the most precise. Regarding this, according to the present invention the time/spectral representation or the spectrogram of an interesting audio signal is scaled using the curves of equal volume reflecting human volume perception in order to determine the melody of the audio signal on the basis of the resulting perception-related time/spectral representation.

Description

The present invention relates to the extraction of a melody underlying an audio signal. Such extraction may be used, for example, to obtain a transcribed representation of a melody underlying a monophonic or polyphonic audio signal, which may also be in an analog form or in a digital sampled form. Thus, melody extractions, for example, enable the generation of ringtones for mobile phones from any audio signal, e.g. Singing, humming, whistling or the like.

Already for some years, beeps of mobile phones are no longer just the signaling of a call alone. Rather, they have become an entertainment factor with growing melodic capabilities of mobile devices and a status symbol among teenagers.

Earlier mobile phones offered in part the possibility to compose monophone ringtones on the device itself. However, this was complicated and often frustrating for musically under-educated users and unsatisfactory in their results. Therefore, this possibility or functionality has largely disappeared from newer phones.

In particular, modern phones that allow polyphonic signaling melodies or ringtones, offer such a wealth of combinations that an independent composition of a melody on such a mobile device is hardly possible. At most, ready-made melody and accompaniment patterns can be recombined to provide a limited amount of independent ringtones.

Such combinability of ready-made melody and accompaniment patterns is implemented, for example, in the Sony-Ericsson T610 phone. In addition, however, the user relies on the purchase of commercially available pre-made ringtones.

It would be desirable to be able to provide the user with an intuitively operable interface for creating his own signaling melody, which does not require a large musical education, but is nevertheless suitable for the implementation of their own polyphonic melodies.

In most keyboards today, there is a so-called automatic accompaniment called functionality to automatically accompany a melody given the chords to be used. Quite apart from the fact that such keyboards provide no way to use an interface to a computer to transfer the provided with a tune melody to a computer and there to convert into a suitable mobile phone format to use them as ringtones in a mobile phone can , the use of a keyboard to generate its own polyphonic signaling melodies for mobile phones for most users, because they are not able to operate this musical instrument.

In the DE 102004010878.1 entitled "Apparatus and Method for Providing a Signaling Tune" whose assignee is the same as the assignee of the present application and which was filed with the German Patent and Trademark Office on Mar. 5, 2004, a method is described with which Use a Java applet and server software to generate monophonic and polyphonic ringtones and send them to a mobile device. However, the procedures proposed there for extracting the melody from audio signals are very error-prone or have only limited applicability. Among other things, it is proposed there to thereby arrive at a melody of an audio signal that characteristic features are extracted from the audio signal to compare the same with corresponding features of pre-stored tunes, and then select as the generated tune that among the pre-stored ones giving the best match. However, this approach inherently restricts the melody recognition to the pre-stored set of tunes.

The DE 102004033867.1 entitled "Method and Apparatus for the Rhythmic Conditioning of Audio Signals" and the DE 102004033829.9 entitled "Method and Apparatus for Producing a Polyphonic Melody" filed with the German Patent and Trademark Office on the same day also deal with the generation of melodies from audio signals, but do not elaborate on the actual melody recognition, but rather the ensuing process of deriving an accompaniment from the melody along with a rhythmic and harmonic-dependent treatment of the melody.

With possibilities of melody recognition deals for example Bello, JP, Towards the Automated Analysis of Simple Polyphonic Music: A Knowledge-based Approach, University of London, Diss., January 2003 , various types of start timing of notes are described, based either on the local energy in the time signal or on an analysis in the frequency domain. In addition, various methods for melody line detection are described. The common feature of these approaches is that they are complicated in that the final melody is obtained indirectly by first processing or tracking several trajectories in the time / spectral representation of the audio signal, and finally, under these trajectories Selection of the melody line or the melody is made.

Also in Martin, KD, A Blackboard System for Automatic Transcription of Simple Polyphonic Music, MIT Media Laboratory Perceptual Computing Section 385, 1996 , a possibility for automatic transcription is described, which is also based on the evaluation of several harmonic tracks in a time / frequency representation of the audio signal or the spectrogram of the audio signal.

M. Karjalainen and Tero Tolonen: "Multi-Pitch and Periodicity Analysis Model for Sound Separation and Auditory Scene Analysis" Proc. IEEE Int. Conf. Acoustics, speech, and signal processing (ICASSP'99), vol. 2, pp. 929-932, Phoenix, Arizona, March 15-19, 1999 shows a system for detecting sounds, wherein the input signal is processed by a filter, which is related to the equal-volume curves (equal-loudness sensitivity filtering). The processed signal is then analyzed in the time domain by autocorrelation to estimate the period of the tones (See D1, Figure 2 ).

G. Monti, M. Sandler: "Automatic Polyphonic Piano Note Extraction Using Fuzzy Logic in a Blackboard System", Proc. of the 5th Int. Conference on Digital Audio Effects (DAFx-02) Hamburg, Germany, September 26-28,2002 , shows how thresholds for masking ("masking thresholds") are calculated from the curves of equal volume (see equations (3) - (8) and illustration 1 ). The signal is then compared to these thresholds to identify the audible components underlying the signal.

In Klapuri, AP: Signal Processing Methods for the Automatic Transcription of Music, Tampere University of Technology, Diss., December 2003 , and Klapuri, AP, Signal Processing Methods for the Automatic Transcription of Music, Tampere University of Technology, Diss., December 2003 . AP Klapuri, "Number Theoretical Means of Resolving a Mixture of Several Harmonic Sounds". In Proceedings European Signal Processing Conference, Rhodes, Greece, 1998 . AP Klapuri, "Sound Onset Detection by Applying Psychoacoustic Knowledge," in Proceedings IEEE International Conference on Acoustics, Speech, and Signal Processing, Phoenix, Arizona, 1999 . AP Klapuri, "Multipitch Estimation and Sound Separation by the Spectral Smoothness Principle," in Proceedings IEEE International Conference on Acoustics, Speech, and Signal Processing, Salt Lake City, Utah, 2001 . Klapuri AP and Astola JT, "Efficient Calculation of a Physiologically-Motivated Representation for Sound", in Proceedings 14th IEEE International Conference on Digital Signal Processing, Santorini, Greece, 2002 . AP Klapuri, "Multiple Fundamental Frequency Estimation based on Harmonicity and Spectral Smoothness", IEEE Trans. Speech and Audio Proc., 11 (6), pp. 804-816, 2003 . Klapuri AP, Eronen AJ and Astola JT, "Automatic Estimation of the Meter of Acoustic Musical Signals", Tempere University of Technology, Institute of Signal Processing, Report 1-2004, Tampere, Finland, 2004, ISSN: 1459-4595, ISBN: 952-15-1149-4 , various methods are described around the automatic transcription of music.

In the context of basic research on the subject extraction of a main melody as a special case of polyphonic transcription is further Baumann, U .: A Method for the Detection and Separation of Multiple Acoustic Objects, Diss., Chair of Human-Machine Communication, Technische Universität München, 1995 to emphasize.

The above-mentioned different approaches to melody recognition or automatic transcription usually make special demands on the input signal. For example, they only allow piano music or only a certain number of instruments or exclude percussive instruments or the like:

The most practicable approach to current modern and popular music is the approach of Goto, as it is, for example, in Goto, M .: A Robust Predominant-FO Estimation Method for Real-time Detection of Melody and Bass Lines in CD Recordings, Proc. IEEE International Conference on Acoustics, Speech and Signal Processing, pp.II-757-760, June 2000 , is described. The aim of this method is to extract a dominant melody and bass line, with the detour to line finding again taking place by selecting among several trajectories, namely using so-called "agents". The process is so expensive.

Also deals with melody detection Paiva RP et al: A Methodology for Detection of Melody at Polyphonic Musical Signals, 116th AES Convention, Berlin, May 2004 , There, too, it is proposed to follow the path of trajectory tracking in the time / spectral representation. The document also deals with the segmentation of the individual trajectories, until they are processed into a sequence of notes.

It would be desirable to have a method of melody extraction or automatic transcription that is more robust and reliable for a wider variety of different audio signals. Such a robust System could provide a high time and cost savings in "query by huming" systems, ie in systems where a user is able to find songs by pre-sums in a database, as an automatic transcription for the system database reference files it is possible. Of course, a robustly functioning transcription could also find use as a recording frontend. Furthermore, it would be possible to use automatic transcription as a supplement to an audio ID system, ie a system that recognizes audio files on a fingerprint contained in them, as in the case of non-recognition by the audio ID system, such as due to a missing fingerprint, the automatic transcription could alternatively be used to evaluate an incoming audio file.

Stably functioning automatic transcription would also allow for the production of similarity relationships associated with other musical features, e.g. Key, harmony and rhythm, such as for a "recomandation engine" or "suggestion engine". In musicology, a stable automatic transcription could create new views and lead to a re-examination of judgments on older music. Also, to preserve copyright by objectively comparing pieces of music, automatic transcription that is stable in use could be used.

In summary, the application of melody recognition or auto-transcription is not limited to the generation of ringtones for mobile phones mentioned above, but can generally serve as a support for musicians and those interested in music.

Therefore, the object of the present invention is to provide a more stable melody recognition scheme which works correctly for a wider variety of audio signals.

This object is achieved by a device according to claim 1, a method according to claim 33 and computer program according to claim 34.

The recognition of the present invention is that the melody extraction or automatic transcription can be made much more stable and possibly even unsuitable if the assumption is sufficiently taken into account that the main melody is that part of a piece of music that the person perceives loudest and most concise. Taking this into account, according to the present invention, the time-over-spectral representation or spectrogram of an audio signal of interest is scaled using the equal volume curves that reflect human loudness perception to determine the melody based on the resulting perceptual time-over-spectral representation of the audio signal.

According to a preferred embodiment of the present invention, the above musicological statement that the main melody is that portion of a piece of music that the person perceives most loudly and succinctly is taken into account in two ways. Namely, according to this embodiment, when determining the melody of the audio signal, first a melody line extending through the time-over-spectral representation is determined, namely by assigning to each time period or frame - uniquely - exactly one spectral component Frequency bin is assigned to the time-over-spectral representation, namely that which leads to the sound result with the maximum intensity. More specifically, according to this embodiment, the spectrogram of the audio signal is first logarithmized, so that the logarithmized spectral values indicate the sound pressure level. Subsequently, the logarithmic spectral values of the logarithmized spectrogram become dependent on their respective value and the spectral component to which they belong perceptually related spectral values. Functions are used that represent the curves of equal volume as sound pressure as a function of spectral components or as a function of the frequency and are assigned to different volumes.

Fig. 1

ein Blockschaltbild einer Vorrichtung zur Erzeugung einer polyphonen Melodie;

Fig. 2

ein Flussdiagramm zur Veranschaulichung der Funktionsweise der Extraktionseinrichtung der Vorrichtung von Fig. 1;

Fig. 3

ein detaillierteres Flussdiagramm zur Veranschaulichung der Funktionsweise der Extraktionseinrichtung der Vorrichtung von Fig. 1 für den Fall eines polyphonen Audioeingangssignals;

Fig. 4

ein exemplarisches Beispiel für eine Zeit/Spektraldarstellung bzw. ein Spektrogramm eines Audiosignals, wie es bei der Frequenzanalyse in Fig. 3 entstehen könnte;

Fig. 5

ein logarithmiertes Spektrogramm, wie es sich nach der Logarithmierung in Fig. 3 ergibt;

Fig. 6

ein Diagramm mit den Kurven gleicher Lautstärke, wie sie der Bewertung des Spektrums in Fig. 3 zu Grunde liegen;

Fig. 7

einen Graphen eines Audiosignals, wie es vor der eigentlichen Logarithmierung in Fig. 3 verwendet wird, um einen Bezugswert für die Logarithmierung zu erhalten;

Fig. 8

ein wahrnehmungsbezogenes Spektrogramm, wie es nach der Bewertung des Spektrogramms von Fig. 5 in Fig. 3 erhalten wird;

Fig. 9

die sich aus dem wahrnehmungsbezogenen Spektrum von Fig. 8 durch die Melodielinienermittlung von Fig. 3 ergebende Melodielinie bzw. -funktion eingezeichnet in der Zeit-/Spektraldomäne;

Fig. 10

ein Flussdiagramm zur Veranschaulichung der allgemeinen Segmentierung von Fig. 3;

Fig. 11

eine schematische Darstellung eines exemplarischen Melodielinienverlaufs in der Zeit-/Spektraldomäne;

Fig. 12

eine schematische Darstellung eines Ausschnitts aus der Melodielinienverlaufsdarstellung von Fig. 11 zur Veranschaulichung der Wirkweise der Filterung in der allgemeinen Segmentierung von Fig. 10;

Fig. 13

der Melodielinienverlauf von Fig. 9 nach der Frequenzbereichseingrenzung in der allgemeinen Segmentierung von Fig. 10;

Fig. 14

eine schematische Zeichnung, in der ein Ausschnitt aus einer Melodielinie gezeigt ist, zur Veranschaulichung der Wirkweise des vorletzten Schritts in der allgemeinen Segmentierung von Fig. 10;

Fig. 15

eine schematische Zeichnung eines Ausschnitts aus einer Melodienlinie zur Veranschaulichung der Wirkweise der Segmenteinteilung in der allgemeinen Segmentierung von Fig. 10;

Fig. 16

ein Flussdiagramm zur Veranschaulichung der Lückenschließung in Fig. 3;

Fig. 17

eine schematische Zeichnung zur Veranschaulichung der Vorgehensweise beim Setzen des variablen Halbtonvektors in Fig. 3;

Fig. 18

eine schematische Zeichnung zur Veranschaulichung der Lückenschließung nach Fig. 16;

Fig. 19

ein Flussdiagramm zur Veranschaulichung des Harmoniemappings bzw. der Harmonieabbildung in Fig. 3;

Fig. 20

eine schematische Darstellung eines Ausschnitts aus dem Melodielinienverlauf zur Veranschaulichung der Wirkweise des Harmoniemappings nach Fig. 19;

Fig. 21

ein Flussdiagramm zur Veranschaulichung der Vibratorerkennung und des Vibratorausgleichs in Fig. 3;

