EP4134954B1 - Method and device for improving an audio signal - Google Patents

Description

The present invention relates to a method for enhancing an audio signal. The method is preferably carried out in real time, so that it is suitable for essentially simultaneous recording and playback of audio signals.

In practice, audio signals are often recorded under unfavorable acoustic conditions using microphones. For example, a desired speech signal component is overlaid by unwanted background noise during the recording, which impairs the quality of the audio signal, particularly with regard to speech intelligibility. In addition, the audio signal can be reverberated due to the spatial conditions or as a result of a large distance between the speaker and the microphone, so that the speech component of the audio signal is difficult to understand when played back over loudspeakers despite amplification. For this reason, the actual advantage of acoustic amplification of the audio signal is often not sufficient to ensure satisfactory speech signal quality and speech intelligibility.

In order to reduce the problems mentioned, it is basically possible to process the audio signal after recording using an audio filter in order to reduce undesired signal components. However, this is associated with difficulties because the audio filter must be matched to the respective audio signal. In practice, this means that an audio filter can achieve good results for a specific audio signal recorded in a specific acoustic environment with a specific microphone, for a different audio signal, which has been recorded under other conditions, but not.

The problems mentioned above are particularly relevant in the field of mobile audio technology used, for example, for trade fairs, because it has to be compatible with a wide variety of acoustic environments and, as a rule, little time is available to optimally adjust the audio processing devices. In addition, there is often no way at all to optimize the audio devices for a particular speaker, for example with regard to the appropriate distance between the speaker and the microphone. In addition, differences between different speakers cause problems. For example, different speakers, who have different voice characteristics (e.g. different speaker volume and frequency composition), in particular due to age and gender differences, cannot be treated with the same audio devices with a constant configuration in such a way that a high voice signal quality is reliably achieved.

It is possible to use a mixer to filter the audio signal and set the acoustic filter parameters manually during recording. However, this is expensive and requires specially trained personnel. In addition, the improvements that can be achieved in this way are variable. Problems exist in particular in the case of strongly changing acoustic recording situations that cannot be compensated for with sufficient speed and reliability.

Methods for improving the voice signal quality are from the documents U.S. 2016 0 019 905 A1 , U.S. 2017 0 047 080 A1 , U.S. 6,295,364 B1 , US 2010 012 1 634 A1 , US 2006 024 7 922 A1 and Schepker etc al., Improving speech intelligibility in noise by SII-dependent preprocessing using frequency-dependent amplification and dynamic range compression, Interspeech 2013.

It is an object of the invention to provide a method for enhancing audio signals which is suitable for different audio signals and in particular a reliable automatic one Enhancing the audio signal in real time. Furthermore, it is an object of the invention to provide a device for enhancing audio signals, which is suitable for automatically enhancing different audio signals, in particular in real time.

According to a first aspect, the object is achieved by a method having the features of claim 1.

A method according to the invention for improving an audio signal has at least the following steps: receiving an audio signal with a plurality of amplitude values, the audio signal having speech at least in sections; detecting speech portions of the audio signal; filtering the audio signal with at least one level filter to reduce signal level variations of the audio signal in the detected speech sections; and filtering the audio signal with at least one equalization filter to reduce spectral variations of the audio signal in the detected speech segments.

The method also includes the following steps: determining a feedback frequency, which represents a feedback of the audio signal; Filtering the audio signal with a feedback filter on the basis of the determined feedback frequency in order to reduce spectral components of the audio signal that represent feedback.

Filtering with the at least one equalization filter includes a step of determining coarse spectral values on the basis of fine spectral values of the audio signal, the coarse spectral values representing the fine spectral values with a lower spectral resolution than the fine spectral values. Furthermore, first equalization weights are determined, which indicate a deviation of the coarse spectral values from predetermined represent reference spectral values. The audio signal is also weighted with the first equalization weights to bring the spectral values into agreement with the reference spectral values.

Determining the feedback frequency includes the steps of: determining a subset of spectral values of the audio signal that violate a predetermined spectral threshold; determining a plurality of first spectral parameter values on the basis of the subset, each of the first spectral parameter values representing a predetermined relation between an associated spectral value of the subset and at least one temporally and/or spectrally adjacent spectral value; and determining the feedback frequency based on the plurality of first spectral parameter values.

It has been shown that the quality of audio signals suffers particularly from insufficient intelligibility of the speech components contained, ie in particular those sections of the audio signal which contain spoken language. Against this background, according to the invention, time sections of the audio signal are detected which contain speech and can be referred to as speech sections. Based on the detected portions, the audio signal is then processed with a level filter and an equalization filter to reduce certain variations in the audio signal. Here, variations can be treated both within an audio signal and between different audio signals.

The level filter is used to reduce signal level variations in order to standardize the level of the audio signal. For example, very loud and quiet parts of the speech signal are attenuated or amplified in sections, resulting in a uniform signal level overall adjusts In practice, different signal levels result, for example, from variable distances between a speaker and the recording microphone and from the acoustic properties of the surrounding room. However, the resulting level variations are compensated by the level filter, so that the subjective signal quality improves.

Additionally or alternatively, an equalization filter is used to reduce spectral variations in the audio signal. On the one hand, spectral variations occur due to different speakers, who impress their own spectral characteristics on the audio signal with their voices. In addition, there is a spectral coloring due to the acoustic environment during the recording and possibly due to the sound equipment used, in particular the microphone and its alignment relative to the speaker.

For a high level of speech intelligibility, it is important that the spectral components in certain frequency ranges that are relevant for speech intelligibility are not masked, or only to a small extent, by other spectral components. However, the acoustic environmental conditions often lead to the speech-relevant parts being variably superimposed by other signal parts in the same or in neighboring frequency ranges, so that the speech-relevant parts cannot always be perceived equally well. Such changes in the signal can be determined from the spectral variations over time and can therefore be treated by a suitable filter. Against this background, the equalization filter is used to reduce spectral variations in the detected speech sections. In this way, the audio signal can be standardized in spectral terms in order to To increase signal quality, especially with regard to good speech intelligibility.

In particular, the method allows a fully automatic signal improvement to take place. A previous or operational manual setting or readjustment of filter parameters is therefore not necessary, i.e. the parameters of the level and/or equalization filter can be permanently set when the method is carried out as intended or are set automatically by a computing unit. In addition, the process ensures excellent signal improvement for a wide variety of audio signals, even in particularly difficult acoustic environments. In other words, the method is particularly robust against acoustic variations of any kind and is therefore particularly suitable for professional use in practice. In addition, the method can be used in real time, i.e. with a latency of less than 20 ms, preferably less than 10 ms, in particular 6 ms.

It is particularly advantageous if both the level filter and the equalization filter are used. In addition, additional filters can also be provided in order to further improve the audio signal, as will be explained below.

It goes without saying that the filtering of the audio signal does not necessarily have to be restricted to the detected speech sections. For example, an equalization filter can also be effective outside of speech sections with regard to special spectral components that are caused, for example, by feedback. However, the audio signal is filtered at least in the speech sections because these are particularly important for speech intelligibility. To the In order to improve the efficiency of the method, certain aspects of the filtering can be restricted to the speech sections.

Embodiments are disclosed in the description, the figures and the dependent claims.

