EP1373908A1

EP1373908A1 - Method and device for analysing a digital audio signal

Info

Publication number: EP1373908A1
Application number: EP02720088A
Authority: EP
Inventors: Sacha c/o Cabinet Hautier Vrazic
Original assignee: Sigtone SA
Current assignee: Sigtone SA
Priority date: 2001-04-05
Filing date: 2002-03-29
Publication date: 2004-01-02
Also published as: FR2823386B1; FR2823386A1; WO2002082106A1

Abstract

The invention concerns a device and a method for analysing a digital audio signal (C) for use for identifying musical notes present in a sound, audio or musical signal. The invention is characterised in that the method consists in: creating filters (7) having different bandwidths each corresponding to a frequency interval including the frequency (fi) of a musical note to be identified, the set covering the frequency range to be analysed; filtering the digital audio signal (C) using the created filters arranged in a filter bank (5) to break down the digital audio signal (C) into as many output signals as there are filters (7) in the bank (5), each of the output signal (Fi) corresponding to a musical note to be identified so as to produce a projection of the digital audio signal (C) into a time-frequency base.

Description

Method and device for analyzing a digital audio signal

The present invention relates to a method and a device for the analysis of a digital audio signal usable for the recognition of musical notes present in a sound, audio or musical signal.

It will find its application particularly in the field of the automatic transcription of sound, audio and musical signals in particular for the construction of partitions when a musician plays his instrument, but also the indexing and the automatic recognition of audio signals.

It also applicable to all types of signals from discs, computer files or any other data medium, but also to signals from a real-time acquisition (via microphone or other sensor). Document FR-A-2,648,566 discloses a method and a device for measuring frequency peaks contained in the spectrum of an analog signal, by threshold comparator.

According to this document, the instants of passage of the signal are determined and stored by a given threshold value using a threshold comparator, a clock and a random access memory.

By calculation, the location of the frequency peaks contained in the spectrum of the analyzed signal is determined.

In general, according to the state of the art, a filter bank is produced for a very wide frequency range, for example a bandwidth of a octave or higher. Musical notes are located by detecting their frequency, using the so-called PITCH TRACKING method.

This technique has the problem of having a low reliability in particular by the fact that the additional step of PITCH TRACKING is present. The present invention overcomes the drawbacks of the techniques known up to now, and has the advantage for this of presenting a method and a device provided with high reliability.

Indeed, we obtain a better precision in the location of the notes, both in space and in frequency, by the fact that we associate a signal with each note whose time evolution is known. No ambiguity can remain.

Another advantage of the invention is to be adaptable to any type of musical instrument, since the frequency range to be analyzed can be adapted according to the range of the instrument. This implementation is also done in real time.

Other objects and advantages will appear during the description which follows which presents a detailed embodiment of the invention which is not, however, limiting.

The present invention relates to a method for analyzing a digital audio signal usable for recognizing musical notes present in a sound, audio or musical signal, characterized by the fact that: - filters with bandwidth widths are created distinct, each corresponding to a frequency interval including the frequency of a single musical note to be recognized, the whole covering the frequency range to be analyzed;

- the digital audio signal is filtered using the filters created organized in a bank to break down the digital audio signal into as many output signals as there are filters present in the bank, each of the output signals corresponding to a musical note to be recognized in order to project the digital audio signal in a time-frequency basis. According to preferred variants of this process:

- the bandwidth of each filter corresponds to a frequency interval symmetrical with the frequency of the musical note to be recognized.

- a preliminary filtering of the digital audio signal is carried out by a finite impulse response filter - a digital audio signal is sub-sampled to reduce its bit rate, the sub-sampling being a function of the frequency range to be analyzed.

- the digital audio signal is obtained by acquisition of a sound signal by microphone and sampling. - the group time of each filter is compensated by a pure delay filter of suitable value.

The invention also relates to a device for analyzing a digital audio signal usable for recognizing musical notes present in a sound, audio or musical signal, characterized in that it comprises a bank of filters in each of which the digital audio signal is injected at the input, said filters having different bandwidths but covering the entire frequency range to be analyzed, each bandwidth including the frequency of a single musical note, to decompose the audio signal digital in as many output signals as musical notes to be recognized, in order to carry out a projection of the digital audio signal in a time-frequency basis.

This device can be presented according to the embodiments introduced below:

- It includes a quantizer for determining the width of the passband of each filter in the bench so that each filter corresponds to one and only one musical note.

- the quantizer is made up of means for calculating the frequency of each musical note in the frequency range to be analyzed and frequency thresholds separating two consecutive notes to assign to each filter a bandwidth between two consecutive frequency thresholds and including the frequency of a note

it includes a subsampler for the prior reduction of the audio-digital signal bit rate, the subsampling being a function of the frequency range to be analyzed,

- it includes a finite impulse response filter for preliminary filtering of the digital audio signal, before subsampling.