Fig. 22

eine schematische Darstellung eines Segmentverlaufs zur Veranschaulichung der Vorgehensweise nach Fig. 21;

Fig. 23

eine schematische Darstellung eines Ausschnitts aus dem Melodielinienverlauf zur Veranschaulichung der Vorgehensweise bei der statistischen Korrektur in Fig. 3;

Fig. 24

ein Flussdiagramm zur Veranschaulichung der Vorgehensweise bei der Onset-Erkennung und - Korrektur in Fig. 3;

Fig. 25

einen Graphen, der eine exemplarische Filterübertragungsfunktion zur Verwendung bei der Onset-Erkennung nach Fig. 24 zeigt;

Fig. 26

einen schematischen Verlauf eines zweiwegegleichgerichteten gefilterten Audiosignals sowie der Hüllkurve desselben, wie sie zur Onset-Erkennung und -Korrektur in Fig. 24 verwendet werden;

Fig. 27

ein Flussdiagramm zur Veranschaulichung der Funktionsweise der Extraktionseinrichtung aus Fig. 1 für den Fall monophoner Audioeingangssignale;

Fig. 28

ein Flussdiagramm zur Veranschaulichung der Tontrennung in Fig. 27;

Fig. 29

eine schematische Darstellung eines Ausschnitts aus dem Amplitudenverlauf des Spektrogramms eines Audiosignals entlang eines Segments zur Veranschaulichung der Funktionsweise der Tontrennung nach Fig. 28;

Fig. 30a

und b schematische Darstellungen eines Ausschnitts aus dem Amplitudenverlauf des Spektrogramms eines Audiosignals entlang eines Segments zur Veranschaulichung der Funktionsweise der Tontrennung nach Fig. 28;

Fig. 31

ein Flussdiagramm zur Veranschaulichung der Tonglättung in Fig. 27;

Fig. 32

eine schematische Darstellung eines Segments aus dem Melodielinienverlauf zur Veranschaulichung der Vorgehensweise der Tonglättung nach Fig. 31;

Fig. 33

ein Flussdiagramm zur Veranschaulichung der Offset-Erkennung und -Korrektur in Fig. 27;

Fig. 34

eine schematische Darstellung eines Ausschnitts aus einem zweiwegegleichgerichteten gefilterten Audiosignals und dessen Interpolation zur Veranschaulichung der Vorgehensweise nach Fig. 33; und

Fig. 35

ein Ausschnitt aus einem zweiwegegleichgerichteten gefilterten Audiosignals und dessen Interpolation für den Fall einer potentiellen Segmentverlängerung.

Preferred embodiments of the present invention will be explained in more detail below with reference to the accompanying drawings. Show it:

Fig. 1

a block diagram of an apparatus for generating a polyphonic melody;

Fig. 2

a flowchart illustrating the operation of the extraction device of the device of Fig. 1 ;

Fig. 3

a more detailed flow chart for illustrating the operation of the extraction device of the device of Fig. 1 in the case of a polyphonic audio input signal;

Fig. 4

an exemplary example of a time / spectral representation or a spectrogram of an audio signal, as in the frequency analysis in Fig. 3 could arise;

Fig. 5

a logarithmic spectrogram, as it appears after logarithmation in Fig. 3 results;

Fig. 6

a graph with the curves of equal volume, as the evaluation of the spectrum in Fig. 3 underlie;

Fig. 7

a graph of an audio signal, as it was before the actual logarithmization in Fig. 3 is used to obtain a reference value for the logarithm;

Fig. 8

a perceptual spectrogram, as shown by the evaluation of the spectrogram of Fig. 5 in Fig. 3 is obtained;

Fig. 9

arising from the perceptual spectrum of Fig. 8 through the melody line detection of Fig. 3 resulting melody line or function plotted in the time / spectral domain;

Fig. 10

a flowchart illustrating the general segmentation of Fig. 3 ;

Fig. 11

a schematic representation of an exemplary melody line curve in the time / spectral domain;

Fig. 12

a schematic representation of a section of the melody line progression of Fig. 11 to illustrate the mode of action of filtering in the general segmentation of Fig. 10 ;

Fig. 13

the melody line course of Fig. 9 after frequency domain confinement in the general segmentation of Fig. 10 ;

Fig. 14

a schematic drawing, in which a section of a melody line is shown, to illustrate the operation of the penultimate step in the general segmentation of Fig. 10 ;

Fig. 15

a schematic drawing of a section of a melody line to illustrate the mode of operation of the segmentation in the general segmentation of Fig. 10 ;

Fig. 16

a flow chart illustrating the gap closure in Fig. 3 ;

Fig. 17

a schematic drawing illustrating the procedure when setting the variable halftone vector in Fig. 3 ;

Fig. 18

a schematic drawing illustrating the gap closure after Fig. 16 ;

Fig. 19

a flowchart for illustrating the Harmoniemappings or the harmony mapping in Fig. 3 ;

Fig. 20

a schematic representation of a section of the melody line course to illustrate the mode of action of Harmoniemappings after Fig. 19 ;

Fig. 21

a flowchart illustrating the vibrator detection and the vibrator compensation in Fig. 3 ;

Fig. 22

a schematic representation of a segment profile to illustrate the procedure according to Fig. 21 ;

Fig. 23

a schematic representation of a section of the melody line course to illustrate the procedure in the statistical correction in Fig. 3 ;

Fig. 24

a flow chart illustrating the procedure for onset detection and correction in Fig. 3 ;

Fig. 25

a graph depicting an exemplary filter transfer function for use in onset detection Fig. 24 shows;

Fig. 26

a schematic course of a two-way rectified filtered audio signal and the envelope thereof, as they are for onset detection and correction in Fig. 24 be used;

Fig. 27

a flowchart for illustrating the operation of the extraction device Fig. 1 in the case of monophonic audio input signals;

Fig. 28

a flow chart illustrating the sound separation in Fig. 27 ;

Fig. 29

a schematic representation of a portion of the amplitude profile of the spectrogram of an audio signal along a segment to illustrate the operation of the sound separation according to Fig. 28 ;

Fig. 30a

and b are schematic representations of a portion of the amplitude profile of the spectrogram of an audio signal along a segment to illustrate the operation of the sound separation according to FIG Fig. 28 ;

Fig. 31

a flow chart illustrating the sound smoothing in Fig. 27 ;

Fig. 32

a schematic representation of a segment from the melody line progression to illustrate the procedure of tone smoothing after Fig. 31 ;

Fig. 33

a flow chart illustrating the offset detection and correction in Fig. 27 ;

Fig. 34

a schematic representation of a section of a two-way rectified filtered audio signal and its interpolation to Illustrating the procedure according to Fig. 33 ; and

Fig. 35

a section of a two-way rectified filtered audio signal and its interpolation for the case of a potential segment extension.

With reference to the following description of the figures, it is pointed out that the present invention is described there merely by way of example with reference to a specific application, namely the generation of a polyphonic ringing melody from an audio signal. However, it should be pointed out explicitly that the present invention is of course not limited to this application, but that a melody extraction or automatic transcription according to the invention can also be used elsewhere, such as e.g. for facilitating searching in a database, merely recognizing pieces of music, enabling copyright protection by objectively comparing pieces of music or the like, or just transcribing audio signals to display the transcription result to a musician.

Fig. 1 shows an embodiment of a device for generating a polyphonic melody from an audio signal containing a desired tune. In other words, shows Fig. 1 a device for the rhythmic and harmonic conditioning and re-instrumentation of a melody representing audio signal and to complement the resulting melody to a suitable accompaniment.

The device of Fig. 1 , indicated generally at 300, includes an input 302 for receiving the audio signal. In the present case, it is assumed by way of example that the device 300 or the input 302, respectively, the audio signal in a time sample representation, such as a WAV file, expected. However, the audio signal could also be present in other form at input 302, such as in uncompressed or compressed form or in a frequency band representation. The device 300 further comprises an output 304 for outputting a polyphonic melody in any format, in the present case starting from an output of the polyphonic melody in MIDI format (MIDI = musical instrument digital interface). Between the input 302 and the output 304, an extraction device 304, a rhythm device 306, a key device 308, a harmony device 310 and a synthesis device 312 are connected in series in this order. Furthermore, the device 300 comprises a melody memory 314. An output of the key device 308 is connected not only to an input of the subsequent harmony device 310, but also to an input of the melody memory 314. Accordingly, the input of the harmony device 310 is not limited to the output of the processing direction but also with an output of the melody memory 314. Another input of the melody memory 314 is provided to receive a provision identification number ID. Another input of the synthesizer 312 is configured to receive style information. The meaning of the style information and the provision identification number is apparent from the following functional description. Extractor 304 and rhythm means 306 together form a rhythm editor 316.

Once above, the structure of the device 300 of Fig. 1 has been described, their operation will be described below.

The extraction device 304 is designed to subject the audio signal received at the input 302 to note extraction or recognition in order to extract a signal from the audio signal To get a score. The note sequence 318, which forwards the extraction device 304 to the rhythm device 306, in the present exemplary embodiment is in a form in which, for each note n, a note start time t _n indicating the beginning of the note, for example, in seconds, a tone or note duration τ _n , which indicates the note duration of the note, for example, in seconds, a quantized note or pitch, ie C, Fis or the like, for example as a MIDI note, a volume Ln of the note and an exact frequency f _{n of} the tone or the note is contained in the note sequence, where n is to represent an index for the respective note in the note sequence, which increases with the order of the successive notes or indicates the position of the respective note in the note sequence.

The auto-transcription performed by the note sequence generation unit 318 will be described later with reference to FIG Fig. 2 - 35 explained in more detail.

The note sequence 318 still represents the melody as it was also represented by the audio signal 302. The note sequence 318 is now fed to the rhythm device 306. The rhythm means 306 is arranged to analyze the supplied note sequence to one bar length, one prelude, i. a clock raster to determine the note sequence and thereby the individual notes of the note sequence of appropriate clock-quantified lengths, such as. whole, half, quarter, eighth notes, etc., for the particular bar, and to match the notes' notes to the bar pattern. The note sequence that the rhythm device 306 outputs, thus represents a rhythmically processed note sequence 324.

At the rhythmically processed note sequence 324, the key device 308 performs a key determination and possibly a key correction. More specifically, that determines Means 308 based on the note sequence 324 a major key of the user melody represented by the note sequence 324 and the audio signal 302 including the tone gender, ie major or minor, of the example sung piece. Thereafter, it also recognizes, at this point, non-scale notes in the note sequence 114 and corrects them to arrive at a harmonic-sounding final result, a rhythmically edited and pitch-corrected note sequence 700, which is passed to the harmony device 310 and corrects a key Represents the form of the user's desired melody.

The functioning of the device 324 with regard to the determination of the key can be carried out in various ways. The key determination may, for example, those in the article Krumhansl, Carol L .: Cognitive Foundations of Musical Pitch, Oxford University Press, 1990 , or in the article Temperley, David: The cognition of basic musical structures. The MIT Press, 2001 , described manner take place.

The harmony device 310 is configured to receive the note sequence 700 from the device 308 and to find a suitable accompaniment for the tune represented by this note sequence 700. For this purpose, device 310 acts or acts in a cyclic manner. Specifically, the means 310 operates on each clock as determined by the clock raster set by the rhythm means 306, such that it provides statistics about the tones of the notes T _n occurring in the respective clock. The statistic of the occurring tones is then compared with the possible chords of the major scale scale as determined by the key device 308. Means 310 then selects, among the possible chords, in particular, that chord whose tones best match the notes that are in the respective measure, as indicated by statistics. To this Thus, for each clock, means 310 determines the chord that best fits the notes or notes, for example, in the respective measure. In other words, the means 310 allocates chord levels of the root key to the pitches found by the means 306 in dependence on the pitch, so that a chord progression forms over the course of the melody. Consequently, at the output of the device 310, in addition to the rhythmically processed and key-corrected note sequence including NL, it also outputs a chord step specification to the synthesis device 312 for each measure.

Synthesizer 312 uses style information that can be entered by a user as indicated by case 702 to perform the synthesis, ie, artificially generate the eventually resulting polyphonic melody. For example, the stylist information allows a user to select from four different styles in which the polyphonic melody can be generated, namely Pop, Techno, Latin or Reggae. For each of these styles, either one or more companion patterns are stored in the synthesis device 312. To generate the accompaniment, the synthesizer 312 now uses the accompaniment pattern (s) indicated by the style information 702. To produce the accompaniment, the synthesis device 312 hangs the accompaniment patterns per cycle together. If the chord determined by the device 310 is a clock around the chord version in which an accompaniment pattern already exists, then the synthesizer 312 simply selects the corresponding accompaniment pattern for the current style for that accompaniment clock. However, if, for a particular clock, the chord determined by the means 310 is not the one in which an accompaniment pattern is stored in the means 312, the synthesizer 312 shifts the notes of the accompaniment pattern by the corresponding semitone, respectively, and changes the sixth and fifth by a semitone in the case of another Sound-like, namely by shifting by a semitone up in the case of a major chord reversed in the case of a minor chord.

Further, the synthesizer 312 orchestrates the melody represented by the note string 700 forwarded from the harmony 310 to the synthesizer 312 to obtain a main melody, and then combines accompaniment and main melody into a polyphonic melody, exemplified herein in the form of MIDI File at output 304 outputs.

The key device 308 is further configured to store the note string 700 in the melody memory 314 under a provision identification number. If the user is dissatisfied with the result of the polyphonic melody at the output 304, he may re-enter the provision identification number together with a new style information into the apparatus of FIG Fig. 1 whereupon the melody memory 314 forwards the sequence 700 stored under the providing identification number to the harmony device 310, which then determines the chords as described above, whereupon the synthesizer 312 uses the new style information depending on the chords a new accompaniment and depending on the note sequence 700 generates a new main melody and joins together to form a new polyphonic melody at the output 304.

The following will now be based on the Fig. 2 - 35 the operation of the extraction device 304 described. It is first referring to the Fig. 2 - 26 the procedure for melody recognition in the case of polyphonic audio signals 302 at the input of device 304 is described.