According to one embodiment, the method includes a step of determining a plurality of spectral values on the basis of the amplitude values, the amplitude values representing the audio signal in a time domain and the spectral values representing the audio signal in a frequency domain. The detection of the speech sections, the filtering with the at least one level filter and/or the filtering with the at least one equalization filter takes place on the basis of the amplitude values and/or the spectral values. The filtering is thus performed on the basis of two different representations of the audio signal, namely time domain and frequency domain values of the audio signal. The efficiency and reliability of the process is increased in this way.

The spectral values can be determined on the basis of the time-domain amplitude values using known frequency-space transforms, such as the Fast Fourier Transform (FFT). The spectral values are preferably formed by the absolute value of the frequency coefficients (spectral amplitude values), which can be determined particularly efficiently by FFT on the basis of the time domain amplitude values. The advantageous use of the amplitude values and the spectral values thus requires comparatively few computer resources.

According to a further embodiment, the detection of the speech sections comprises at least the following steps: determining at least a first energy parameter value on the basis of the amplitude values, wherein the first energy parameter value represents an average energy of a portion of the speech signal; determining at least one second spectral parameter value based on spectral values of the audio signal, the at least one second spectral parameter value representing a harmonic spectral structure of the portion; and detecting the segment as a speech segment if the at least one first energy parameter value violates a first energy parameter threshold and/or the at least one second spectral parameter value violates a spectral parameter threshold. The detection of speech sections based on time domain and spectral parameters has proven to be particularly useful for reliably detecting both noise-like sections (e.g. consonants) and tonal sections (e.g. vowels) and evaluating them by threshold comparison to distinguish between speech sections and noise sections.

The threshold values mentioned (energy parameter threshold value and spectral parameter threshold value) can in principle be permanently set. However, the reliability of the detection of speech sections can be improved in a particular way by adapting the first energy parameter threshold value and/or the first spectral parameter threshold value as a function of time. For example, the signal level of the audio signal can be used to set the thresholds to ensure that the thresholds are aligned with the current energy level.

According to a further embodiment, the filtering of the audio signal with the at least one level filter comprises at least the following: determining at least one level parameter value on the basis of the amplitude values, the level parameter value having a represents mean level of the audio signal for a detected speech segment; determining at least one compensation weight based on the at least one level parameter value; and weighting the audio signal with the at least one compensation weight to reduce the signal level variations of the audio signal.

The at least one level parameter value can generally comprise a plurality of level parameter values which indicate the level for detected speech sections of different lengths. Advantageously, first and second level parameter values can be determined, the first level parameter values representing the mean level of the audio signal with a first time resolution and the second level parameter values representing the mean level of the audio signal with a second time resolution. The first and second time resolution differ from each other. In this way, short-term and long-term effects of human auditory perception can be advantageously taken into account. In particular, brief level peaks (clipping) can be detected by level parameter values with a short time resolution and used for filtering. In addition, moderate level variations, which only become perceptible after a minimum duration, can be recorded using level parameter values with greater time resolution. The compensation weight for the level filter is then determined based on the first and second level parameter values.

The first level parameter values are preferably formed on the basis of a plurality of consecutive energy averages. These can be smoothed to obtain first loudness values that form the first level parameter values. The second level parameter values are preferably formed by second volume values. these can are in turn formed on the basis of a plurality of consecutive energy averages, with a larger number of energy averages being smoothed than the first volume values, so that the second level parameter values each indicate the level for a longer period of time than the first level parameter values. The second time resolution is therefore preferably greater than the first time resolution.

In order to determine loudness values, at least some of the amplitude values are preferably grouped into time segments of the audio signal. Then the loudness values for at least some of the time periods are determined based on the grouped amplitude values, each of the loudness values representing the loudness of one of the time periods of the audio signal.

The terms "energy" and "level" each represent an intensity or magnitude of the amplitude values. Level values can therefore basically be viewed as energy values of the audio signal and vice versa, with a different unit for both values being possible but not mandatory (e.g. the normalized logarithmic unit dB can be provided for the level in contrast to the energy). However, the term "level" creates a functional reference to the level filter in particular. The term "loudness" represents the intensity of the amplitude values considering the auditory perceptibility.

According to a further advantageous embodiment, first compensation weights and second compensation weights are determined, with the first compensation weights being determined in order to reduce signal level variations with at least one level that is greater than a predetermined level threshold value, with the second compensation weights being determined in order to reduce the signal level of the Adjust audio signal to a predetermined value. In this way, on the one hand, excessive level values are treated, which can potentially be perceived as distortion that reduces quality. In addition, the audio signal in the detected speech sections is set to a basic level so that moderate volume fluctuations from the listener's point of view can be compensated. The first compensation weights are preferably determined on the basis of the first level parameter values and the second compensation weights are determined on the basis of the second level parameter values. In this way, the filtering can be carried out in a particularly hearing-friendly manner.

As mentioned above, the filtering with the at least one equalization filter includes a step of determining coarse spectral values on the basis of fine spectral values of the audio signal, the coarse spectral values representing the fine spectral values with a lower spectral resolution than the fine spectral values. Furthermore, first equalization weights are determined, which represent a deviation of the coarse spectral values from predetermined reference spectral values. The audio signal is also weighted with the first equalization weights to bring the spectral values into agreement with the reference spectral values. The fine spectral values are preferably formed by the spectral values mentioned above, which can be efficiently determined in particular by FFT. With a sampling rate of 48 kHz and a block length of 1024, for example, the spectral resolution of these spectral values is significantly higher than the resolution that can be resolved by the human ear. In contrast, the frequency resolution of the coarse spectral values preferably corresponds to the resolution of human hearing, so that on this basis an aurally appropriate equalization is made possible. The reference spectral values used for this purpose represent a reference spectrum for achieving a high voice quality of audio signals. The coarse spectral values can be obtained, for example, by octave band filtering of the fine spectral values.

According to a further embodiment, the filtering with the at least one equalization filter includes a weighting of the audio signal with second equalization weights, the second equalization weights being predetermined. Thus, in addition to or as an alternative to the first equalization weights, second equalization weights can be provided which, in contrast to the first equalization weights, are not determined dynamically but are fixed in advance. The second equalization weights can be used to weaken spectral components, for example, which are always a hindrance to high voice quality and can therefore be given a negative gain factor.

According to a further embodiment, the method includes filtering the audio signal with at least one compressor in order to reduce a dynamic range of the audio signal. For the at least one compressor, a plurality of different sets of parameters can be provided, which are selected as a function of an amount of the audio signal and are used as a basis for filtering with the at least one compressor. The multiple parameter sets can advantageously differ from one another in terms of a degree of compression. For example, the multiple sets of parameters may include a first set of parameters to reduce the dynamic range of the audio signal with a first degree of compression, wherein the multiple sets of parameters include a second set of parameters to reduce the dynamic range of the audio signal with a second degree of compression that is greater than the first degree of compression . For particularly good results, the plurality of parameter sets preferably have a third parameter set in order to match the dynamic range of the audio signal with a third Reduce compression level that is less than the first compression level. In this way, quality-reducing distortions that can be caused by strong compression can be avoided particularly effectively. The compressor can be thought of as a special level filter because a reduction in dynamic range is accompanied by a reduction in level and level variations.