- it includes a microphone and a sampler for the analysis of sound signals.

- It includes a group time compensator for each filter consisting of a pure delay filter of suitable value. - the bench filters are Chebychev Type I filters of suitable order with which the digital equivalent is associated.

The attached drawings are given as indicative and non-limiting examples. They represent an embodiment. They will allow the invention to be easily understood.

Figure 1 is a block diagram of the device according to the invention in a particular embodiment.

FIG. 2 illustrates the distribution of the passbands of the bank filters in the frequency range to be analyzed. FIG. 3 shows the frequency representation of the filtering device according to the invention.

FIG. 4 shows an example of an audio signal evolving over time, then, successively, two output signals corresponding to two filtering channels each revealing a musical note. With reference to FIG. 1, the method according to the invention begins with an acquisition of a sound signal A by means of a microphone 1, then with a digitization of the audio electrical signal B thus obtained by means of a sampler 2 .

The sampler 2 delivers a digital audio signal C. Without departing from the scope of the invention, the digital signal C can be obtained from other sources and in particular from a compact digital disk, a computer file or any other source. .

The sampling frequency of sampler 2 may be the standard frequency for audio compact discs, that is to say 44.1 kHz.

Advantageously, the digital audio signal C is then filtered using a finite impulse response filter constituting anti-aliasing filtering (known by the term anti-aliasing).

This first processing of the signal C is followed by the filter 3 by a sub-sampling at the level of a sub-sampler marked 4 in FIG. 1.

The purpose of sub-sampling is to reduce the bit rate of the signal to be processed and its parameters depend on the range of the musical instrument or the range of frequencies to be analyzed, in order to respect the frequency of NYQUIST.

Likewise, the impulse response of filter 3 depends on the sub-sampling rate used, and therefore on the range of frequencies to be analyzed. For the remainder of the description, E is called the signal coming from the sub-sampler 4 which is the working signal usable for the step of filtering by filter bank which will be described below.

In fact, the working signal E is then processed by a filter bank 5 composed by a plurality of filters 7 whose bandwidth is distinct and predetermined, and adapted to cover the entire frequency range to be analyzed.

In the preferred application for recognizing musical notes, the frequency range to be analyzed corresponds to the range of the instrument. As for the passband widths of each filter 7, they are determined to include the frequency fi of a note without interfering with the frequency of other notes. Thus, each filter 7 is associated with a note in the frequency range to be analyzed.

As shown diagrammatically in FIG. 1, the working signal E is decomposed or projected through the filter bank 5, in order to obtain output signals Fi each revealing the time and frequency location of a musical note.

As indicated above, the selectivity of each filter 7 is important and is achieved by determining the order of each of the filters 7. For a musical analysis, the bandwidth of each filter 7 is determined by a chromatic quantizer 6 capable of fixing the terminals of the bandwidth of the filter 7. The bandwidth of a filter 7 includes the frequency fi of the musical note associated with it. It is constructed by a quarter-tone interval on either side of the frequency fi for recognition of 12 notes per octave. More specifically, Figure 2 gives an illustration of bandwidths as a function of frequency. The thresholds of each filter are determined by arithmetic mean of the frequencies of two successive notes.

The quantizer performs this operation. Indeed, we can easily calculate the frequency of each musical note according to the geometric sequence whose progression is as follows: ι π + i ⁼ X 2 The thresholds are then defined by: fi + fi + l

Yi =

2 where yι is a threshold of rank i and fj and f _{i + 1} the frequencies of two successive notes of rank i and i + 1. The quantizer determines the thresholds up to the maximum frequency of the frequency range to be analyzed.

An example is given below of a table revealing the frequency calculation of the notes for two octaves.

These two octaves include 24 notes requiring the calculation of the parameters of 24 filters whose bandwidths are obtained by the chromatic quantizer 6.

As many signals Fi are obtained as there are notes to be analyzed, that is to say 24 in the present case.

Preferably, the filters 7 are CHEBYCHEV type I filters of order 10. The order of the filter is adapted as a function of the instrument and of the sub-sampling rate. This type of filter is by analog origin and it is therefore necessary to calculate its digital equivalent for a software implementation, which does not exclude 0 a hardware implementation.

This equivalent is obtained, after calculation of the poles of the analog transfer function of the filter then, by calculation of the discrete equivalent filter by bilinear transformation.

We obtain the digital equivalent filter of the form as defined below:

In this formula, the index j indicates the channel number, that is to say the note to be analyzed and a and b are the coefficients of the filters.

By way of illustration, an example is given below of the coefficients calculated 0 for this discrete equivalent filter for recognition of notes over two octaves.