Fig. 2 shows first the rough procedure in the melody extraction or autotranscription. Starting point is the reading or the input of the audio file in a step 750, which, as described above, may be present as a WAV file. The device 304 then performs a frequency analysis on the audio file in a step 752 to thereby provide a time / frequency representation or spectrogram of the audio signal contained in the file. In particular, step 752 includes decomposing the audio signal into frequency bands. In the course of a windowing, the audio signal is subdivided into preferably time-overlapping time segments which are then spectrally decomposed in each case in order to obtain a spectral value for each of a set of spectral components for each time interval or each frame. The set of spectral components depends on the choice of the transformation underlying the frequency analysis 752, a specific embodiment of which will be described below with reference to FIG Fig. 4 is explained.

After step 752, means 304 determines a weighted amplitude spectrum or a perceptual spectrogram, respectively, in step 754. The detailed procedure for determining the perceptual spectrogram will be described in the following Fig. 3 - 8 explained in more detail. The result of step 754 is to rescale the spectrogram obtained from the frequency analysis 752 using the equal volume curves that reflect the human perceptual sensation to fit the spectrogram to the human perceptual perception.

The processing 756 subsequent to step 754 uses, among other things, the perceptual spectrogram obtained from step 754 to finally obtain the melody of the output signal in the form of a musical segmented melody line, ie in a form in which groups of consecutive frames interleave each other same pitch, these groups being temporally spaced over one or more frames spaced from each other, so do not overlap and thus correspond to note segments of a monophonic melody.

In Fig. 2 For example, processing 756 is decomposed into three substeps 758, 760 and 762. In the first sub-step, the perceptual spectrogram is used to obtain a time / fundamental frequency representation from the same, and in turn to use this time / fundamental frequency representation to determine a melody line such that each frame in a unique manner has exactly one spectral component Frequency bin is assigned. The time / fundamental representation accounts for the division of sounds into partial tones by first delogarithmizing the perceptual spectrogram from step 754 to summate for each frame and for each frequency bin the delogarithmized perceptual spectral values at that frequency bin and the overtones to the respective frequency bin. The result is a sound spectrum per frame. From this sound spectrum, the determination of the melody line is carried out by selecting for each frame the fundamental tone or the frequency or that frequency bin at which the sound spectrum has its maximum. The result of step 758 is thus, so to speak, a melody line function which uniquely assigns exactly one frequency bin to each frame. This melody line function in turn defines a melody line progression in the time / frequency domain or a two-dimensional melody matrix spanned by the possible speech components or bins on one side and the possible frames on the other side.

The following substeps 760 and 762 are provided to segment the continuous melody line, thus giving single notes. In Fig. 2 the segmentation is divided into two sub-steps 760 and 762, depending on whether the segmentation takes place in input frequency resolution, ie in Frequenzbinauflösung, or whether the Segmentation takes place in halftone resolution, ie after quantization of the frequencies to semitone frequencies.

The result of processing 756 is processed in step 764 to generate a sequence of notes from the melody line segments, each note being assigned a note start time, a note duration, a quantized pitch, an exact pitch, and so on.

Having now made reference to above Fig. 2 the operation of the extraction device 304 of Fig. 1 is more generally described, reference is made in the following Fig. 3 its operation in more detail in the case where the music represented by the audio file at input 302 is of polyphonic origin. The distinction between polyphonic and monophonic audio signals stems from the observation that monophonic audio signals often originate from less well-trained individuals and therefore have musical inadequacies that require a slightly different approach to segmentation.

In the first two steps 750 and 752 is correct Fig. 3 With Fig. 2 An audio signal is first provided 750 and then subjected to a frequency analysis 752. For example, according to an embodiment of the present invention, the WAV file is in a format since the individual audio samples are sampled at a sampling frequency of 16 kHz. The individual samples are present, for example, in a 16-bit format. Furthermore, it is assumed in the following by way of example that the audio signal is present as a mono-file.

The frequency analysis 752 can then be carried out, for example, by means of a warped filter bank and an FFT (Fast Fourier Transformation). In particular, in the frequency analysis 752, the sequence of audio values is first windowed with a window length of 512 samples, where with a Hopsize of 128 samples, ie the windowing is repeated every 128 samples. Together with the sampling rate of 16 kHz and the quantization resolution of 16 bits, these parameters represent a good compromise between time and frequency resolution. In these exemplary settings, a time frame corresponds to a duration of 8 milliseconds.

The warped filter bank is used according to a special embodiment for the frequency range up to about 1550 Hz. This is necessary to achieve a sufficiently good resolution for low frequencies. For a good halftone resolution enough frequency bands should be available. At a lambda value of -0.85 at 16 kHz sampling rate, approximately two to four frequency bands correspond to one semitone on a frequency of 100 Hz. For low frequencies, each frequency band can be assigned a semitone. For the frequency range up to 8 kHz the FFT is used. The frequency resolution of the FFT is sufficient from about 1,550 Hz for a good halftone representation. Here approx. Two to six frequency bands correspond to one semitone.

In the implementation described above by way of example, the transient response of the warped filter bank must be taken into account. Preferably, therefore, a temporal synchronization in the combination of the two transformations is made. For example, the first 16 frames of the filterbank output are discarded, as well as the last 16 frames of the output spectrum FFT are disregarded. With a suitable design, the amplitude level of the filter bank and FFT is identical and requires no adaptation.

Fig. 4 shows by way of example an amplitude spectrum or a time / frequency representation or a spectrogram of an audio signal, as obtained by the preceding embodiment of a combination of a warped filter bank and an FFT. Along the horizontal axis in Fig. 4 is the time t in seconds s worn away while along the vertical axis, the frequency f is in Hz. The height of the individual spectral values is gray scale. In other words, the time / frequency representation of an audio signal is a two-dimensional field spanned by the possible frequency bins or spectral components on one side (vertical axis) and the time segments or frames on the other side (horizontal axis), each Position of this field is assigned to a specific tuple of frame and Frequenzbin a spectral value or an amplitude.

According to a specific embodiment, the amplitudes in the spectrum of Fig. 4 is still postprocessed in the frequency analysis 752, as the amplitudes calculated by the warped filterbank may sometimes not be accurate enough for subsequent processing. The frequencies that are not exactly at the center frequency of a frequency band, have a lower amplitude value than frequencies that correspond exactly to the center frequency of a frequency band. In addition, in the output spectrum of the warped filter bank crosstalk to adjacent frequency bands, which are also referred to as bins or frequency bins.

To correct the erroneous amplitudes of the effect of crosstalk can be exploited. This error affects a maximum of two adjacent frequency bands in each direction. According to one embodiment, therefore, in the spectrogram of Fig. 4 within each frame adds the amplitudes of adjacent bins to the amplitude value of a middle bin, and this for all bins. Because there is a risk that incorrect amplitude values will be calculated when two audio frequencies are particularly close together in a music signal, and thus phantom frequencies are generated having greater values than the two original sine parts, according to a preferred embodiment only the amplitude values of the directly adjacent neighbor bins will be generated added to the amplitude of the original signal component.

This represents a compromise between accuracy and the occurrence of side effects caused by the addition of the directly adjacent bins. Despite the lower accuracy of the amplitude values, this trade-off is acceptable in the context of melody extraction, since the change in the calculated amplitude value can be neglected in the addition of three or five frequency bands. In contrast, the emergence of phantom frequencies is much more important. The generation of phantom frequencies increases with the number of simultaneous sounds in a piece of music. When searching for the melody line, this can lead to wrong results. The calculation of the exact amplitudes is preferably carried out both for the warped filter bank and for the FFT, so that the music signal is subsequently represented over the entire frequency spectrum by an amplitude level.

The above embodiment for a signal analysis from a combination of a warped filter bank and an FFT enables a hearing-friendly frequency resolution and the presence of sufficient frequency bins per semitone. For more details on the implementation, refer to the thesis of Claas Derboven entitled "Implementation and investigation of a method for the detection of sound objects from polyphonic audio signals", developed at the Technical University Ilmenau in 2003, and the diploma thesis of Olaf Schleusing entitled " Investigation of frequency range transformations for metadata extraction from audio signals ", developed at the Technical University Ilmenau in 2002, referenced.

p₀ gibt hierbei den Bezugsschalldruck an, d.h. den Lautstärkepegel, der den kleinsten wahrnehmbaren Schalldruck bei 1.000 Hz besitzt.As mentioned above, the analysis result of frequency analysis 752 is a matrix of spectral values. These spectral values represent the volume by the amplitude. However, the human volume perception possesses a logarithmic division. It is therefore useful to the amplitude spectrum adapt this classification. This is done in a logarithmization 770 following step 752. In logarithmization 770, all spectral values are logarithmized to the level of the sound pressure level, which corresponds to the logarithmic perception of loudness of humans. More specifically, in the logarithmization 770 to the spectral value p in the spectrogram obtained from the frequency analysis 752, p is mapped to a sound pressure level value and a logarithmic spectral value L, respectively $$\mathrm{L}\left[\mathrm{dB}\right]\mathrm{=}\mathrm{20}\log \left(\frac{\mathrm{p}}{{\mathrm{p}}_{\mathrm{0}}}\right)$$

p ₀ indicates the reference sound pressure, ie the volume level, which has the smallest perceptible sound pressure at 1,000 Hz.

Within logarithmization 770, this reference value must first be determined. While the smallest perceptible sound pressure p _{0 is} used as the reference value in analog signal analysis, this law is not easily transferred to digital signal processing. In order to determine the reference value, according to one exemplary embodiment, therefore, a trial audio signal is used for this purpose, as described in US Pat Fig. 7 is illustrated. Fig. 7 shows the rehearsal audio signal 772 over time t, where in the Y direction the amplitude A is plotted in the smallest representable digital units. As can be seen, the sample audio signal or reference signal 772 is present with an amplitude value of one LSB or with the smallest representable digital value. In other words, the amplitude of the reference signal 772 only oscillates by one bit. The frequency of the reference signal 772 corresponds to the frequency of the highest sensitivity of the human Hearing threshold. However, other determinations of the benchmark may be more beneficial on a case-by-case basis.

In Fig. 5 is exemplarily the result of the logarithmization 770 of the spectrogram of Fig. 4 shown. If, due to the logarithmization, a part of the logarithmic spectrogram is in the negative value range, these negative spectral or amplitude values are set to 0 dB to avoid non-meaningful results in the further processing in order to obtain positive results over the entire frequency range , For the sake of brevity, please note that in Fig. 5 the logarithmized spectral values in the same way as in Fig. 4 are shown, that is arranged in a spanned by the time t and the frequency f matrix and grayscale depending on the value, namely the darker the greater the respective spectral value.

The volume rating of humans is frequency dependent. Therefore, the logarithmic spectrum, as it results from the logarithm 770, must be evaluated in a subsequent step 772 in order to adapt to this frequency-dependent evaluation of the human. For this purpose, the curves of equal volume 774 are used. In particular, the score 772 is therefore necessary to match the different amplitude scores of the musical sounds across the frequency scale of human perception since, according to human perception, the amplitude values of low frequencies experience a lower score than amplitudes of higher frequencies.

For the curves 774 of equal volume, the curve characteristic from DIN 45630 Part 2, Deutsches Institut für Normung eV, Grundlagen der Sound measurement, normal curves of equal volume, 1967, used. The graph history is in Fig. 6 is shown. Like it out Fig. 6 As can be seen, the equal volume curves 774 are respectively assigned to different volume levels indicated in phon. In particular, these curves 774 represent functions that assign each frequency a sound pressure level in dB such that all the sound pressure levels that are on the respective curve correspond to the same volume level of the respective curve.

verwendet werden. Zwischen diesem Kurvenverlauf und der Hörschwelle unter DIN-Norm sind allerdings Abweichungen im tief- und hochfrequenten Wertbereich vorhanden. Zur Anpassung können die Funktionsparameter der Ruhe-Hörschwelle nach der obigen Gleichung verändert werden, um den Verlauf der niedrigsten Lautstärkekurve der oben genannten DIN-Norm von Fig. 6 zu entsprechen. Danach wird diese Kurve vertikal in Richtung höherer Lautstärkepegel in Abständen von 10 dB verschoben und die Funktionsparameter an die jeweilige Charakteristik der Funktionsgraphen 774 angepasst. Die Zwischenwerte werden in 1-dB-Schritten durch lineare Interpolation ermittelt. Vorzugsweise kann die Funktion mit dem höchsten Wertebereich einen Pegel von 100 dB bewerten. Dies ist ausreichend, da eine Wortbreite von 16 Bit einem Dynamikbereich von 98 dB entspricht.Preferably, the equal volume curves 774 are present in the device 204 in analytic form, of course, it would also be possible to provide a look-up table which assigns a volume level value to each pair of frequency bin and sound pressure level quantization value. For the volume curve with the lowest volume level, for example, the formula $$\frac{{\mathrm{L}}_{{\mathrm{T}}_{}}}{}$$$$\frac{{{\mathrm{}}_{\mathrm{4}}}_{}}{\mathrm{dB}}\mathrm{=}\mathrm{3}\mathrm{.}\mathrm{64}{\left(\frac{\mathrm{f}}{\mathrm{kHz}}\right)}^{\mathrm{-}\mathrm{0}\mathrm{.}\mathrm{8th}}\mathrm{-}\mathrm{6}\mathrm{.}\mathrm{5}{\exp }^{\left(\mathrm{-}\mathrm{0}\mathrm{.}\mathrm{6}{\left(\frac{\mathrm{f}}{\mathrm{kHz}}\mathrm{-}\mathrm{3}\mathrm{.}\mathrm{3}\right)}^{\mathrm{2}}\right)}\mathrm{+}{\mathrm{10}}^{\mathrm{-}\mathrm{3}}{\left(\frac{\mathrm{<figure-callout\; id="4"\; label="f\; kHz"\; filenames="imgf0005.png,imgf0006.png"\; state="\{\{state\}\}">f\; </figure-callout>}}{\mathrm{kHz}}\right)}^{\mathrm{4}}$$

be used. However, deviations in the low- and high-frequency value range are present between this curve and the hearing threshold under DIN standard. For adaptation, the function parameters of the resting hearing threshold can be changed according to the above equation to the curve of the lowest volume curve of the above-mentioned DIN standard of Fig. 6 correspond to. Thereafter, this curve is shifted vertically in the direction of higher volume levels at intervals of 10 dB, and the function parameters are adapted to the respective characteristic of the function graphs 774. The intermediate values are determined in 1 dB increments by linear interpolation. Preferably, the function with evaluate the highest value range to a level of 100 dB. This is sufficient, since a word width of 16 bits corresponds to a dynamic range of 98 dB.