As mentioned above, the method further comprises the following steps: determining a feedback frequency, which represents a feedback of the audio signal; Filtering the audio signal with a feedback filter on the basis of the determined feedback frequency in order to reduce spectral components of the audio signal that represent feedback. The spectral values already present are preferably used to determine the feedback frequency, so that they do not have to be determined again for this purpose. Feedback occurs when reproduced signal components are recorded again by the microphone and amplified, resulting in an unstable system state that is acoustically perceptible through strong resonance, e.g. through humming or a shrill whistling sound. The feedback filter counteracts such coupling effects so that the signal quality is not affected. The feedback filter can be viewed as a special equalization filter.

As mentioned above, determining the feedback frequency preferably comprises the following steps: determining a subset of spectral values of the audio signal that violate a predetermined spectral threshold value; determining a plurality of first spectral parameter values based on the subset, each of the first spectral parameter values having a represents a predetermined relation between an associated spectral value of the subset and at least one temporally and/or spectrally adjacent spectral value; and
determining the feedback frequency based on the plurality of first spectral parameter values. The computing effort for determining the feedback frequency can be greatly reduced by the threshold-based preselection of spectral values, so that the real-time capability of the method is enhanced. A predetermined relation between spectral values can be formed, in particular, by mathematically linking the spectral values, for example by using mathematical operators such as division or addition. In this way, certain properties of the spectrum that are typical of a feedback frequency can be efficiently detected.

It is of particular advantage that if the specific feedback frequency disappears between successive time sections of the audio signal, the effectiveness of the feedback filter is gradually reduced over a number of time sections. In this way, it is ensured in a particularly reliable manner that any feedback frequency is effectively removed from the signal. In addition, any noticeable filter fluctuations are avoided. The step-by-step reduction of the feedback filter preferably takes place according to the scheme of a finite automaton.

According to a further embodiment, a pause filter is provided for filtering the audio signal in order to reduce the audio signal in areas outside of the detected speech sections. In this way, for example, temporal masking effects caused by background noise can be weakened.

Furthermore, the audio signal can be filtered with a noise filter in order to reduce the audio signal in areas with amplitude values that violate a predetermined noise threshold value. In particular, very small amplitude values that are below a threshold value and are irrelevant for good signal quality can be essentially completely removed in this way. This counteracts the occurrence of feedback. A noise gate filter is preferably used.

According to a further embodiment, the audio signal is filtered with a bandpass filter. A lower limit frequency of the bandpass filter is preferably in a range from 50 to 100 Hz. An upper limit frequency of the bandpass filter is preferably in a range from 8000 to 10000 Hz.

The method aspects described above can be stored as instructions in a non-volatile memory. If the instructions are executed by an arithmetic unit, the arithmetic unit is prompted by the instructions to execute the method described according to one embodiment. In general, the method can thus be partially or fully implemented by a computer.

According to a second aspect, the object of the invention is achieved by a device having the features of the independent device claim.

A device according to the invention for improving an audio signal that has speech comprises, on the one hand, at least one input interface for detecting an audio signal. The input interface has a connector for a microphone to capture the audio signal. On the other hand, at least one output interface is provided for outputting the audio signal. The Output interface has a connection for an audio playback device, eg a public address system with one or more sound transducers. The device also has a computing unit for executing a method for improving the audio signal. The method is designed according to one of the preceding embodiments.

The device is preferably designed as a compact audio device, so that it is particularly suitable for mobile use.

The device preferably has a non-volatile memory in which commands for executing the method are stored. For this purpose, the memory can be coupled to the processing unit.

The computing unit preferably includes an analog-to-digital converter and a digital-to-analog converter. The enhancement of the audio signal can thus be based at least in part on a digital version of the audio signal. The method can thus be carried out particularly efficiently on the one hand. On the other hand, a high filtering quality can be guaranteed.

The input or output interface can be implemented as a wired interface in order to ensure compatibility with other professional audio devices and to minimize transmission losses. However, it is also conceivable for the interfaces to be wireless in each case, in which case the interfaces can also be combined to form a common wireless interface for this purpose.

Further embodiments of the device are in the dependent claims, the following description and the drawings described. However, it goes without saying that the method features described can also be implemented in a corresponding manner in the device, in particular by means of a corresponding configuration of the computing unit. Conversely, device features described here can also form part of the process with regard to their function as process features.

According to one embodiment, the device also includes a preamplifier for the audio signal, which can be coupled to the input interface. In this way, the audio signal can advantageously be amplified to a predetermined level range before sampling. A plurality of predetermined amplification values can be provided for the preamplifier, with one of the amplification values preferably being selected automatically or by an operator of the device and the amplification being used as a basis.

The device preferably has an electrical supply for the input interface. This enables electrical supply of a connected sound transducer, e.g. a microphone, via the input interface in the sense of a so-called phantom power supply.

According to a further embodiment, the device also has a switching device which can be coupled to the input interface, the output interface and/or the processing unit in order to optionally connect the input interface to the output interface via the processing unit. In other words, the computing unit can be bypassed. In this way, an output of the audio signal can also be guaranteed in the event of a malfunction of the processing unit.

In order to enable the device to function reliably even in continuous operation, the device is preferably provided with a cooling device. All components of the device, including the processing unit, can thus be accommodated in a compact housing, with e.g.

The computing unit can advantageously have a single-board computer, so that the device can be made particularly compact overall. The device can also have a housing, in which in particular all electrical components of the device can be accommodated in order to be protected from external influences in this way. The computing unit can have one or more processors and a memory in which instructions for executing the method can be stored.

In order to configure the device, the device preferably has at least one external communication interface. For example, the device can be equipped with a network interface, eg an Ethernet interface, or a bus interface to be connected via a network or directly to a user terminal, for example a PC or a mobile terminal such as a laptop. A connection to wireless end devices can also take place via the Internet in order to enable a connection to a central server (cloud). The control interface can also be in the form of a wireless interface, so that the device can be connected directly to a mobile end device (eg via Bluetooth or a local wireless network). Communication with the device, for example for the purpose of configuration, can thus be particularly done comfortably. In particular, control data, for example filter parameters for executing the described method for improving an audio signal, can be set via the communication interface. In particular, this can be done remotely so that configuration by the end user of the device can be avoided entirely. In addition or as an alternative, the communication interface can be designed to transmit the audio signal to a mobile terminal device or a central server. In this way, the audio signal can be stored in the end device or in a cloud, for example for documentation purposes. For transmission to a central server, the communication interface is preferably designed as an Ethernet interface, which also enables transmission of audio signals (eg using Dante, Milan, AES (Advanced Encryption Standard).

In addition, a firmware of the device can be updated via a communication interface of the device. A communication interface is preferably provided in the form of a separate bus interface, which is used in particular to connect a storage medium, e.g. a mass storage device in the form of a USB stick or the like. On the one hand, configuration and/or update data can be stored on the storage medium, which are transmitted to the device in order to update the locally stored data. In addition, the audio signal can be output to the storage medium for recording purposes and stored in the storage medium. For this purpose, the device is preferably equipped with an operating interface in order to be able to control the recording of the audio signal directly on the device.

A non-claimed aspect of the disclosure relates to a method for selectively enhancing a first audio signal using an audio processing means, wherein the first audio signal comprises at least portions of speech and the method comprises at least the following steps: determining whether the audio processing means has a predetermined health status; If the audio processing means has the predetermined fitness state, performing a method of enhancing the first audio signal using the audio processing means to provide a second audio signal; If the audio processing means does not have the predetermined health status, providing the first audio signal. The method thus enables the audio processing means to be used selectively depending on its health status. Malfunctions of the audio processing means therefore do not lead to no audio signal being output and user satisfaction being impaired. In the event of a malfunction, at least the first audio signal is provided, so that a usable audio signal is available, for example for public address systems, and in this way a basic functionality is retained. The method can be implemented in particular by a switching device, which can be implemented in a device, for example as a switchable relay. Alternatively, the switching functionality can also be implemented by the computing unit itself. However, a separate switching device has the advantage of protection against a complete failure of the processing unit, in which no transmission of the signal can take place.