By processing the working signal by means of the filters 7 thus defined, a decomposition of the digital audio signal C is obtained into a wavelet packet, each channel obtained corresponding to a musical note.

Given that the bandwidth of the filters 7 is not uniform, a separate group time is present at the output. This group time is not uniform, it is therefore necessary to compensate for this disparity of delay in order to obtain a phase catch-up at the output.

This delay compensation is carried out by means of a delay compensator filter 80 positioned after the filter bank 5.

Of course, the value of the pure delay filter 8 corresponding to each filter 7 is determined as a function of the theoretical group time of the filter 7.

The overall transfer function obtained is as follows: H; = _ where the term d, = max delay c. is the estimated delay to be inserted in channel i (i is the analysis channel for the nth note). η is the estimated group time for channel i.

An example of the group time is given below for a filter bank of 24 filters, in milliseconds

An example of frequency representation of the device according to the invention is given in FIG. 3.

The first graph is an illustration of the evolution of the amplitude of the signal (in decibel) as a function of frequency. The second graph is an illustration of the evolution of the phase (in degrees) as a function of frequency.

As indicated above, the number and the parameters of the filters 7 are adapted to the frequency range to be analyzed.

Thus, for larger range instruments, a larger number of filters and output signals Fi is necessary.

Real-time implementation is possible using the overlap save method by FOURRIER transform function which convolves each impulse response by the signal.

REFERENCES

A. sound signal

B. electrical audio signal C. digital audio signal

D. filtered signal

E. working signal Fj. fi output signals. Frequency of a note 1. microphone

2. sampler

3. filter

4. subsampler

5. filter bank 6. quantifier

7. filter

8. group time compensator filter

Claims

1. Method for analyzing a digital audio signal (C) usable for recognizing musical notes present in a sound, audio or musical signal, characterized in that:

- Filters (7) are created having distinct bandwidths each corresponding to a frequency interval including the frequency (fi) of a single musical note to be recognized, the assembly covering the frequency range to be analyzed;

the digital audio signal (C) is filtered using the filters created organized in a bank (5) to decompose the digital audio signal (C) into as many output signals as there are filters (7) present in the bank (5), each of the output signals (Fi) corresponding to a musical note to be recognized in order to carry out a projection of the digital audio signal (C) in a time-frequency basis.

2. Method according to claim 1, characterized in that the bandwidth of each filter (7) corresponds to a frequency interval symmetrical around the frequency (fi) of the musical note to be recognized.

3. Method according to claim 1 or 2, characterized in that a preliminary filtering of the digital audio signal (C) is carried out by a filter (3) with finite impulse response, and that a signal sub-sampling is carried out digital audio (C) to reduce its bit rate, the sub-sampling being a function of the frequency range to be analyzed.

4. Method according to any one of claims 1 to 3, characterized in that the digital audio signal (C) is obtained by acquisition of a sound signal (A) by microphone (1) and sampling.

5. Method according to any one of claims 1 to 4, characterized in that the group time of each filter (7) is compensated by a pure delay filter (8) of suitable value.

6. Device for analyzing a digital audio signal (C) usable for the recognition of musical notes present in a sound, audio or musical signal, characterized in that it comprises a bank (5) of filters (7 ) into each of which the digital audio signal (C) is injected as an input, said filters (7) having different bandwidths but covering the entire frequency range to be analyzed, each bandwidth including frequency (fi ) a single musical note, to decompose the digital audio signal (C) into as many output signals (Fi) as musical notes (fi) to recognize, in order to produce a projection of the digital audio signal (C) in a time-frequency basis.

7. Device according to claim 6, characterized in that it comprises a quantizer (6) for determining the width of the passband of each filter (7) of the bench (5) so that at each filter ( 7) corresponds to one and only one musical note.

8. Device according to claim 7, characterized in that the quantizer (6) consists of means for calculating the frequency (fi) of each musical note in the frequency range to be analyzed and frequency thresholds separating two consecutive notes to assign to each filter (7) a bandwidth between two consecutive frequency thresholds and including the frequency (fi) of a note.

9. Device according to claim 7 or 8, characterized in that it comprises a subsampler (4) for the prior reduction of the bit rate of the audio-digital signal (C), the subsampling being a function of the range frequency to analyze, and that it comprises a finite impulse response filter (3) for preliminary filtering of the digital audio signal (C), before the subsampling.

10. Device according to any one of claims 8 to 9, characterized in that it comprises a microphone (1) and a sampler (2) for the analysis of sound signals (A).

11. Device according to any one of claims 8 to 10, characterized in that it comprises a compensator for the group time of each filter (7) constituted by a pure delay filter (8) of suitable value.

12. Device according to any one of claims 8 to 11, characterized in that the filters (7) of the bench (5) are Chebychev Type I filters of suitable order with which the digital equivalent is associated.