Based on the equal volume curves 774, in step 772 means 304 forms each logarithmic spectral value, ie, each value in the array of Fig. 5 depending on the frequency f or frequency bin to which it belongs and its value representing the sound pressure level, on a perceptual spectral value representing the volume level.

The result of this procedure for the case of the logarithmic spectrum of Fig. 5 is in Fig. 8 shown. As can be seen, in the spectrogram of Fig. 8 low frequencies no longer of great importance. Higher frequencies and their overtones are emphasized by this rating. This also corresponds to human perception for evaluating the volume for different frequencies.

The above-described steps 770-774 illustrate possible substeps of step 754 Fig. 2 represents.

The procedure of Fig. 3 After evaluation 772 of the spectrum, in a step 776, it proceeds with a fundamental frequency determination or with the calculation of the total intensity of each sound in the audio signal. For this purpose, in step 776 the intensities of each fundamental tone and the associated harmonics are added up. From a physical point of view, a sound consists of a fundamental tone under the corresponding partial tones. The partial tones are integer multiples of the fundamental frequency of a Sound. The partial or overtones are also called harmonics. In order to sum up the intensity of the same and the associated harmonics for each fundamental tone, a harmonic grid 778 is used in step 776 in order to search for each possible fundamental tone, ie each frequency bin, for overtones that are an integral multiple of the respective one Fundamental tones are. Thus, at a particular frequency bin as a fundamental, further frequency bins corresponding to an integer multiple of the frequency bin of the fundamental are assigned as harmonic frequencies.

In step 776, the intensities in the spectrogram of the audio signal at the respective fundamental tone and its harmonics are then added up for all possible fundamental tone frequencies. In this case, however, a weighting of the individual intensity values is carried out, since due to several sounds occurring simultaneously in a piece of music, there is the possibility that the fundamental tone of a sound is obscured by an overtone of another sound with a lower-frequency fundamental tone. Also overtones of a sound can be obscured by overtones of another sound.

Nevertheless, to determine the associated sounds of a sound, a tone model based on the principle of the model of Mosataka Goto and adapted to the spectral resolution of the frequency analysis 752 is used in step 776, the tone model of Goto in Goto, M .: A Robust Predominant-F0 Estimation Method for Real-time Detection of Melody and Bass Lines, in CD Recordings, Proc. IEEE International Conference on Acoustics, Speech and Signal Processing, Istanbul, Turkey, 2000 , is described.

Starting from the possible fundamental frequency of a sound, harmonic grid 778 assigns the respective harmonic frequencies for each frequency band or frequency bin. According to a preferred embodiment, overtones are searched for fundamental frequencies in only one particular frequency bin range, such as from 80 Hz to 4,100 Hz, and harmonics are considered only up to the 15th order. The overtones of different sounds can be assigned to the sound model of several fundamental frequencies. By this effect, the amplitude ratio of a sought sound can be changed considerably. To attenuate this effect, the amplitudes of the partial tones are evaluated with a halved Gaussian filter. The keynote receives the highest value. All the following partial tones receive a lower weighting according to their order, the weighting decreasing in a Gauss-shaped manner, for example, with increasing order. Thus, an overtone amplitude of another sound that obscures the actual overtone has no particular effect on the overall result of a sought-after voice. As the frequency resolution of the spectrum becomes lower for higher frequencies, a bin with the corresponding frequency does not exist for each higher order overtone. Due to the crosstalk to the adjacent bins of the frequency environment of the sought overtone, the amplitude of the sought harmonic can be simulated relatively well by means of a Gaussian filter over the nearest frequency bands. Therefore, overtone frequencies or the intensities thereon do not have to be determined in units of frequency bins, but interpolation can also be used to accurately determine the intensity value at the overtone frequency.

However, the summation over the intensity values is not performed directly on the perceptual spectrum from step 772. Rather, first in step 776, the perceptual spectrum of Fig. 8 first delogarithmized with the aid of the reference value from step 770. The result is a delogarithmized perceptual spectrum, ie an array of delogarithmized perceptual spectral values for each frequency bin and frame tuple. Within this delogarithmized perceptual spectrum, for each possible root, the spectral value of the fundamental and the possibly interpolated spectral values are added together using the harmonic grid 778 of the associated harmonic, giving a sound intensity value for the frequency range of all possible fundamental frequencies, for each frame - in the previous example In other words, the result of step 776 is a sound spectrogram, step 776 itself corresponding to a level addition within the spectrogram of the audio signal. The result of step 776 is entered, for example, in a new matrix having one row for each frequency bin within the frequency range of possible fundamental frequencies and one column for each frame, wherein in each matrix element, ie at each column and row crossing, the result of the summation is entered as the fundamental tone for the corresponding frequency bin.

Next, in a step 780, a preliminary determination of a potential melody line is made. The melody line corresponds to a function over time, namely one Function that assigns exactly one frequency band or one frequency bin to each frame. In other words, the melody line determined in step 780 defines a track along the domain of definition of the sound spectrogram or matrix of step 776, the track never being ambiguous along the frequency axis.

The determination is made in step 780 such that the maximum amplitude is determined for each frame over the entire frequency range of the sound spectrogram, i. the largest summation value. The result, i. the melody line largely corresponds to the basic course of the melody of the audio track underlying the audio signal 302.

The evaluation of the spectralogram with the equal volume curves in step 772 and the search for the maximum intensity sound result in step 780 take into account the musicological statement that the main melody is that portion of a song that the person perceives loudest and most concisely.

The above-described steps 776-780 illustrate possible substeps of step 758 Fig. 2 represents.

In the potential melody line from step 780 are segments that do not belong to the melody. In melody pauses or between melody notes, dominant segments such as those from the bass progression or other accompaniment instruments are found. These melody pauses need to go through the later steps in Fig. 3 be eliminated. In addition, short, individual elements that can not be assigned to any area of the title arise. They become, for example by means of a 3x3 average filter, as will be described below.

After the determination of the potential melody line in step 780, a general segmentation 782 is first performed in a step 782 which ensures that parts of the potential melody line are eliminated which prima facie can not belong to the actual melody line. In Fig. 9 For example, the result of the melody line determination of step 780 is exemplary of the case of the perceptual spectrum of Fig. 8 shown. Fig. 9 shows the melody line plotted over the time t or over the sequence of frames along the x-axis, along the y-axis, the frequency f and the frequency bins are displayed. In other words, in Fig. 9 the melody line from step 780 is represented in the form of a binary image array, which is also sometimes referred to as a melody matrix and has one row for each frequency bin and one column for each frame. All points of the array where the melody line is not located have a value of 0 or are white, while the points of the array where the melody line is located have a value of 1 or are black. These points are thus located on frequency bin and frame tuples, which are associated with each other by the melody line function of step 780.

At the melody line of Fig. 9 , in the Fig. 9 labeled 784, step 782 of general segmentation now operates, with reference to a possible implementation Fig. 10 is explained in more detail.

The general segmentation 782 begins in a step 786 with the filtering of the melody line 784 in the frequency / time domain in a representation in which the melody line 784 as in FIG Fig. 9 shown as a binary track in an array, which is spanned by the frequency bins on one side and the frames on the other side. The pixel array of Fig. 9 For example, consider an x by y pixel array, where x is the number of frames and y is the number of frequency bins.

Step 786 is now intended to remove smaller outliers or artifacts in the melody line. Fig. 11 shows by way of example in a schematic form a possible course of a melody line 784 in a representation according to FIG Fig. 9 , As can be seen, the pixel array shows regions 788 in which there are sporadic black pixel elements corresponding to portions of the potential melody line 784 which, due to their temporal brevity, are certainly not part of the actual melody and therefore should be removed.

In step 786, therefore, the pixel array is replaced by Fig. 9 respectively. Fig. 11 in which the melody line is represented in binary form, first generates a second pixel array by entering for each pixel a value corresponding to the summation of the binary values at the corresponding pixel and the pixel adjacent to this pixel. Be on this Fig. 12a Referenced. There is an exemplary excerpt from the course of a melody line in the binary image of Fig. 9 or Fig. 11 shown. The exemplary section of Fig. 12a includes five rows corresponding to different frequency bins 1-5 and five columns AE corresponding to different adjacent frames. The course of the melody line is in Fig. 12a symbolized by the fact that the corresponding pixel elements representing parts of the melody line are hatched. According to the embodiment of FIG. Fig. 12a Therefore, the frequency bin 4, the frame C, the frequency bin 3, etc. are assigned to the frame B by the melody line. Of course, the frame A is also assigned a frequency bin by the melody line, but this is not among the five frequency bins from the section of Fig. 12a ,

In the filtering in step 786, the binary value of the same as well as the binary value of the neighboring pixels is first summed up for each pixel 790, as already mentioned. This is for example in Fig. 12a As an example of the pixel 792, it illustrates which figure at 794 has a square drawn around the pixels adjacent to the pixel 792 and the pixel 792 itself. Consequently, a sum value of 2 would result for the pixel 792, since in the area 794 around the pixel 792 there are only 2 pixels belonging to the melody line, namely the pixel 792 itself and the pixel C3, ie at the frame C and the bin 3. This summation is repeated by shifting the area 794 for all other pixels, resulting in a second pixel image, also sometimes referred to below as an intermediate matrix.

This second pixel image is then subjected to a pixel-by-pixel mapping, wherein in the pixel image all summation values from 0 or 1 to zero and all summation values greater than or equal to 2 are mapped to one. The result of this mapping is for the exemplary case of Fig. 12a in Fig. 12a represented by numbers of "0" and "1" in the individual pixels 790. As can be seen, the combination of 3x3 summation and subsequent mapping to "0" and "1" using the Threshold 2 causes the melody line to "smear". The combination acts as a kind of low-pass filter, which would be undesirable. Therefore, as part of step 786, the first pixel image, ie Fig. 9 respectively. Fig. 11 , or in Fig. 12a the pixel image, which is illustrated by the hatched pixels, with the second pixel array, ie the one in Fig. 12a is illustrated by the zeros and ones multiplied. This multiplication prevents low-pass filtering of the melody line by the filtering 786 and further ensures the uniqueness of the assignment of frequency bins to frames.

The result of the multiplication for the clipping Fig. 12a is that filtering 786 does not change anything on the melody line. This is also desirable at this point since the melody line is obviously contiguous in this area and the filtering from step 786 is only intended to remove outliers or artifacts 788.

Fig. 12b Therefore, to illustrate the mode of operation of the filtering 786, there is shown another exemplary excerpt from the melody matrix of FIG Fig. 9 respectively. Fig. 11 , As can be seen there, the combination of summation and threshold mapping results in an intermediate matrix in which two separated pixels P4 and R2 receive a binary value of 0, although at these pixel positions the melody matrix has a binary value of 1 as indicated by the Hatching in Fig. 12b to recognize that is to show at these pixel positions the melody line. These isolated "outliers" of the melody line are therefore removed by the filtering in step 786 after multiplication.

After step 786, general segmentation 782 is followed by step 796, in which portions of melody line 784 are removed by neglecting those portions of the melody line that are not within a predetermined frequency range. In other words, in step 796, the value range of the melody line function is restricted to the predetermined frequency range from step 780. In other words, in step 796, all pixels of the melody matrix of Fig. 9 respectively. Fig. 11 set to zero, which are outside the predetermined frequency range. In the case of a polyphonic analysis as adopted herein, a frequency range is, for example, 100-200-200-10000 Hz, and preferably 150-1050 Hz. In the case of a monophonic analysis as described with reference to FIGS Fig. 27 ff., a frequency range is, for example, from 50-150 to 1,000-1,100 Hz and preferably from 80 to 1,050 Hz. The limitation of the frequency range to this bandwidth contributes to the observation that melodies in popular music are usually represented by vocals, the is in this frequency range as well as human language.

To illustrate step 796, see Fig. 9 By way of example, a frequency range of 150 to 1050 Hz is indicated by a lower limit frequency line 798 and an upper limit frequency line 800. Fig. 13 FIG. 12 shows the melody line filtered by step 786 and clipped by step 796, which is used to distinguish in FIG Fig. 13 is provided with the reference numeral 802.

After step 796, in a step 804, removal of portions of the melody line 802 of too small an amplitude occurs, wherein the extraction means 304 in this case on the logarithmic spectrum Fig. 5 from step 770. More specifically, the extractor 304 beats in the logarithmic spectrum of each frequency bin and frame tuple through which the melody line 802 passes Fig. 5 after the corresponding logarithmic spectral value, and determines whether the corresponding logarithmic spectral value is less than a predetermined percentage of the maximum amplitude or logarithmic spectral value in the logarithmic spectrum of Fig. 5 is. In the case of polyphonic analysis, this percentage is preferably between 50 and 70% and preferably 60%, while in monophonic analysis this percentage is preferably between 20 and 40% and preferably 30%. Parts of the melody line 802 for which this is the case are neglected. This approach takes into account the fact that a melody usually always has approximately the same volume, or that sudden extreme volume fluctuations are unlikely to be expected. In other words, in step 804, all pixels of the melody matrix of Fig. 9 and 17 are set to zero at which the logarithmic spectral values are less than the predetermined percentage of the maximum logarithmic value.

Step 804 is followed, in a step 806, by a separation of those sections of the remaining melody line at which the course of the melody line in the frequency direction changes abruptly so as to have only a short halfway uniform melody curve. To this, too explain, refer to Fig. 14 which shows a section of the melody matrix over AM consecutive frames, with the frames arranged in columns as the frequency increases along the column direction from bottom to top. The frequency bin resolution is in Fig. 14 for the sake of clarity not shown.