The methods disclosed herein can preferably be carried out with the device described. However, it is also possible to carry out the method in whole or in part on any computer, in particular a central server. For example, the audio signal can be captured locally and transmitted to a server where signal enhancement is performed. The enhanced signal can then be sent to a local receiver for playback with a sound transducer.

Fig. 1

ein Blockdiagramm zur Illustration eines Verfahrens zur Verbesserung eines Audiosignals;

Fig. 2

ein Blockdiagramm zur Illustration eines Verfahrens zum Detektieren von Sprachabschnitten eines Audiosignals;

Fig. 3

Frequenzgänge von Oktavfiltern zur Bestimmung von Grobspektralwerten für ein Entzerrfilter für das Verfahren nach Fig. 1;

Fig. 4

ein Blockdiagramm zur Illustration eines Verfahrens zum Bestimmen einer Rückkopplungsfrequenz;

Fig. 5

ein Blockdiagramm zur Illustration eines schrittweisen Reduzierens eines Rückkopplungsfilters;

Fig. 6

eine schematische Darstellung eines Geräts zur Audiosignalverbesserung;

Fig. 7

eine Anordnung mit dem Gerät von Fig. 7.

The invention is explained in more detail below purely by way of example with reference to the drawings. The drawings show in detail:

1

a block diagram to illustrate a method for enhancing an audio signal;

2

a block diagram for illustrating a method for detecting speech portions of an audio signal;

3

Frequency responses of octave filters for determining coarse spectral values for an equalization filter for the method 1 ;

4

a block diagram to illustrate a method for determining a feedback frequency;

figure 5

a block diagram to illustrate a stepwise reduction of a feedback filter;

6

a schematic representation of a device for audio signal enhancement;

7

an arrangement with the device of 7 .

In the figures, identical or corresponding elements are marked with the same reference symbols.

A method for enhancing an audio signal is described below with reference to 1 described.

An analog audio signal is captured with a microphone (not shown) (step 10), the audio signal having a plurality of speech sections and a plurality of noise sections. The speech sections have speech and form a speech signal component. The noise sections are formed by all other sections that do not have speech, especially in pauses in speaking.

In step 12, the audio signal is pre-amplified, ie electronically amplified as an analog signal with an amplification factor. For a corresponding preamplifier (in 1 not shown), a fixed gain can be set. Alternatively, a user can select one of a number of preset gain values as a function of a recording-related basic level in order to relieve a subsequent level filter for reducing level variations.

The pre-amplified audio signal is converted in step 14 from an analogue signal to a digital signal. This is preferably done using an analog-to-digital converter which samples the analog signal at a predetermined sampling rate, e.g., 48,000 Hz. Alternatively, step 14 can also take place after step 16, which is explained below.

The audio signal is processed with a level filter in step 16 in order to compensate for variations in the signal level. For this purpose, the level filter is operated as a function of first filter data 44, which is based on of the audio signal at the output of the level filter can be determined in step 18. They include first volume values, detected speech sections and detected level peaks. Level peaks are detected signal levels that are greater than a predetermined level threshold value, in which the signal overdrives (clipping).

The volume values are determined for individual blocks of the audio signal, which preferably each have a length of 64 sample values. For each block, a first loudness value is determined by summing the squared sample values of the block and then taking the square root of the sum. So-called RMS values (Root Mean Square) are formed in this way, each of which represents an average energy of the underlying block of sampled values.

The RMS values of several blocks are preferably used for the level filter. To detect level peaks, the RMS values of the current block and the previous block are evaluated together, with a level peak being detected if at least one of the two RMS values exceeds a predetermined threshold value, for example −3 dB. If a level peak is detected, this information is taken into account as part of the first filter data 44 in step 16 . In response to a detected peak in level, the gain of the level filter is reduced sharply and rapidly in step 16, for example at a rate of -3 dB within 200 ms. This effectively removes level peaks. Level peaks are preferably filtered regardless of whether the relevant section of the audio signal is a speech section or not.

The level filter of step 16 is further configured to adjust the level of the audio signal to a predetermined value. For this purpose, the RMS values of the current block and a large number of several previous blocks, for example 30 previous blocks, are used. The RMS values are smoothed across the blocks considered, removing short-term fluctuations that are irrelevant to human perception (except for the level peaks, which are treated separately). The median of the RMS values under consideration is preferably formed for smoothing in order to obtain second volume values which indicate the current signal level in an aurally correct manner. A compensation weight is then determined, which represents the difference between a predetermined reference value and the current second volume value. For example, the current volume value can be subtracted from a reference volume of -20 dB to form a compensation weight. The compensation weight is then weighted, e.g. multiplied, with the audio signal to bring the loudness in line with the reference loudness.

The maximum change in the compensation weight over time is preferably limited, for example to 5 dB per second. This avoids unnatural fluctuations in the volume of the audio signal.

In addition, the adjustment of the signal level with reference to the reference volume is preferably only carried out in those sections of the audio signal which have been detected as speech sections. The information as to which sections have been detected as speech sections is made known to the level filter of step 16 as part of the filter data 44 .

The detection of speech segments takes place in step 18 and is explained below with reference to 2 explained.

Speech sections are detected on the basis of amplitude values 54 and spectral values 56, with the amplitude values 54 representing the audio signal in the time domain and the spectral values 56 representing the audio signal in the frequency domain. The amplitude values 54 are formed by the sample values of the digital audio signal after step 14. The spectral values 56 are determined block by block using fast Fourier transformations (FTP) on the basis of the amplitude values 54 . In principle, however, other frequency transformations can also be used. The block length for determining the spectral values 56 is preferably 1024 amplitude values (sampling values), with adjacent blocks preferably overlapping by half and the relevant amplitude values of each block being weighted with a Hann window before the transformation, in order to avoid unwanted spectral components caused by the block boundaries become to reduce. Furthermore, the spectral values 56 are weighted with a predetermined factor, so that the spectral values 56 are normalized to a range between 0 and 1. The factor depends in particular on the window used. In the case of the preferred Hann window, a factor of 0.00391 can advantageously be used.

In step 58 of 2 three parameters are determined and each checked to determine whether an associated threshold value criterion is violated. A first parameter value is formed by the RMS value described above based on the amplitude values 54 . The first parameter value may also be referred to as Short Time Energy (STE) because it represents the average energy over a relatively short length block of 64 amplitude values. If the first parameter value exceeds an associated threshold (step 62), the first parameter value indicates a speech portion, otherwise a noise (non-speech) portion. High RMS values can be caused in particular by consonants and thus indicate speech.

A second parameter value is determined on the basis of the spectral values 56 and indicates the form of a harmonic overtone structure of the frequency spectrum. In particular, the second parameter value represents a measure of the spectral flatness of the frequency spectrum represented by the spectral values 56 (Spectral Flatness, SF). The second parameter value is preferably determined by dividing the geometric mean of the spectral values 56 and the arithmetic mean of the spectral values 56 . The second parameter value is then compared to an associated threshold (step 62). If the threshold is exceeded, the second parameter value indicates a speech section, otherwise a noise section. High values of the second parameter indicate noise-like blocks that are atypical for speech. In contrast to the first parameter, the second parameter refers to a significantly longer block length of 1024 due to the spectral values, so that the usually significantly shorter consonants are not significant compared to an otherwise tonal characteristic.

A third parameter value is also determined, which indicates whether a maximum of the spectral values 56 lies in a predetermined frequency range. For this purpose, it is preferably determined whether the spectral value, the amount of which forms a maximum compared to the other spectral values 56 of a block (step 58), is in a frequency range between 70 and 250 Hz, ie it is checked whether the maximum spectral value represents a frequency that greater than a lower frequency threshold and is less than an upper frequency threshold (step 62). If true, the third parameter value indicates a speech portion, otherwise a noise portion. The fundamental frequency of speech is generally in the range between 70 and 250 Hz, so that a maximum of the spectral values 56 in this range indicates speech.

Adaptive threshold values are preferably provided for the first and second parameter values in order to compensate for variable distances between a respective speaker and the recording microphone. The threshold value is determined adaptively for a block of interest based on the parameter values of several previous blocks (step 60), the previous blocks preferably comprising detected speech sections and noise sections. For example, the first parameter values of 30 previous blocks classified as speech section and the first parameter values of thirty previous blocks classified as noise section are used to determine the threshold value for the first parameter value. The first parameter values are summed up for each section type and the sums obtained are subtracted from each other. The result is weighted with a weighting factor to get the associated threshold for the first parameter value of the current block. This ensures that the threshold value is adjusted to the current level of the first parameter value in order to avoid incorrect classifications. The weighting factor is preferably set between 0 and 1 and controls the sensitivity of the detection.

The threshold value for the second parameter is preferably also determined according to the principle of the threshold value for the first parameter. In this case, however, the calculation rule is inverted, since the second parameter indicates language with a decreasing amount and is therefore im is inversely correlated with speech compared to the first parameter. Consequently, the sum of the second parameter values for speech sections is subtracted from the sum of the second parameter values for noise sections and given a weighting factor, preferably between 0 and 1, which controls the sensitivity of the detection.

In step 64, the three parameters are evaluated together and it is determined whether or not the parameter values violate the associated threshold criterion. If two of the three parameter values indicate a speech section, i.e. violate the associated threshold criterion, the block in question is provisionally detected as a speech section.

In order to avoid strongly fluctuating detection results, in particular implausible alternating changes between speech sections and noise sections, a change between a speech section and a noise section and vice versa is only permitted if a predetermined number of consecutive blocks have been classified as speech section or noise section (step 66 and 68). For example, after a block detected as a noise section, five immediately consecutive blocks must be provisionally detected as a speech section in order to finally detect these blocks as a speech section (step 70). Otherwise, the blocks are still detected as noise sections (step 72). Conversely, after a block detected as a speech section, for example eight immediately consecutive blocks must be provisionally detected as a noise section in order to finally detect these blocks as noise sections (step 72). Otherwise, the blocks are still detected as sections of speech (step 70).

Further steps of the procedure are described below 1 explained. In step 20, the audio signal is weighted with a fixed amplification factor in order to compensate in advance for level losses caused by subsequent filters. For example, the signal can be amplified by 3 to 6 dB.

In step 22 the audio signal is filtered with a noise filter adapted to reduce very quiet portions of the audio signal. It is assumed here that very quiet signal sections do not contain any relevant information and in this respect can at most negatively affect the perceived voice quality. In particular, the risk of feedback is reduced by reducing the signal level in very quiet signal sections. In particular, a so-called noise gate can be used as a noise filter, which is adapted to suppress quiet signal sections. A threshold value, which is compared with the current signal level, is used as the criterion for recognizing quiet signal sections. If the current signal level falls below the threshold, the noise filter is activated. The threshold value is preferably well below the reference volume set in step 16 . For example, the threshold can be -55 dB. If the threshold value is not reached, the audio signal is reduced with a ratio in the range of 5 to 10. Values in the range of 10 ms or 100 ms are preferably used as rise time (attack time) and decay time (release time).

In step 24, second filter parameters 46 are determined, which are used for the subsequent steps 32, 34 and 36. The second filter parameters 46 include, on the one hand, the speech sections 52 already detected in step 18. Octave spectral values 48 are also determined, which in comparison to the spectral values 56 have a coarser Have spectral resolution that is modeled on human auditory perception. For this purpose, the spectral values 56 determined, for example, by means of FFT are filtered with an octave filter bank. The octave filter bank comprises a total of eight filters that overlap in the spectral range and are 3 are represented by way of example by magnitude frequency responses 37 over the frequency F and the magnitude G. The frequency responses 37 have their respective maximum at a filter-specific mid-frequency fc and fall towards smaller and larger frequency values. The center frequencies fc are preferably 63, 125, 250, 500, 1000, 2000, 4000 and 8000 Hz. The cut-off frequencies (magnitude frequency response of -3 dB) can be calculated generically on the basis of the respective center frequency fc. The lower cutoff frequency is 32fc/45 and the upper cutoff frequency is 45fc/32. For filtering, the weighted spectral values falling into a respective filter are summed up, with the weights each representing the absolute value frequency response at the frequency of the spectral value in question.

In step 24, feedback frequencies 50 are also determined, which are used as part of the filter data 46 for a feedback filter, which is used in step 34. The determination of the feedback frequencies is based on 4 explained in more detail.

A maximum value analysis is used to select a number of candidates from the spectral values 56, which represent possible feedback frequencies. For example, those spectral values can be sought out as candidates from the spectral values 56 which in each case have the highest absolute value of all spectral values of a block and are adjacent to spectral values with a similar absolute value. The candidates thus represent the maxima of pronounced extrema of the spectrum. For each candidate, three parameter values are determined (step 74) and compared to a respective threshold (step 78). The threshold values are preferably permanently set for each parameter because the parameters are generally insensitive to a voice signal volume that is low compared to the background noise.

A first parameter represents the ratio between the magnitude of the candidate and the associated harmonics (Peak-to-Harmonic Ratio, PHPR). Preferably, the first two harmonics are used, i.e. the spectral values that represent double and triple the frequency compared to the candidate. High PHPR values indicate a feedback frequency (feedback frequency) because speech usually has a clear overtone structure with harmonics.

A second parameter represents the ratio between the magnitude of the candidate and the magnitude of immediately neighboring spectral values (Peak-to-Neighboring Ratio, PNPR). The first three adjacent spectral values in each frequency direction are preferably used. High PNPR values indicate a feedback frequency because speech tends to have less steep frequency maxima.

A third parameter represents the course of the absolute value of the candidate over time (Interframe Magnitude Slope Deviation, IMSD). The mean increase in the absolute value of the candidate and a number of adjacent spectral values is preferably determined over five previous blocks. Positive IMSD values of, for example, 0.5 dB typically indicate a feedback frequency, because the amount of the fundamental frequency of speech does not usually increase over several blocks.

For more information on the calculation of the parameters, refer to the Publication, TV Waterschoot, M. Moonen, "Comparative Evaluation of Howeling Detection Criteria in Notch-Filter-Based Howling Supression", Journal of the Audio Engineering Society, Vol. 58, pp. 923-940, 2010 , referenced.