The melody line as revealed by step 804 is in Fig. 14 exemplified by the reference numeral 808. As can be seen, the melody line 808 in the frames AD remains constant on a frequency bin, to then show a frequency hopping between the frames D and E that is greater than a semitone distance HT. Between the frames E and H, the melody line 808 then remains constant again on a frequency bin, in order then to fall back from frame H to frame I by more than one semitone interval HT. Such a frequency hopping, which is greater than a semitone distance HT, also occurs between the frames J and K. From then on, the melody line 808 between frames J and M again remains constant on a frequency bin.

To carry out the steps 806, the device 304 now scans the melody line in a frame-wise manner, for example from the front to the back. In this case, the device 304 checks for each frame whether a frequency jump greater than the semitone distance HT takes place between this frame and the subsequent frame. If so, means 304 marks these frames. In Fig. 14 the result of this marking is exemplarily illustrated by the fact that the corresponding frames are surrounded by a circle, here the frames D, H and J. In a second step, means 304 now checks between which of the marked frames less than a predetermined number of frames are arranged, wherein in the present case, the predetermined number is preferably three. Overall, this extracts portions of the melody line 808 where the same jumps between immediately consecutive frames less than a semitone but is less than four frame elements long. Between frames D and H there are three frames in the present exemplary case. This means nothing else than that the melody line 808 does not jump over the frames E - H by more than a semitone. However, there is only one frame between the marked frames H and J. This means nothing else than that in the area of the frames I and J, the melody line 808 jumps more than one semitone both forward and backward in the time direction. This section of the melody line 808, namely in the area of the frames I and J, is therefore neglected in the subsequent processing of the melody line. In the current melody matrix, therefore, the corresponding melody line element is set to zero on frames I and J, ie it turns white. This exclusion can therefore comprise at most three consecutive frames, which corresponds to 24 ms. However, tones shorter than 30 ms rarely occur in today's music, so that the exclusion after step 806 does not lead to a deterioration of the transcription result.

After step 806, processing within general segmentation 782 proceeds to step 810, where device 304 divides the remaining remnants of the former potential melody line from step 780 into a sequence of segments. In the division into segments, all elements in the melody matrix are combined into a segment or a trajectory, which are directly adjacent. To illustrate this, shows Fig. 15 a section of the melody line 812, as they are after the Step 806 results. In Fig. 15 only the individual matrix elements 814 from the melody matrix along which the melody line 812 extends are shown. For example, to check which matrix elements 814 are to be segmented, the device 304 scans them in the following manner. The device 304 first of all checks whether the melody matrix at all has a marked matrix element 814 for a first frame. If not, means 304 proceeds to the next matrix element and again checks the next frame for the presence of a corresponding matrix element. Otherwise, that is, if a matrix element that is part of the melody line 812 is present, the device 304 checks the next frame for the presence of a matrix element that is part of the melody line 812. If so, means 304 further checks if that matrix element is directly adjacent to the matrix element of the previous frame. Immediately adjacent is one matrix element to another if they are directly adjacent one another in the row direction, or if they are diagonally corner to corner. If there is a neighborhood relationship, means 304 checks for the presence of a neighborhood relationship also for the next frame. Otherwise, ie in the absence of a neighborhood relationship, a currently recognized segment on the previous frame ends and a new segment begins on the current frame.

The in Fig. 15 The section of the melody line 812 shown represents an incomplete segment in which all the matrix elements 814 that are part of, or run along, the melody line are immediately adjacent to each other.

The segments found in this way are numbered consecutively, resulting in a sequence of segments.

The result of the general segmentation 782 is thus a sequence of melody segments, each melody segment covering a sequence of immediately adjacent frames. Within each segment, the melody line jumps from frame to frame by at most a predetermined number of frequency bins, in the preceding exemplary embodiment by at most one frequency bin.

After the general segmentation 782, the device 304 proceeds with the melody extraction at step 816. Step 816 is to close the gaps between adjacent segments to address the case that, due to, for example, percussive events in the melody line determination in step 780, other sound components have been inadvertently detected and filtered out in the general segmentation 782. The gap closure 816 will be referred to Fig. 16 in more detail, wherein the gap closure 816 relies on a halftone vector, which is determined in a step 818, the determination of the halftone vector referring to Fig. 17 will be explained in more detail.

Since the gap closure 816 uses the halftone vector, reference will first be made below Fig. 17 the determination of the variable halftone vector explained. Fig. 17 FIG. 12 shows the patchy melody line 812 resulting from the general segmentation 782 in a shape plotted in the melody matrix. Upon determination of the halftone vector in step 818, the device now sets 304 determines which frequency bins the melody line 812 passes through and how often or in how many frames. The result of this approach, illustrated with the case 820, is a histogram 822 indicating, for each frequency bin f, the frequency with which it is traversed by the melody line 812 and how many matrix elements of the melody matrix that are part of the melody line 812 , are arranged at the respective Frequenzbin. From this histogram 822, device 304 then determines in a step 824 the frequency bin with the maximum frequency. This is in Fig. 17 indicated by an arrow 826. Starting from this frequency bin 826 of the frequency f ₀ , the device 304 then determines a vector of frequencies f _i which, relative to one another and above all to the frequency f _{0, have} a frequency spacing which corresponds to an integer multiple of a half tone length HT. The frequencies in the halftone vector will be referred to as halftone frequencies hereinafter. Sometimes, halftone cutoff frequencies will also be referred to below. These are located exactly between adjacent halftone frequencies, ie exactly centered on this. A halftone interval, as is customary in music, is defined as 2 ^{1/12 of} the ^frequency of use f ₀ . By determining the halftone vector in step 818, the frequency axis f along which the frequency bins are plotted can be divided into halftone regions 828 extending from halftone cutoff frequency to the adjacent halftone cutoff frequency.

On this division of the frequency axis f into halftone areas, the gap closure is based on the following Fig. 16 is explained. As previously mentioned, in gap closure 816, attempts are made to close gaps between adjacent segments of melody line 812 which are unintentionally in melody line recognition 780 and general segmentation 782, respectively, as described above. The gap closure is carried out segment by segment. For a current reference segment, within gap closure 816, it is first determined in a step 830 whether the gap between the reference segment and the subsequent segment is less than a predetermined number of p frames. Fig. 18 shows by way of example a section of the melody matrix with a section of the melody line 812. In the example considered, the melody line 812 has a gap 832 between two segments 812a and 812b, of which the segment 812a is the aforementioned reference segment. As can be seen, the gap in the exemplary case of Fig. 18 six frames.

In the present exemplary case, with the preferred sampling frequencies, etc. given above, p is preferably 4. Thus, in the present case, the gap 832 is not less than four frames, whereupon the processing proceeds to step 834 to check if the gap 832 is smaller is equal to q frames, where q is preferably 15. This is the case in the present case, so processing continues at step 836, where it is checked whether the facing segment ends of the reference segment 812a and the successor segment 812b, ie the end of the segment 812a and the beginning of the successor segment 812b, are in a same or in relation to each other adjacent halftone areas lie. In Fig. 18 To illustrate the situation, the frequency axis f is divided into halftone areas, as determined in step 818. As can be seen, lie in the case of Fig. 18 the facing segment ends of the segments 812a and 812b in one and the same halftone area 838.

For this positive check case, in step 836, the gap closure processing continues at step 840, where it is checked which amplitude difference in the perceptual spectrum of step 772 exists at the positions of the end of the reference segment 812a and the beginning of the successor segment 812b. In other words, in step 840, in the perceptual spectrum of step 772, means 304 looks up the respective perceptual spectral values at the positions of the end of segment 812a and the beginning of segment 812b and determines the absolute value of the difference of the two spectral values. Further, means 304 determines in step 840 whether the difference is greater than a predetermined threshold r, preferably being 20-40% and preferably 30% of the perceptual spectral value at the end of the reference segment 812a.

If the determination in step 840 gives a positive result, the gap closure proceeds to step 842. There, means 304 determines a gap closure line 844 in the melody matrix which directly connects the end of the reference segment 812a and the beginning of the successor segment 812b. The gap closure line is preferably rectilinear, as it is in Fig. 18 is shown. More specifically, the connecting line 844 is a function across the frames over which the gap 832 extends, the function assigning a frequency bin to each of these frames so that a desired connecting line 844 results in the melody matrix.

Along this line of connection, then, means 304 determines the corresponding perceptual spectral values from the perceptual spectrum from step 772 by looking up the respective frequency bin and frame tuples of gap closure line 844 in the perceptual spectrum. Via these perceptual spectral values along the gap closure line, means 304 determines the mean and compares it in step 842 with the corresponding averages of perceptual spectral values along reference segment 812a and successor segment 812b. If both comparisons indicate that the mean for the gap closure line is greater than or equal to the average of the reference or successor segment 812a or b, then the gap 832 is closed in a step 846 by entering or closing the gap closure line 844 in the melody matrix the corresponding matrix elements thereof are set to 1. At the same time, in step 846, the list of segments is changed to merge the segments 812a and 812b into a common segment, whereupon the gap closure for the reference segment and the successor segment is completed.

A gap closure along the gap closure line 844 also occurs if, at step 830, the gap 832 is less than 4 frames long. In this case, in a step 848, the gap 832 is closed, as in the case of step 846 along a direct and preferably straight gap closure line 844 connecting the facing ends of the segments 812a-812b, whereupon the gap closure for the two segments is finished and continues with the subsequent segment, as far as such exists. Although this in Fig. 16 is not shown, the gap closure in step 848 will still be made dependent on a condition corresponding to that of step 836, ie, that the two mutually facing segment ends lie in the same or adjacent halftone areas.

If any of the steps 834, 836, 840 or 842 result in a negative check result, the gap closure ends for the reference segment 812a and is performed again for the follower segment 812b.

The result of the gap closure 816 is thus a possibly shortened list of segments or a melody line, which may have gap closure lines in the melody matrix in some places. As was apparent from the previous discussion, at a gap of less than 4 frames, a connection between adjacent segments in the same or adjacent halftone area is always made.

The gap closure 816 is followed by a harmony map 850, which is intended to eliminate errors in the melody line that have arisen by incorrectly determining the wrong root note of a sound in determining the potential melody line 780. In particular, the harmony mapping 850 operates segment by segment to shift individual segments of the melody line resulting from gap closure 816 by an octave, fifth, or major third, as will be described in more detail below. As the following description will show, the conditions for this are strict in order not to erroneously shift a segment wrong in frequency. The Harmoniemapping 850 is in the following detailed reference Fig. 19 and Fig. 20 described.

As already mentioned, the harmony mapping 850 is performed segment by segment. Fig. 20 shows by way of example a section of the melody line, as it has revealed after the gap closure 816. This melody line is in Fig. 20 provided with the reference numeral 852, wherein in the section of Fig. 20 three segments can be seen from the melody line 852, namely the segments 852a-c. The representation of the melody line is again as a track in the melody matrix, but again it is recalled that the melody line 852 is a function that uniquely assigns a single frequency bin to the individual frames - meanwhile not all of them - so that the in Fig. 20 show traces shown.

The segment 852b located between the segments 852a and 852c appears to be cut out of the melody line progression as it would result from the segments 852a and 852c. In particular, in the present case, the segment 852b with no frame gap follows the reference segment 852a by way of example, as indicated by a dashed line 854. Likewise, by way of example, the time range covered by the segment 852b is intended to be directly adjacent to the time range covered by the segment 852c, as indicated by a dashed line 856.

In Fig. 20 are now in the melody matrix or in the time / frequency representation further dashed, dot-dashed and dot-dash lines shown, which also result from a parallel displacement of the segment 852b along the frequency axis f. In particular, a dash-dot line 858 by four semitones, ie a major third, to the segment 852b shifted to higher frequencies. Dashed line 858b is shifted down twelve halftones of frequency direction f, ie one octave. Again a third line 858c is dash-dotted to this line and a quint line 858d is shown as a dot-and-dash line, ie a line shifted by seven semitones towards higher frequencies relative to the line 858b.

Like it Fig. 20 As can be seen, the segment 852b appears to have been erroneously determined in the context of the melody line determination 780, since it would be less abruptly inserted between the adjacent segments 852a and 852c when shifted one octave down. The task of the Harmoniemappings 850 is therefore to check whether a shift to such "outliers" should take place or not, since such frequency jumps occur less frequently in a melody.

The harmony mapping 850 begins with the determination of a melody centroid line by means of a mean value filter in a step 860. In particular, step 860 comprises calculating a moving average of the melody curve 852 with a certain number of frames over the segments in the time direction t, the window length being 80 - 120, for example and preferably 100 frames at the above-mentioned frame length of 8 ms, ie correspondingly different number of frames at a different frame length. More precisely, to determine the melody centroid line, a window of length 100 frames is frame-shifted along the time axis t. In this case, all the frequency bins associated with frames within the filter window by the melody line 852 are averaged and this average value for the frame is entered in the middle of the filter window, which causes repetition of successive frames in the filter frame Case of Fig. 20 results in a melody centroid line 862, a function that uniquely assigns a frequency to the individual frames. The melody centroid line 862 may extend over the entire time range of the audio signal, in which case the filter window at the beginning and end of the piece must be "squashed" accordingly, or only over a range from the beginning and the end of the audio piece half the filter window width is spaced.

In a subsequent step 864, the device 304 checks whether the reference segment 852a is adjacent to the successor segment 852b along the time axis t. If this is not the case, the processing with the subsequent segment as the reference segment is performed again (866).

In the present case of Fig. 20 however, the check at step 864 results in a positive result, whereupon the processing proceeds to step 868. In step 868, the successor segment 852b is virtually shifted to obtain the octave, fifth, and / or third lines 858a-d. The selection of major thirds, fifths and octaves is advantageous in pop music, as there is usually used a major chord, in which the highest and the lowest tone of a chord have a spacing of a major third plus a minor third of a fifth. Alternatively, the above procedure is of course also applicable to minor keys, in which chords of minor third and then major third occur.

In a step 870, then, means 304 in the spectrum evaluated with equal loudness curves or the perceptual spectrum from step 772, respectively, look up the minimum perceptual spectral value along reference segment 852a and the octave, quintet, and / or third line, respectively 858a-d. In the exemplary case of Fig. 20 Consequently, there are five minimum values.