If all three parameters for determining the feedback frequency for a candidate in question violate the associated threshold criterion, the feedback frequency is preferably determined as a maximum of the spectrum in the region of the candidate in question. For this purpose, the spectrum is interpolated on the basis of the candidate and the adjacent spectral values with an interpolation function (e.g. by parabolic interpolation) and then the maximum of the interpolation function is formed. In particular, this maximum can lie between two spectral values, so that the interpolated maximum is more precise. The feedback frequency determined in this way is used as part of the filter data 50 for the feedback filter (step 34).

In order to relieve the computer resources, it is preferred not to subject the underlying candidate to the parameter analysis again for a predetermined period of time after a successfully determined feedback frequency if the candidate is identified as such again. For example, the same candidates are not checked again within a 1 second time window to determine whether or not they represent a feedback frequency. Instead, the feedback frequency determined for the temporally previous candidate is adopted for the subsequent, same candidate, because there is a high probability that the same feedback frequency will also be used for the subsequent candidate would be determined. Only after the predetermined time has elapsed is a relevant candidate checked again.

For each specific feedback frequency, a so-called bell filter (peak filter) is provided in the feedback filter, the center frequency of which is set to the specific feedback frequency. The Q value of the filters is preferably set to a fixed value. In addition, the gain of the filter is preferably adjusted adaptively, as shown below with reference to FIG figure 5 is explained.

the inside figure 5 The algorithm illustrated implements a finite state machine (FSM) which is initially in an inactive state 90, ie the bell filter has a gain of 0 dB and does not affect the audio signal. At a newly determined feedback frequency, an active state 92 is entered, in which the bell filter is operated with full (negative) gain. After a first predetermined time X has elapsed, a change is made to a first reduction state 94 if the feedback frequency has not been determined again by then and the active state is therefore retained (feedback 96). In the first reduction state, the bell filter has a reduced gain, for example 2/3 of full gain. The feedback filter is thus operated with reduced effectiveness. After a second predetermined time Y has elapsed, a change is made to a second reduction state 98 if the feedback frequency has not been determined again by then and the active state is retained (feedback 96).

After the second predetermined time Y has elapsed again, a change is made to the third reduction state 100, in which the bell filter for the Change to the inactive state is scheduled for the next filter run.

The time-dependent adaptation of the feedback filter is advantageous for several reasons. On the one hand, it ensures that a specific feedback frequency is filtered for a sufficiently long time. Feedback typically lasts for at least a few 100 ms, so long enough filtering is required to effectively suppress the feedback. In addition, due to the gradual reduction of the feedback filter, audible distortion of the audio signal is reduced.

In step 26, the audio signal is filtered with a two-stage compressor to remove peak levels that can lead to audible distortion. A first compressor stage is activated at a signal level above a first threshold and filters the audio signal with a first filter that reduces moderate level peaks with a low degree of compression (e.g. ratio 20, rise time 10 ms, decay time 100 ms). The second compressor stage is activated at a signal level above a second threshold, which is greater than the first threshold. The audio signal is then filtered with a second filter to remove extreme peaks particularly effectively. For this purpose, a stronger degree of compression is selected (e.g. ratio 1000, rise time 0.1 ms, decay time 5 ms). The second compressor stage provides an emergency filter to ensure that all amplitude values are below a critical maximum value

In step 28, the audio signal is bandpass filtered to remove potential spurious signals. To this end, preferably all spectral components that at least predominantly do not represent speech are reduced. Speech signal components are predominantly on the Frequency range between 70 and 8000 Hz is limited so that spectral components outside this frequency range can be filtered. A double-cascaded second-order high-pass filter is preferably combined with a double-cascaded second-order low-pass filter as the band-pass filter. The high-pass filter and the low-pass filter preferably each have an edge steepness of 24 dB per octave. The limit frequencies are preferably in the range between 60 and 80 Hz (lower limit frequency) and between 8000 and 10000 Hz (upper limit frequency). Furthermore, the Q values of the filters should extend over an octave and have values in the range of 1.4, for example.

In step 30, the audio signal is filtered with a second compressor in order to reduce the dynamic range of the audio signal. This makes the subjective loudness more uniform and speech intelligibility is improved. A filter with a relatively mild degree of compression, which is in particular lower than the degree of compression of the first compressor from step 28, is used as the compressor. For example, a low ratio can be selected, which should not exceed the value of three. In addition, longer rise and decay times in the range of 0.5 and 1 second are preferably provided.

In step 32, the audio signal is filtered with an equalizer to reduce spectral variations. For this purpose, the equalizer is operated with eight bell filters whose center frequencies correspond to those of the octave band filters from 3 correspond, which are used to determine the octave spectral values. The Q values of the bell filters are preferably set to cover about an octave each. A separate amplification factor is provided for each bell filter, which is determined as a function of the octave spectral values 48 and predefined reference spectral values. The reference spectral values correspond in their spectral resolution to the octave spectral values, so that each octave spectral value is assigned a reference spectral value.

The reference spectral values together form a reference spectral curve, the shape of which is correlated with a high level of speech intelligibility and can be determined, for example, by spectral evaluation of a large number of undisturbed speech signals, e.g. on the basis of a mean value of the octave-filtered spectrum. Each octave spectral value is compared to an associated reference spectral value in order to determine an amplification factor which represents the deviation between the octave spectral value and the associated reference spectral value. If a relevant octave spectral value is below the associated reference spectral value, for example, an amplification factor for the bell filter of this spectral range is determined such that weighting the octave spectral value with the weighting factor at least approximately results in the reference spectral value. The gain factors are adjusted in this way to bring the frequency spectrum of the audio signal into agreement with the reference spectral curve and thus reduce spectral variations within the audio signal and between different audio signals. For example, characteristics of different speakers or spectral influences are compensated for by different microphone positions in favor of high speech intelligibility.

In order to avoid distortions, the amplification factors are preferably limited above and below. In addition, the change in the amplification factors over time is also limited.

The bell filters for filtering the audio signal in step 32 are preferably used only for filtering blocks that are classified as Speech section have been detected. Thus, the fitting of the spectrum to the reference spectral curve is limited to speech sections. Any distortions and inefficient use of computing resources are thus avoided.

The filtering with the equalizer or bell filters in step 32 can cause undesired variations in the signal level. In order to compensate for such variations, the audio signal is preferably weighted with a correction factor which is determined as the mean value of the sign-inverted weighting factors.

In step 36 the audio signal is filtered with a pause filter in order to reduce the signal level in areas outside the detected speech sections, i.e. in speech pauses, and in this way to reduce background noise. For this purpose, the speech sections detected in step 18 or 24 are used as filter data 52 . Those sections of the audio signal which have not been detected as speech sections form noise sections which are filtered by the pause filter. The audio signal is preferably weighted in the detected noise sections with a fixed negative gain factor of, for example -3 dB.

In step 38 the audio signal is filtered with a further equalizer in order to compensate for the effects of the different filtering. A filter bank consisting of 23 bell filters between 50 Hz and 10 kHz is preferably used for this purpose. The filters preferably each extend over a third of an octave, with the Q value being adjustable to 4.3. A fixed negative gain factor is preferably provided for each bell filter.

In step 40 the audio signal can be analyzed for test purposes during a development phase. This option is purely optional and not necessary for a later application of the method in practice.

In step 42, the now improved audio signal is first transformed into an analog signal by means of a digital-to-analog converter and then made available via an output interface. From there, the audio signal can be picked up for playback via a sound reinforcement system. It is also conceivable for the digital audio signal to be output instead of an analogue version, provided the sound reinforcement system has a digital signal input for the audio signal.

Regarding 6 an audio device 102 is described below, which is set up to carry out the method of 1 to execute. The audio device 102 has a housing 104 indicated schematically. The external dimensions of the housing 104 are preferably no larger than a few centimeters, for example a maximum of 10 centimeters, so that the housing 104 is compact overall and is also particularly suitable for mobile applications.

The audio device 102 has an input interface 112 for receiving an analog audio signal and an output interface for outputting the enhanced audio signal. The device also has a USB-C interface 110 and an Ethernet interface 108 . The USB-C interface 110 can generally be embodied as a power supply interface for connecting to an external power supply. It does not necessarily have to be designed according to the USB-C standard.

Additionally or alternatively, one or more wireless interfaces can be provided in order to wirelessly receive audio signals and/or control signals and/or electrical energy from outside and/or transmit them to a receiver (not shown).

The input interface 112 and the output interface 106 are preferably each designed as XLR interfaces, so that conventional sound transducers can be connected directly to the audio device 102 via XLR connectors.

The audio device 102 can thus be used in particular in an in 7 shown arrangement in which the input interface 112 is connected to a microphone 134 for detecting an audio signal from a speaker, not shown. Furthermore, the output interface 106 is connected via an amplifier 130 to a loudspeaker 132 or a public address system with multiple loudspeakers in order to reproduce the audio signal enhanced by the audio device 102 . Speaker 132 and microphone 134 are in the same room, such as a conference room or the like. The signal improvement takes place in real time, so that the audio signal recorded with the microphone 134 can be played back essentially simultaneously via the loudspeaker 132 and thus ensures an acoustically advantageous amplification of the audio signal.

The audio device 102 also includes a manual interface 128, which is 6 is indicated only schematically and is generally set up to receive control data for the audio device 102 directly at the audio device 102 by manual input from a user.

The audio signal is first recorded with the microphone 134 and fed to a preamplifier 116 via the input interface 112 . Depending on the position of a switching device 118, the audio signal then reaches the output interface 106 either via a processing unit 114 or directly. For this The processing unit 114 can receive a specification from the outside via the interfaces 108, 110 and/or 128 that defines whether the audio signal is to be routed through the processing unit 114 and improved by it or not. As an alternative or in addition, the arithmetic unit 114 can determine its functionality for executing the method for improving the audio signal by means of a self-diagnosis and set the switch position of the switching device 118 depending on the test. For example, in a basic setting, switching device 118 can connect input interface 112 directly to output interface 106 via preamplifier 116, with switching device 118 only being switched over if processing unit 114 is fully functional, including the necessary power supply, in order to connect input interface 112 to processing unit 114 connect to. This ensures that the audio signal can be tapped from the output interface 106 independently of any malfunction of the computing unit 114 and a failure of the energy supply. The audio device 102 is therefore particularly well suited for professional use.

The preamplifier 116 can be operated with variable gain. For this purpose, a respective amplification value can be set by the arithmetic unit 114 . This can, for example, be selected directly on the device 102 by means of the interface 128 from a predetermined number of different amplification values, for example three amplification values. The selection of the amplification value can be conveyed visually to the operator by means of an illuminated display, eg by means of a number of LED diodes, on the audio device 102 . By setting the preamplification appropriately, large level variations can preferably already be compensated for in the analog signal, so that digital noise due to high amplification of the digital signal can be avoided.

On the one hand, the interface 110 is provided for the energy supply of the audio device 102, which can be connected to a mains source by means of an associated supply cable in order to operate the audio device 102 in mains operation. Alternatively, the audio device 102 can be supplied via an energy store integrated in the housing 104, for example an electric battery 126. The rechargeable battery 126 is coupled to the interface 110 and can be charged via it. Instead of the USB-C interface 110, another type of interface can also be provided for the power supply.

To protect against overvoltage or incorrect polarity, the device 102 is preferably equipped with an electrical protective device 120, which protects the electrical consumers of the audio device 102 from voltage damage. These include in particular the computing unit 114, a fan 124 for cooling the computing unit 114 and a phantom power supply device 122 which is coupled to the input interface 112. The phantom power device 122 is used to supply the microphone 134 connected to the input interface 112 with electricity, for example with a microphone supply voltage of 48 volts. The phantom power device 122 has a voltage converter, not shown in detail, in order to convert the supply voltage of the audio device 102, which is provided via the USB-C interface 110, for example 5 volts, into the microphone supply voltage.

The computing unit 114 is preferably designed as a single-board computer, so that the audio device 102 can be designed to be compact from this point of view and can also be produced inexpensively. The computing unit 114 is configured in particular via a bus interface 107, which is preferably of the USB-A type. For this purpose, the interface 107 is connected to a server or directly to a mobile terminal device connected (not shown) to access the computing unit 114 from the outside and optionally one or more configuration parameters for the method of 1 (e.g. threshold values, rise and decay times). A configuration via the USB-C interface 110 is also conceivable.

Alternatively, it is possible to connect a USB stick or the like to the interface 107, with the desired configuration data or new firmware being stored on the USB stick. The data are then transmitted to the processing unit 114 automatically or after initiation by an operator via the interface 107 in order to update the configuration parameters or the firmware accordingly. This process can be performed by an end user of the device.

However, a detailed configuration of filter parameters by the end user is preferably not required. All the necessary configuration parameter values are already stored in an internal memory of the computing unit (not shown), so that the method ensures good results fully automatically under almost all usual acoustic environmental conditions. For special acoustic environments, the configuration parameter set can be adjusted remotely or locally via the interface 107 by a trained expert, for example. This means that there is no setup effort for the end user. For commissioning in the application of 7 it is only necessary to connect the audio device 102 to the microphone 134 and the loudspeaker 132 via the interfaces 112 and 106 provided. Then the audio device 102 can be used directly in terms of a plug-and-play functionality. If battery operation is not desired, the audio device 102 is connected via the USB-C interface 110 connected to a power source (not shown) to power the audio device 102 electrically.

The audio device 102 also has a manual operating interface 113 (eg with a manually operable button) and a visual display device 109 (eg an LED). A user of the audio device 102 can control a recording of the audio signal provided at the output interface 106 via the operating interface 113 . For example, the user first connects a USB stick or the like to the interface 107. The USB stick is detected by the computing unit 114 and the user is shown on the display device 109 by activating a first display mode that the audio device 102 is ready to record. The user interface 113 is then actuated in order to store the audio signal (in its digital form) in the USB stick. The display device 109 indicates the successful start of the recording by activating a second display mode (eg flashing LED). The audio signal is then stored continuously in a file on the USB stick. When the storage capacity is full, the recording will stop automatically. This is indicated to the user by activating a third display mode on the display device 109 . The recording can optionally be ended prematurely by operating the user interface 107 again.