These minimum values are used in subsequent step 872 to select one or none among the octave, fifth, and / or third shift lines 858a-d, depending on whether the octave, fifth, and / or octave shift lines or third-line minimum value has a predetermined reference to the minimum value of the reference segment. In particular, an octave line 858b is selected below the lines 858a-858d if the minimum value is at most 30% less than the minimum value for the reference segment 852a. A quint-line 858d is selected if the minimum value determined for it is at most 2.5% smaller than the minimum value of the reference segment 852a. One of the triplets 858c is used if the corresponding minimum value for that line is at least 10% greater than the minimum value for the reference segment 852a.

Of course, the above-mentioned values used as criteria for selection from lines 858a-858b can be varied, although they provided very good results for pop music pieces. Likewise, it is not absolutely necessary to determine the minimum values for the reference segment or the individual lines 858a-d, but the individual average values could also be used, for example. The advantage of the diversity of the criteria for the individual lines is that this can take into account a probability that an octave, fifth or third jump has erroneously occurred in the melody line determination 780, or that such a jump in the melody was actually wanted.

In a subsequent step 874, the device 304 shifts the segment 852b to the selected line 858a-858d, if such was selected in step 872, provided that the shift points in the direction of the melody centerline 862 as viewed from the follower segment 852b , In the exemplary case from Fig. 20 if the latter condition were satisfied, as long as the third line 858a was not selected in step 872.

After Harmoniemapping 850 takes place in a step 876 a Vibratoerkennung and vibrato compensation, the operation of which reference to the Fig. 21 and 27 is explained in more detail.

Step 876 is performed segment by segment for each segment 878 in the melody line as it results after harmony mapping 850. In Fig. 22 For example, an exemplary segment 878 is shown enlarged in a representation in which the horizontal axis corresponds to the time axis and the vertical axis corresponds to the frequency axis, as was the case in the previous figures. In a first step 880, the reference segment 878 is first examined in the context of the vibrato detection 876 for local extrema. Here again it is recalled that yes, the melody line function and thus also the part of the segment corresponding to the interest clearly maps the frames over this segment unambiguously on frequency bins in order to form the segment 888. This segment function is examined for local extrema. In other words, in step 880, the reference segment 878 is examined for those locations where it has local extremal locations along the time axis with respect to the frequency direction, ie, locations where the slope of the melody line function is zero. These posts are in Fig. 22 exemplified with vertical lines 882 indicated.

In a subsequent step 884, it is checked whether the extrema sites 882 are arranged such that temporally adjacent local extreme sites 882 are located at frequency bins having a frequency spacing greater or less equal to a predetermined number of bins, e.g., 15 to 25, preferably but 22 bins referring to Fig. 4 described implementation of the frequency analysis or a Number of bins per semitone range of about 2 to 6, is. In Fig. 22 is exemplified by a double arrow 886 the length of 22 frequency bins. As can be seen, extremals 882 satisfy criterion 884.

In a subsequent step 888, the device 304 checks whether, between the adjacent extreme digits 882, the time interval is always less than or equal to a predetermined number of time frames, the predetermined number being 21, for example.

If the check in step 888 turns out to be positive, as in the example of Fig. 22 That is, if the number of extrema 882 is greater than or equal to a predetermined number, which is preferably 5 in the present case, it is checked in a step 892, as shown by the double arrow 890, which should be 21 frames long , In the example of Fig. 22 this is given. Thus, even if the check in step 892 is positive, in a subsequent step 894 the reference segment 878 or the detected vibrato is replaced by its mean value. The result of step 894 is in Fig. 22 displayed at 896. More specifically, in step 894, the reference segment 878 on the current melody line is removed and replaced by a reference segment 896 which extends over the same frames as the reference segment 878 but along a constant frequency bin corresponding to the average of the frequency bins through which the replaced reference segment 878 was. If the result of one of the checks 884, 888 and 892 is negative, the vibrato detection or compensation ends for the relevant reference segment.

In other words, vibrato detection and vibrato equalization are performed according to Fig. 21 a vibrato detection by stepwise feature extraction, which searches for local extrema, namely local minima and maxima, with a restriction on the number of allowed frequency bins of the modulation and a limitation in the temporal distance of the extrema, where as a vibrato only a group of at least 5 extrema is considered. A recognized vibrato is then replaced in the melody matrix by its mean value.

After the vibrato detection in step 876, a statistical correction is performed in step 898, which also accounts for the observation that in a tune short and extreme pitch variations are not to be expected. The statistical correction of 898 will be referred to Fig. 23 explained in more detail. Fig. 23 shows by way of example a section of a melody line 900, as it may result from the vibrato recognition 876. Again, the course of the melody line 900 is shown registered in the melody matrix, which is spanned by the frequency axis f and the time axis t. In the statistical correction 898, a melody centerline for the melody line 900 is first determined similar to the step 860 in the harmony mapping. For determination, as in the case of step 860, a window 902 of predetermined length, such as 100 frames in length, is frame-shifted along the time axis t to calculate, frame by frame, an average of the frequency bins containing the melody line 900 within the window 902, the average being assigned to the frame in the middle of the window 902 as a frequency bin, resulting in a point 904 of the melody centerline to be determined. The resulting melody centerline is in Fig. 23 indicated by the reference numeral 906.

This will cause a second window to appear in Fig. 23 not shown, shifted along the time axis t frame-wise, which has, for example, a window length of 170 frames. For each frame, the standard deviation of the melody line 900 to the melody center line 906 is determined. The resulting standard deviation for each frame is multiplied by 2 and added by 1 bin. This value then becomes the respective frequency bin for each frame, which is the Melody centerline 906 at this frame traverses, adds, and subtracts to obtain upper and lower standard deviation lines 908a and 908b. The two standard deviation lines 908a and 908b define an allowed area 910 between them. As part of the statistical correction 898, all segments of the melody line 900 that are completely outside of the approval area 910 are now removed. The result of the statistical correction 898 is thus a reduction in the number of segments.

The step 898 is followed by a halftone mapping 912. The halftone mapping is performed frame-wise by resorting to the halftone vector, step 818, which defines the halftone frequencies. The halftone mapping 912 functions such that, for each frame on which the melody line resulting from step 898 is present, it is checked in which of the halftone areas the frequency bin lies in which the melody line passes through the respective frame Frequency bin the melody line function the respective frame maps. The melody line is then changed such that in the respective frame the melody line is changed to the frequency value corresponding to the semitone frequency of the semitone area in which the frequency bin through which the melody line passed was.

Instead of the frame-wise halftone mapping or quantization, segment-wise semitone quantization can also be carried out, for example by assigning only the frequency mean value per segment to one of the halftone areas and thus to the corresponding halftone area frequency in the previously described manner, which then over the entire time length of the corresponding segment the frequency is used.

Steps 782, 816, 818, 850, 876, 898 and 912 thus correspond to step 760 in FIG Fig. 2 ,

Upon halftone mapping 912, an onset detection and correction per segment is performed in step 914. This will be referred to the Fig. 24 - 26 explained in more detail.

The goal of the onset detection and correction 914 is to correct the individual segments of the melody line resulting from the halftone mapping 912, which correspond more and more to the individual notes of the searched tune, with respect to their starting times. For this purpose, the incoming or in step 750 provided audio signal 302 is used again, as will be described in more detail below.

In a step 916, first the audio signal 302 is filtered with a bandpass filter corresponding to the halftone frequency to which the respective reference segment has been quantized in step 912, or with a bandpass filter having cutoff frequencies between which the quantized halftone frequency of the respective segment lies , Preferably, the bandpass filter is used as one having cutoff frequencies corresponding to the halftone cutoff frequencies f _u and f _{o of} the halftone region in which the considered segment is located. Still further preferably, as the band-pass filter, an IIR band-pass filter with the cutoff frequencies f _u and f _o associated with the respective halftone region is filtered as filter cutoff frequencies or with a Butterworth bandpass filter whose transfer function in Fig. 25 is shown.

Subsequently, in a step 918, a two-way rectification of the audio signal filtered in step 916 is performed, whereupon, in a step 920, the time signal obtained in step 918 is interpolated and the interpolated time signal is convolved with a Hamming window, whereby an envelope of the two-way rectified and filtered audio signals, respectively, is determined.

Steps 916-920 will be referred to Fig. 26 once again illustrated. Fig. 26 Numeral 922 shows the two-way rectified audio signal, as it does after step 918, in a graph plotting horizontally the time t in virtual units and vertically the amplitude of the audio signal A in virtual units. Further, in the graph, the envelope 924 resulting in step 920 is shown.

Steps 916-920 are only one way of generating the envelope 924 and of course can be varied. In any case, envelopes 924 are generated for the audio signal for all those semitone frequencies or halftone areas in which segments or note segments of the current melody line are arranged. For each such envelope 924, the following steps of Fig. 24 executed.

First, in a step 926, potential start times are determined as the locations of locally maximum rise of the envelope 924. In other words, inflection points in the envelope 924 are determined in step 926. The times of the turning points in the case of Fig. 26 are illustrated with vertical bars 928.

For the following evaluation of the determined potential starting times or potential increases, a downsampling to the time resolution of the preprocessing is carried out, if necessary in the context of step 926, which is described in Fig. 24 not shown. It should be noted that in step 926 not all potential start times or all inflection points have to be determined. It is also not necessary that all determined or determined potential starting times must be supplied to the subsequent processing. Rather, it is possible to determine or further process only those inflection points as potential starting times, which are arranged in temporal proximity before or in a time range of one of the segments corresponds to the melody line located in the halftone area underlying the determination of the envelope 924.

In a step 928 it is now checked whether, for a potential start time, it lies before the segment start of the same corresponding segment. If so, processing continues at step 930. Otherwise, however, i. if the potential start time is past the existing start of the segment, step 928 is repeated for a next potential start time, or step 926 for a next envelope determined for another half-tone range, or segmented onset detection and correction is performed for a next segment ,

In step 930, a check is made as to whether the potential start time is more than x frames before the beginning of the corresponding segment, where x is between 8 and 12, for example, and preferably 10 with a frame length of 8 ms, changing the values for other frame lengths accordingly would. If this is not the case, ie if the potential start time or the determined start time is 10 frames before the segment of interest, the gap between the potential start time and the previous segment start is closed or the previous start of the segment is corrected to the potential start time in a step 932 , If necessary, the predecessor segment is correspondingly shortened or its segment end is changed to the frame before the potential start time. In other words, step 932 includes extending the reference segment forward to the potential start time and possibly shortening the length of the precursor segment at the end thereof to avoid overlapping the two segments.

However, if the check at step 930 indicates that the potential start time is closer than x frames before the beginning of the corresponding segment, it is checked in step 934 whether step 934 is being traversed the first time for that potential start time. If this is not the case here, the processing for this potential start time and the relevant segment ends here, and the processing of the onset detection continues in step 928 for another potential start time or in step 926 for a further envelope.

Otherwise, however, in a step 936, the previous segment start of the segment of interest is virtually moved forward. In doing so, the perceptual spectral values in the perception-related spectrum are looked up, which are located at the virtually shifted segment start times. If the fall of these perceptual spectral values in the perceptual spectrum exceeds a certain value, the frame in which this transgression has taken place is provisionally used as segment start of the reference segment and step 930 is repeated again. If then the potential start time is not more than x frames before the beginning of the corresponding segment determined in step 936, the gap is also closed in step 932, as described above.

The effect of the onset detection and correction 914 is thus that individual segments in the current melody line are changed in their time extent, namely extended forward or shortened back.

The step 914 is then followed by a length segmentation 938. In segmental length segmentation 938, all segments of the melody line which, because of halftone mapping 912, now appear in the melody matrix as horizontal lines lying at semitone frequencies, are scanned through, and those segments from the melody line removed, which are smaller than a predetermined length. For example, segments are removed that are less than 10-14 frames long, and preferably 12 frames and less long, again assuming a frame length of 8ms above or adjusting the numbers of frames accordingly. 12 frames at 8 milliseconds correspond to 96 milliseconds time resolution, which is less than about 1/64 note.

Steps 914 and 938 thus correspond to step 762 Fig. 2 ,

The melody line held in step 938 then consists of a somewhat reduced number of segments having one and the same semitone frequency over a certain number of consecutive frames. These segments are clearly attributable to musical segments. This melody line is then entered in a step 940 corresponding to the above-described step 764 of FIG Fig. 2 corresponds, converted into a notation or into a midi file. In particular, each segment still in the melody line after the length segmentation 938 is examined to find the first frame in the respective segment. This frame then determines the note start time of the note corresponding to that segment. For the note, the note length is then determined from the number of frames over which the corresponding segment extends. The quantized pitch of the note results from the halftone frequency, which is constant in each segment due to step 912.

The MIDI output 914 by means 304 then provides the note sequence, based on which the rhythm means 306 performs the operations described above.

The preceding description with reference to the Fig. 3 - 26 referred to melody recognition in means 304 for the case of polyphonic audio pieces 302. However, it is known that that the audio signals 302 are of a monophonic type, as is the case, for example, in the case of pre-whistling for the generation of ringing tones, as has been described above, can be compared with the procedure of FIG Fig. 3 slightly modified procedure may be preferable in that it avoids errors that may arise in the approach of Fig. 3 due to musical imperfections in the source audio signal 302.

Fig. 27 FIG. 4 shows the alternative mode of operation of device 304, which is similar to that of monophonic audio signals Fig. 3 is preferable, but in principle also for polyphonic audio signals would be applicable.

Until step 782, the procedure is the same Fig. 27 with that of Fig. 3 Therefore, for these steps, the same reference numerals as in the case of Fig. 3 be used.

Unlike the procedure after Fig. 3 after step 782, the procedure follows Fig. 27 a sound separation is performed in step 950. The reason for performing the sound separation in step 950, referring to FIG Fig. 28 can be explained in more detail, with reference to Fig. 29 for a portion of the frequency / time space of the spectrogram of the audio signal, the nature of the spectrogram, as determined by frequency analysis 752, for a predetermined segment 952 of the melody line resulting after general segmentation 782 , as a root and for their overtones illustrated. In other words, in Fig. 29 the exemplary segment 952 has been shifted along the frequency direction f by integer multiples of the respective frequency to determine overtone lines. Fig. 29 now shows only those parts of the reference segment 952 and corresponding overtone lines 954a-g, where the Spectrogram from step 752 has spectral values that exceed an exemplary value.

As can be seen, the amplitude of the fundamental of the reference segment 952 obtained in the general segmentation 782 is consistently above the exemplary value. Only the overtones arranged above indicate an interruption approximately in the middle of the segment. The patency of the root has caused the segment in general segmentation 782 not to split into two notes, although there is likely to be a note boundary at about the middle of segment 952. Errors of this kind occur primarily only in monophonic music, which is why the sound separation only in the case of Fig. 27 is carried out.

The sound separation 950 will now be referred to below Fig. 28 . Fig. 29 and Fig. 30a, b explained in more detail. The tone separation begins at step 958, starting from the melody line obtained in step 782, with the search for that overtone or tone lines 954a-954g along which the spectrogram obtained by frequency analysis 752 has the most dynamic amplitude response. Fig. 30a shows in a graph in which the x-axis of a time axis t and the y-axis of the amplitude or the value of the spectrogram corresponds, such an amplitude characteristic 960 for one of the upper tone lines 954a - 954g. The dynamics for the amplitude curve 960 is determined from the difference between the maximum spectral value of the curve 960 and the minimum value within the curve 960. Fig. 30a will exemplify the amplitude curve of the spectrogram along that harmonic line 450a - 450g, which has the greatest dynamics among all these amplitude curves. At step 958, preferably only the 4th through 15th order overtones are considered.

In a following step 962, the positions in the amplitude curve with the greatest dynamics are then at which a local amplitude minimum falls below a predetermined threshold, identified as potential separation points. This will be in Fig. 30b illustrated. In the exemplary case of Fig. 30a resp. b only falls below the absolute minimum 964, which of course also represents a local minimum, the threshold value, which in Fig. 30b exemplified by dashed line 966. In Fig. 30b Thus, there is only one potential separation point, namely the time or frame at which the minimum 964 is located.

In a step 968, those which are located in a boundary area 970 around the segment start 972 or in a boundary area 974 around the segment end 976 are then sorted out among the possibly multiple separation sites. For the remaining potential separation points, in step 978 the difference between the amplitude minimum at the minimum 964 and the mean of the amplitudes of the local maxima 980 and 982 adjacent to the minimum 964 is formed in the amplitude curve 960. The difference is in Fig. 30b illustrated with a double arrow 984.

In a subsequent step 986, it is checked whether the difference 984 is greater than a predetermined threshold value. Otherwise, in a step 988, the reference segment at the potential separation point or minimum 964 is separated into two segments, one of which is separated from the second Segment start 972 extends to the frame of the minimum 964, and the other between the frame of the minimum 964 and the subsequent frame and the segment end 976. The list of segments is extended accordingly. Another way of separating 988 is to provide a gap between the two emerging segments. For example, in the area in which the amplitude curve 960 is below the threshold value, in Fig. 30b for example, over the time range 990.

Another problem that occurs primarily in monophonic music is that the individual notes are subject to frequency fluctuations that complicate subsequent segmentation. Therefore, subsequent to tone separation 950, tone smoothing is performed in step 992, referring to FIG Fig. 31 and 32 is explained in more detail.

Fig. 32 FIG. 8 shows schematically a segment 994, as it is in the melody line, which results on the sound separation 950. The representation in Fig. 32 is such that in Fig. 32 for each tuple of frequency bin and frame traversed by segment 994, a digit is provided on the corresponding tuple. The assignment of the digit will be referred to below Fig. 31 explained in more detail. As can be seen, segment 994 varies in the exemplary case of FIG Fig. 32 across 4 frequency bins and spans 27 frames.

The purpose of tone smoothing is now to select, among the frequency bins between which segment 994 oscillates, the one which is to be assigned to segment 994 constantly for all frames.

The tone smoothing begins in step 996 with the initialization of a counter variable i to 1. In a subsequent step 998, a counter value z is initialized to 1. The counter variable i has the meaning of the numbering of the frames of the segment 994 from left to right in FIG Fig. 32 , The counter variable z has the meaning of a counter that counts over how many consecutive frames the segment 994 is in one and the same frequency bin. In Fig. 32 For ease of understanding the following steps, the value for z for the individual frames is shown in the form of the numbers representing the course of segment 994 in FIG Fig. 32 represent.

In a step 1000, the counter value z is then accumulated to a sum for the frequency bin of the ith frame of the segment. For each frequency bin in which the segment 994 oscillates, there exists a sum or an accumulation value. In this case, the counter value could be weighted according to a varying exemplary embodiment, such as a factor f (i), where f (i) is a function that increases steadily with i, thus the shares to be totalized at the end of a segment, ie the voice, for example better tuned to the tone, to weight more strongly compared to the transient at the beginning of a note. Below the horizontal timeline is in square brackets in Fig. 32 an example of such a function f (i) is shown, where in Fig. 32 i increases along the time and indicates how many positions a given frame occupies among the frames of the considered segment, and successive values which occupy the function shown by way of example for successive sections, again indicated by small vertical bars along the time axis, with numbers shown in these square brackets. As can be seen, the exemplary weighting function increases with i from 1 to 2.2.

In a step 1002, it is checked if the i-th frame is the last frame of the segment 994. If this is not the case, the counter variable i is incremented in a step 1004, ie it is moved to the next frame. In a subsequent step 1006, it is checked whether the segment 994 in the current frame, ie the ith frame, is in the same frequency bin as it was in the (i-1) th frame. If so, in a step 1008 the counter variable z is incremented, whereupon processing continues again at step 1000. However, if the segment 994 is not in the same frequency bin in the i-th frame and the (i-1) th frame, processing proceeds to 1 with the initialization of the counter variable z in step 998.

Finally, if it is determined in step 1002 that i-th frame is the last frame of segment 994, then for each frequency bin in which segment 994 is located, a sum resulting in Fig. 32 at 1010.

In step 1012, upon determination of the last frame in step 1002, the frequency bin for which the accumulated sum 1010 is greatest is selected. In the exemplary case of Fig. 32 this is the second lowest frequency bin among the four frequency bins in which segment 994 is located. In a step 1014, the reference segment 994 is then smoothed by swapping it to a segment in which each of the frames where the segment 994 was located is assigned the selected frequency bin. The sound smoothing off Fig. 31 is repeated segment by segment for all segments.

In other words, tone smoothing, in other words, serves to equalize the singing and singing of tones from lower or higher frequencies, and accomplishes this by finding a value over the time course of a tone corresponding to the frequency of the settled tone. To determine the frequency value of the oscillating signal, all elements of a frequency band are counted up, after which all the incremented elements of a frequency band, which are located on the note segment, are added up. Then the tone is entered over the time of the note segment in the frequency band with the largest sum.

After tone smoothing 992, a statistical correction 916 is then performed, performing the statistical correction of those Fig. 3 in particular corresponds to step 898. The statistical correction 1016 is followed by a halftone mapping 1018 that corresponds to the halftone mapping 912 Fig. 3 corresponds and also uses a halftone vector determined at a halftone vector detection 1020, which corresponds to that of Fig. 818.

Steps 950, 992, 1016, 1018, and 1020 thus correspond to step 760 Fig. 2 ,

The halftone mapping 1018 is followed by an onset identifier 1022, which essentially corresponds to that of FIG Fig. 3 , step 914, corresponds. Only in step 932 is it preferably prevented that gaps are closed again, or segments imposed by the tone separation 950 are closed again.

The onset detector 1022 is followed by offset detection and correction 1024, which will be described below with reference to FIG FIGS. 33-35 is explained in more detail. In contrast to the onset recognition, the offset detection and correction is used to correct the end of the note. The offset detection 1024 serves to prevent the reverberation of monophonic pieces of music.

In a step 1026 similar to step 916, first the audio signal is filtered with a bandpass filter corresponding to the halftone frequency of the reference segment, whereupon, in a step 1028 corresponding to step 918, the filtered audio signal is two-way rectified. Further, in step 1028, an interpretation of the rectified time signal is performed. This approach is sufficient in the case of offset detection and correction to determine approximately an envelope, thereby eliminating the more complicated step 920 of onset detection.

Fig. 34 shows in a graph in which along the x-axis, the time t is plotted in virtual units and along the y-axis of the amplitude A in virtual units, the interpolated time signal, for example with a Reference numeral 1030 and, for comparison thereto, the envelope, as determined in the Onseterkennung in step 920, with a reference numeral 1032.

In a step 1034, in the time segment 1036 corresponding to a reference segment, a maximum 1040 of the interpolated time signal 1030 is determined, in particular the value of the interpolated time signal 1030 at the maximum 1040. In a step 1042, a potential note end time is then determined as the time wherein the rectified audio signal has dropped in time to the maximum 1040 to a predetermined percentage of the value at the maximum 1040, the percentage in step 1042 being preferably 15%. The potential note end is in Fig. 34 illustrated with a dashed line 1044.

In a subsequent step 1046, it is then checked whether the potential note end 1044 is behind the segment end 1048 in time. If not, as it is in Fig. 34 by way of example, the reference segment is shortened from the time range 1036 to end at the potential note end 1044. However, if the note end is earlier than the end of the segment, as exemplified in Fig. 35 5, it is checked in a step 1050 whether the time interval between potential note end 1044 and segment end 1048 is less than a predetermined percentage of the current segment length a, the predetermined percentage preferably being 25% in step 1050. If the result of the check 1050 is positive, an extension 1051 of the reference segment of length a takes place, in order to end now at the potential note end 1044. However, to prevent overlap with the subsequent segment, step 1051 may also be contingent upon an impending overlap, in order not to be performed in this case, or just up to the beginning of the successor segment, possibly with a certain distance therefrom.

However, if the check in step 1050 is negative, no offset correction is performed and step 1034 and the following steps are repeated for another reference segment of equal semitone frequency or step 1026 for other semitone frequencies is continued.

After offset detection 1024, step 1052 is followed by step 938 Fig. 3 corresponding length segmentation 1052 is performed, followed by a MIDI output 1054 following step 940 Fig. 3 equivalent. Step 762 off Fig. 2 correspond to steps 1022, 1024 and 1052.

Referring to the previous description of Fig. 3 - 35 is still pointed to the following. The two alternative melody extraction approaches presented herein include various aspects that need not all be included simultaneously in an effective melody extraction approach. It should first be pointed out that, in principle, steps 770-774 could also be combined with each other by converting the spectral values of the spectrogram from the frequency analysis 752 into the perceptual spectral values using only a single lookup in a lookup table.

In principle, of course, it would also be possible to omit steps 770-774 or just steps 772 and 774, but this should lead to a deterioration of the melody line determination in step 780 and thus to a deterioration in the overall result of the melody extraction method as a whole.

In basic frequency determination 776, a tone model from Goto was used. However, other sound models or other weightings of the overtone components would also be possible and could be adapted, for example, to the origin or origin of the audio signal, as far as this or these is known, such as when in the embodiment of the ringtone generation, the user is set to a Vorsumsmen.

With regard to the determination of the potential melody line in step 780, it is pointed out that although the above-mentioned musicological statement states that for each frame only the fundamental frequency of the loudest sound component had been selected, it is still possible to choose not only one to narrow down the clear selection of the largest share for each frame. As with Paiva, for example, the determination of the potential melody line 780 could include assigning multiple frequency bins to the same frame. Subsequently, a finding of multiple trajectories could be performed. This means allowing a selection of multiple fundamental frequencies or multiple sounds for each frame. Of course, the subsequent segmentation would then have to be carried out differently and, in particular, the subsequent segmentation would be somewhat more complicated since several trajectories or segments would have to be considered and found. Conversely, in this case, some of the above-mentioned steps or sub-steps in segmentation could also be adopted for this case of determining trajectories that may overlap in time. In particular, the steps 786, 796 and 804 from the general segmentation could easily be applied to this case as well. Step 806 could be transferred to the case where the melody line consists of time-overlapping trajectories, if this step took place after the trajectories were identified. The identification of trajectories could be similar to step 810, but modifications would have to be made such that multiple trajectories overlapping in time can also be tracked. The gap closure could also be carried out in a similar way for those trajectories between which there is no gap in time. That too Harmoniemapping could be done between two temporally consecutive trajectories. Vibrato detection and / or vibrato compensation could be readily applied to a single trajectory as well as to the previously mentioned non-overlapping melody line segments. Onset detection and correction could also be applied to trajectories. The same applies to tone separation and tone smoothing as well as offset detection and correction as well as statistical correction and length segmentation. However, allowing for the temporal overlap of trajectories of the melody line in the determination in step 780 at least required that the time overlap of trajectories be sometime eliminated before the actual note sequence output. The advantage of determining the potential melody line in the manner described in the foregoing, with reference to FIG Fig. 3 and 27 is that the number of segments to be examined after the general segmentation is limited in advance to the most essential, and that even the melody line determination itself in step 780 is extremely simple and yet leads to a good melody extraction or transcription.

The general segmentation implementation described above does not have to have all substeps 786, 796, 804, and 806, but may also include a selection thereof.

In gap closure, the perceptual spectrum was used in steps 840 and 842. In principle, however, it is also possible to use in these steps the logarithmic spectrum or the spectrogram obtained directly from the frequency analysis, but the use of the perceptual spectrum in these steps has given the best result in terms of melody extraction. The same applies to step 870 from the harmony mapping.

With regard to harmoniemapping, it is pointed out that it could be provided there, during the displacement 868 of the successor segment, to apply the displacement just in the direction of the melody centerline line, so that the second condition could be omitted in step 874. Referring to step 872, it is to be noted that uniqueness among the selection of the various octave, fifth, and / or third lines could be achieved by creating a priority ranking among them, e.g. Octave line before fifth line before third line and below lines of the same line type (octave, quint, or third line) that is closer to the original position of the successor segment.