REFERENCE LIST

10

Capturing an audio signal with a microphone

12

preamp

14

Capturing the audio signal at an input interface

16

Electronic Amplifier (First Level Filter)

18

input analysis

20

Software amplifier (second level filter)

22

noise filter

24

interim analysis

26

bandpass

28

First Compressor

30

Second Compressor

32

First equalization filter and third level filter

34

feedback filter

36

pause filter

37

magnitude frequency response

38

Second equalization filter

40

initial analysis

42

Providing the audio signal at an output interface

44

First filter data

46

Second filter data

48

octave volumes

50

feedback frequencies

52

Detected speech segments

54

amplitude values

56

spectral values

58

parameter calculation

60

threshold calculation

62

Compare to thresholds

64

Determine if thresholds violated

66

Determine number of consecutive sections

68

Compare with minimum number

70

Detection as a section of speech

72

Detection as noise section

74

Search frequency candidates

76

parameter calculation

78

Comparison with threshold values

80

branch

82

interpolation

84

Storage feedback frequency

86

Clearing the feedback frequency

88

End

90

inactive state

92

active state

94

First state of reduction

96

return

98

Second state of reduction

100

Third state of reduction

102

audio device

104

Housing

106

output interface

107

USB-A interface

108

Ethernet interface

109

display device

110

USB-C interface

112

input interface

113

Manual operator interface

114

unit of account

116

preamp

118

switching device

120

protective device

122

phantom power

124

Fan

126

energy storage

128

Manual operator interface

130

amplifier

132

speaker

134

microphone

f

frequency

G

Amount

FC

center frequency

Claims (12)

1. A method for enhancing an audio signal, in particular in real time, the method comprising at least the following steps:

   - Receiving an audio signal having multiple amplitude values, the audio signal having speech at least in sections;

   - Detecting speech sections of the audio signal (18, 24);

   - filtering the audio signal with at least one level filter (16) to reduce signal level variations of the audio signal in the detected speech portions;

   - determining a feedback frequency (50) representing a feedback of the audio signal;

   - filtering the audio signal with a feedback filter (34) based on the determined feedback frequency (50) to reduce spectral components of the audio signal representing feedback; and

   - filtering the audio signal with at least one equalization filter (32) to reduce spectral variations of the audio signal in the detected speech portions, the filtering with the at least one equalization filter (32) comprising:

   - determining coarse spectral values (48) based on fine spectral values (56) of the audio signal, wherein the coarse spectral values (48) represent the fine spectral values (56) at a lower spectral resolution than the fine spectral values (56);

   - determining first equalization weights representing a deviation of the coarse spectral values (48) from predetermined reference spectral values;

   - weighting the audio signal with the first equalization weights to bring spectral values of the audio signal into agreement with the reference spectral values;

   wherein determining the feedback frequency (50) comprises:

   - Determining a subset of spectral values of the audio signal that violate a predetermined spectral threshold (74);

   - determining a plurality of first spectral parameter values based on the subset, each of the first spectral parameter values representing a predetermined relationship between an associated spectral value of the subset and at least one temporally and/or spectrally adjacent spectral value (76); and

   - determining the feedback frequency (50) based on the plurality of first spectral parameter values (78, 80, 82, 84).
2. The method of claim 1,
   further comprising determining a plurality of spectral values (56) based on the amplitude values (54), wherein the amplitude values (54) represent the audio signal in a time domain and wherein the spectral values (56) represent the audio signal in a frequency domain, and wherein detecting the speech portions (18, 24), filtering with the at least one level filter (16) and/or filtering with the at least one equalization filter (32) is based on the amplitude values (54) and/or the spectral values (56).
3. Method according to claim 1 or 2,

   wherein detecting the speech segments (18, 24) comprises:

   - Determining at least a first energy parameter value based on the amplitude values (54), wherein the first energy parameter value represents an average energy of the audio signal for a plurality of the amplitude values (54);

   - determining at least one second spectral parameter value based on spectral values (56) of the audio signal, the at least one second spectral parameter value representing a harmonic spectral structure of the audio signal for a plurality of the spectral values (56); and

   - detecting a portion of the audio signal as a speech portion when the at least one first energy parameter value violates a first energy parameter threshold value and/or the at least one second spectral parameter value violates a spectral parameter threshold value (62, 64),

   in particular wherein the energy parameter threshold value and/or the spectral parameter threshold value is adjusted as a function of time.
4. A method according to any one of the preceding claims,
   wherein filtering the audio signal with the at least one level filter (16) comprises:

   - Determining at least one level parameter value based on the amplitude values (54), wherein the level parameter value represents an average level of the audio signal for a detected speech portion;

   - determining at least one compensation weight based on the at least one level parameter value;

   - weighting the audio signal with the at least one compensation weight to reduce signal level variations of the audio signal.
5. A method according to claim 4,

   wherein the at least one level parameter value comprises first and second level parameter values for a plurality of detected speech segments, wherein the first level parameter values represent the average level of the audio signal at a first time resolution, wherein the second level parameter values represent the average level of the audio signal at a second time resolution, wherein the second time resolution is greater than the first time resolution, and wherein the at least one compensation weight is determined based on the first and second level parameter values,

   in particular wherein the first level parameter values are formed by average energy values and/or first loudness values and the second level parameter values are formed by second loudness values.
6. Method according to claim 4 or 5,

   wherein said at least one compensation weight comprises first compensation weights and second compensation weights, said first compensation weights being determined to reduce signal level variations having at least one level greater than a predetermined level threshold value,

   wherein the second compensation weights are determined to adjust the signal level of the audio signal to a predetermined value.
7. A method according to any one of the preceding claims,
   wherein filtering with the at least one equalization filter (32) comprises:

   weighting the audio signal with second equalization weights (38),

   wherein the second equalization weights are predetermined.
8. A method according to any one of the preceding claims, further comprising:

   filtering the audio signal with at least one compressor (28, 30) to reduce a dynamic range of the audio signal,

   in particular wherein a plurality of mutually different parameter sets are provided for the at least one compressor (28, 30), which parameter sets are selected in dependence on a magnitude of the audio signal and are used as a basis for the filtering with the at least one compressor, wherein the plurality of parameter sets differ from one another in a degree of compression.
9. Method according to any one of the preceding claims,
   wherein, when the determined feedback frequency (50) disappears between successive time periods of the audio signal, the effectiveness of the feedback filter (34) is gradually reduced (94, 98, 100) over a plurality of time periods.
10. Method according to any one of the preceding claims,
    further comprising:

    - Filtering the audio signal with a pause filter (36) to reduce the audio signal in regions outside of the detected speech segments; and/or

    - filtering the audio signal with a noise filter (22) to reduce the audio signal in regions having amplitude values that violate a predetermined noise threshold; and/or

    - filtering the audio signal with a bandpass filter (26), wherein a lower cut-off frequency of the bandpass filter preferably lies in a range of 50 to 100 Hz, and wherein an upper cut-off frequency of the bandpass filter preferably lies in a range of 8000 to 10000 Hz.
11. Device for enhancing an audio signal, in particular in real time, said audio signal comprising speech,
    the device (102) comprising:

    - at least one input interface (112) for detecting an audio signal, the input interface (112) having a connector for a microphone (134);

    - at least one output interface (106) for outputting the audio signal, the output interface (106) having a connector for an audio playback device (130, 132); and

    - a computing unit (114) for performing a method for enhancing the audio signal according to any of the preceding claims.
12. Device according to claim 11,

    further comprising:

    - a preamplifier (116) for the audio signal, the preamplifier (116) being couplable to the input interface (112); and/or

    - an electrical supply (122) for the input interface (112); and/or

    - a switching means (118) couplable to the input interface (112), the output interface (106) and the computing unit (114); and/or

    - a cooling device (124);

    and/or wherein the computing unit (114) comprises a single board computer; and/or

    wherein the device (102) comprises a housing (104); and/or

    at least one external communication interface (108, 110).

Images