With regard to the onset detection and the offset detection, it is pointed out that the determination of the envelope or of the interpolated time signal used instead in the case of offset detection could also be carried out differently. It is only essential that in the onset and offset detection the audio signal is used back, which is filtered by a bandpass filter with a transmission characteristic around the respective semitone frequency, to the rise of the envelope of the resulting filtered signal the note start time or on the basis of Drop of the envelope to recognize the end of the note.

Regarding the flowcharts under the FIGS. 8-41 It should be noted that they show the operation of the melody extractor 304 and that each of the steps represented by a block in these flowcharts may be implemented in a corresponding subunit of the device 304. The implementation of the individual steps can be implemented in hardware, as an ASIC circuit part, or in software, as a subroutine. In particular, in these figures, the inscriptions inscribed in the blocks roughly indicate which ones The respective step corresponds to the respective block, while the arrows between the blocks illustrate the sequence of steps in the operation of the device 304.

In particular, it should be noted that, depending on the circumstances, the inventive scheme can also be implemented in software. The implementation may be on a digital storage medium, in particular a floppy disk or a CD with electronically readable control signals, which may cooperate with a programmable computer system such that the corresponding method is executed. In general, the invention thus also consists in a computer program product with program code stored on a machine-readable carrier for carrying out the method according to the invention when the computer program product runs on a computer. In other words, the invention can thus be realized as a computer program with a program code for carrying out the method when the computer program runs on a computer.

Claims (34)

1. A device for the extraction of a melody underlying an audio signal (302), represented by a note sequence, comprising:

   means (750) for providing a time-over-spectral representation of the audio signal (302);

   means (354; 770, 772, 774) for a frequency-dependent scaling of amplitudes of the time-over-spectral representation using curves of equal volume reflecting the human frequency-dependent volume perception in order to obtain a perception-related time-over-spectral representation; and

   means (756) for determining a melody of the audio signal on the basis of the perception-related time-over-spectral representation.
2. The device according to claim 1, wherein means (750) for providing is implemented such that it provides a time-over-spectral representation comprising for each of a plurality of spectral components a spectral band with a sequence of spectral values.
3. The device according to claim 2, wherein means for frequency-dependent scaling comprises:

   means (770) for logarithmizing the spectral values of the time-over-spectral representation in order to indicate the sonic pressure level, whereby a logarithmized time-over-spectral representation is obtained; and

   means (772) for mapping the logarithmized spectral values of the logarithmized time-over-spectral representation, depending on their respective value and the spectral component to which they belong, to perception-related spectral values in order to obtain the perception-related time-over-spectral representation.
4. The device according to claim 3, wherein means (772) for mapping is implemented in order to perform the mapping based on functions (774) representing the curves of equal volume, associating a logarithmic spectral value to each spectral component indicating a sonic pressure level, and are associated with different volumes.
5. The device according to claim 3 or 4, wherein means (750) for providing is implemented such that the time-over-spectral representation in every spectral band comprises a spectral value for every time section of a sequence of time sections of the audio signal.
6. The device according to claim 5, wherein means (756) for determining is implemented to
   delogarithmize (776) the spectral values of the perception-related spectrum, in order to obtain a delogarithmized perception-related spectrum with delogarithmized perception-related spectral values,
   sum up (776), for each time section and for each spectral component, the delogarithmized perception-related spectral value of the respective spectral component and the delogarithmized perception-related spectral value of those spectral components representing a partial tone to the respective spectral component, in order to obtain a spectral sound value, whereby a time-over-sound illustration is obtained, and
   generate (780) a melody line by uniquely allocating the spectral component to each time section for which the summing-up for the corresponding time section results in the greatest spectral sound value.
7. The device according to claim 6, wherein means for determining is implemented to differently weight the delogarithmized perception-related spectral values of the respective spectral components and that of those spectral components illustrating a partial tone to the respective spectral component in the summing-ups (780), so that the delogarithmized perception-related spectral values of partial tones of higher order are weighted less.
8. The device according to claim 6 or 7, wherein means for determining comprises:

   means (782, 816, 818, 850, 876, 898, 912, 914, 938; 782, 950, 992, 1016, 1018, 1020, 1022, 1024. 1052) for segmenting the melody line (784) in order to obtain segments.
9. The device according to claim 8, wherein means for segmenting is implemented in order to prefilter (786) the melody line in a state, as which the melody line is indicated in binary form in a melody matrix of matrix position which is spanned by the spectral components on the one side and the time sections on the other side.
10. The device according to claim 9, wherein means for segmenting is implemented in order to sum up the entry into this and neighboring matrix positions for each matrix position (792) when prefiltering (786), compare the resulting information value to a threshold value and enter the comparative result at a corresponding matrix position in an intermediate matrix and subsequently multiply the melody matrix and the intermediate matrix in order to obtain the melody line in a prefiltered form.
11. The device according to one of claims 7 to 10, wherein means for segmenting is implemented to leave a part of the melody line unconsidered (796), which is outside a predetermined spectral value (798, 800), during a subsequent part of the segmentation.
12. The device according to claim 11, wherein means for segmenting is implemented such that the predetermined spectral range reaches from 50 - 200 Hz to 1000 - 1200 Hz.
13. The device according to one of claims 8 to 12, wherein means for segmenting is implemented to leave a part of a melody line unconsidered (804) in a subsequent part of the segmentation at which the logarithmized time-over-spectral representation comprises logarithmized spectral values which are less than a predetermined percentage of the maximum logarithmized spectral value of the logarithmized time-over-spectral representation.
14. The device according to one of claims 8 to 13, wherein means for segmenting is implemented in order to leave parts of the melody line unconsidered (806) in a subsequent part of the segmentation at which, according to the melody line, less than a predetermined number of spectral components associated with neighboring time sections have a distance to each other which is smaller than a semitone distance.
15. The device according to one of claims 11 to 14, wherein means for segmenting is implemented in order to division the melody line (812) reduced by the unconsidered parts into segments (812a, 812b) such that the number of the segments is as small as possible and neighboring time sections of a segment are associated with spectral components according to the melody line whose distance is smaller than a predetermined measure.
16. The device according to claim 15, wherein means for segmenting is implemented to
    close (816) a gap (832) between neighboring segments (12a, 812b), in order to obtain a segment from the neighboring segments when the gap is smaller than a first number of time sections (830), and when with the time sections of the neighboring segments (12a, 812b) which are closest to the respective other one of the neighboring segments spectral components are associated by the melody line, which are in a same semitone area (838) or in adjacent semitone areas (836),
    to only close the gap (836) in the case that the same is greater than or equal to the first number of time sections but smaller than a second number of time sections which is larger than the first number (834), when
    spectral components are associated with the time sections of the neighboring segments (812a, 812b), which are closest to the respective other one of the neighboring segments, by the melody line, which lie in the same semitone area (838) or in adjacent semitone areas (836),
    the perception-related spectral values at those time sections are different (840) by less than a predetermined threshold value; and
    an average value of all perception-related spectral values along a connecting line (844) between the neighboring segments (812a, 812b) is greater than or equal to the average values of the perception-related spectral values along the two neighboring segments (842).
17. The device according to claim 16, wherein means for segmenting is implemented in order to determine those spectral components (826) within the scope of the segmentation which are associated with the time sections according to the melody line most frequently, and to determine (824) a set of semitones relative to this spectral component, which are separated from each other by semitone boundaries which in turn define the semitone areas (828).
18. The device according to claim 16 or 17, wherein means for segmenting is implemented to
    perform the closing of the gap by means of a straight connecting line (844).
19. The device according to one of claims 15 to 18, wherein means for segmenting is implemented to
    temporarily shift (868) a follower segment (852b) of the segments which is directly neighboring (864) to a reference segment (852a) of the segments without a time section lying in between, in the spectrum direction in order to obtain a line of an octave, fifth and/or third;
    select (872) one or none of the line of the octave, fifth and/or third depending on whether a minimum among the perception-related spectral values along the reference segment (852a) has a predetermined relation to a minimum among the perception-related spectral values along the line of the octave, fifth and/or third; and
    if the line of the octave, fifth and/or third is selected, shift (874) the follower segment finally onto the selected line of the octave, fifth and/or third.
20. The device according to one of claims 15 to 19, wherein means for segmenting is implemented to
    determine all local extremes (882) of the melody line in a predetermined segment (878);
    determine a sequence of neighboring extremes among the determined extremes for which all neighboring extremes are arranged at spectral components which are less than a first predetermined measure (886) separate from each other and at time sections which are separate from each other by less than a second predetermined measure (890), and
    change the predetermined segment (878) so that the time sections of the sequence of extremes and the time sections between the sequence of extremes are associated with (894) the average value of the spectral components of the melody line at these time sections.
21. The device according to one of claims 15 to 20, wherein means for segmenting is implemented in order to determine the spectral component (832) within the scope of the segmentation which is associated most frequently to the time sections according to the melody line and to determine a set of semitones relative to this spectral component (832) which are separated from each other by semitone boundaries which in turn define the semitone areas, and wherein means for segmenting is implemented to
    change (912) for each time section in each segment the spectral component associated with the same to a semitone of the set of semitones.
22. The device according to claim 21, wherein means for segmenting is implemented in order to perform the change to the semitones such that this semitone among the set of semitones comes closest to the spectral component to be changed.
23. The device according to claim 21 or 22, wherein means for segmenting is implemented to
    filter the audio signal with a band pass filter (916) comprising a transmission characteristic around the common semitone of a predetermined segment in order to obtain a filtered audio signal (922);
    examine (918, 920, 926) the filtered audio signal (922) in order to determine at which points of time an envelope (924) of the filtered audio signal (922) comprises inflection points, wherein these points of time represent candidate initial points of time,
    depending on whether a predetermined candidate initial point of time is less than a predetermined time period before the first segment (928, 930), elongate the predetermined segment to the front by one or several further time sections (932), in order to obtain an elongated segment which ends approximately at the predetermined candidate initial point of time.
24. The device according to claim 23, wherein means for segmenting is implemented in order to shorten a preceding segment to the front when elongating (932) the predetermined segment, when by this an overlapping of the segments across one or several time sections is prevented.
25. The device according to claim 23 or 24, wherein means for segmenting is implemented to
    depending on whether the predetermined candidate initial point of time is more than the first predetermined time duration before the first time section of the predetermined segment (930), trace in the perception-related time-over-spectral representation the perception-related spectral values along an elongation of the predetermined segment in the direction of the candidate initial point of time up to a virtual point of time where the same decrease by more than a predetermined gradient (936), and to then, depending on whether the predetermined candidate initial point of time is more than the first predetermined time duration before the virtual point of time, elongate (932) the predetermined segment to the front by one or several further time sections in order to obtain the elongated segment which approximately ends at the predetermined candidate initial point of time.
26. The device according to one of claims 23 to 25, wherein means for segmenting is implemented to discard segments (938) which are shorter than a predetermined number of time sections after performing the filtering, the determination and the supplementation.
27. The device according to one of claims 1 to 26, further comprising means (940) for converting the segments into notes, wherein means for converting is implemented in order to allocate to each segment a note initial point of time which corresponds to the first time section of the segment, a note duration that corresponds to the number of the time sections of the segment multiplied by a time section time duration, and a tone pitch corresponding to an average of the spectral components through which the segment runs.
28. The device according to one of claims 15 to 27, wherein means for segmenting is implemented to
    determine overtone segments (954a - g) for a predetermined one (952) of the segments,
    determine (958) the tone segment among the overtone segments along which the time-over-spectral representation of the audio signal comprises the greatest dynamic,
    establish (962) a minimum (964) in the course (960) of the time-over-spectral representation along the predetermined overtone segment;
    examine (986) whether the minimum fulfills a predetermined condition, and
    if this is the case, separate (988) a predetermined segment at the time section where the minimum is located into two segments.
29. The device according to claim 28, wherein means for segmenting is implemented to compare (986), in the examination whether the minimum fulfills a predetermined condition, the minimum (964) to an average value of neighboring local maxima (980, 982) of the course (960) of the time-over-spectral representation along the predetermined overtone segment, and perform the separation (988) of the predetermined segment into the two segments depending on the comparison.
30. The device according to one of claims 15 to 29, wherein means for segmenting is implemented to
    for a predetermined segment (994), allocate a number (z) to each time section (i) of the segment, such that for all groups of directly neighboring time sections to which the same spectral component is associated by the melody line, the numbers associated with the directly neighboring time sections are different numbers from 1 up to the number of directly neighboring time sections,
    for every spectral component which is associated with one of the time sections of the predetermined segment, add up (1000) the numbers of those groups, which time sections of the same the respective spectral component is associated with,
    determine (1012) a smoothing spectral component as the spectral component for which the largest up-summing results; and
    change (1014) the segment by associating the certain smoothing spectral component to each time section of the predetermined segment.
31. The device according to one of claims 15 to 30, wherein means for segmenting is implemented to
    filter (1026) the audio signal with a band pass filter comprising a band pass around the common semitone of a predetermined segment in order to obtain a filtered audio signal;
    localize (1034) in an envelope of the filtered audio signal a maximum (1040) in a time window (1036) corresponding to the predetermined segment;
    determine (1042) a potential segment end as the point of time at which the envelope first fell to a value after the maximum (1040) which is smaller than a predetermined threshold value,
    if the potential segment end is (1046) temporally before an actual segment end of the predetermined segment, shorten (1049) the predetermined segment.
32. The device according to claim 31 wherein means for segmenting is implemented to,
    if (1046) the potential segment end is temporally behind the actual segment end of the predetermined segment, elongate (1051) the predetermined segment if the temporal distance between the potential segment end (1044) and the actual segment end (1049) is not greater than a predetermined threshold value (1050).
33. A method for extracting a melody underlying an audio signal (302), represented by a note sequence, comprising:

    providing (750) a time-over-spectral representation of the audio signal (302);

    frequency-dependent scaling (754; 770, 772, 774) of the amplitudes of the time-over-spectral representation using curves of equal volume reflecting human frequency-dependent volume perception in order to obtain a perception-related time-over-spectral representation; and

    on the basis of the perception-related time-over-spectral representation, determining (756) a melody of the audio signal.
34. A computer program having a program code for performing the method according to claim 33, when the computer program runs on a computer.

Images