WO2025233243A1

WO2025233243A1 - Apparatus and method for processing a pre-processed audio input signal to obtain an activity output signal

Info

Publication number: WO2025233243A1
Application number: PCT/EP2025/062083
Authority: WO
Inventors: Sanya TAYAL; Markus Schnell; Martin Müller; Goran MARKOVIC; Sebastian BOLTEN; Christian Uhle
Original assignee: Fraunhofer Gesellschaft zur Foerderung der Angewandten Forschung eV
Current assignee: Fraunhofer Gesellschaft zur Foerderung der Angewandten Forschung eV
Priority date: 2024-05-08
Filing date: 2025-05-02
Publication date: 2025-11-13
Anticipated expiration: 2026-11-08
Also published as: EP4648049A1

Abstract

Disclosed is an apparatus for processing a pre-processed audio input signal to obtain an activity output signal; comprising a feature extractor to receive the pre-processed audio input signal comprising a spectral energy and/or a noise-free spectral energy in at least one frequency band of a plurality of frequency bands. The feature extractor extracts at least one feature depending on the spectral energy and/or depending on the noise-free spectral energy. A temporal context extractor receives the at least one extracted feature and extracts at least one temporal context feature based on the at least one extracted feature, or receives at least one further extracted feature and extracts at least one temporal context feature based on the at least one further extracted feature. A feature combiner receives and combines the features using the plurality of static combination weights to obtain the output signal.

Description

Apparatus and method for processing a pre-processed audio input signal to obtain an activity output signal

Description

Disclosed is an apparatus for processing a pre-processed audio input signal to obtain an activity output signal, a method of processing a pre-processed audio signal to obtain an activity output signal, and a corresponding computer program.

Voice Activity Detection (VAD) algorithm is commonly used in audio and speech coding to support the discontinuous transmission (DTX) mode which allows efficient coding by significantly reducing the bits to be transmitted for unwanted parts like silence and background noise. This is useful because the activity rate of speech during a conversation is usually around 50% so the efficiency can be almost doubled by not sending the information present in these frames. Only some parametric information can be transmitted to generate the comfort noise (CNG) at the decoder end that is used in place of the actual signal to provide perception of continuity during an ongoing call. Many codecs intended for use in multimedia applications employ a Signal Activity Detector (SAD) where the focus is not solely the detection of human speech but also other acoustic signals, depending upon the application, like music, applause or background noise with multiple speakers. However, both SAD and VAD algorithms share the same core principle of classification of audio signals into “active” (the useful signal to be detected) and “inactive” (noise or irrelevant signals). The definition of “active” signal in SAD is just broader. An effective SAD algorithm would also work well as VAD if the input signal is speech.

A very basic approach in VAD or SAD is the use of energy thresholding, which is effective primarily in high Signal-to-Noise Ratio (SNR) environments. However, its efficacy reduces significantly in conditions where noise levels are similar to that of speech. For such cases, computing energy relative to the noise levels is a more effective approach that involves the use of SNR as a key feature. This SNR feature has been already implemented in the VAD2 of the AMR codec [2][3] in 1998 and can be found in several newer VAD implementation of codecs such as G.718 [4] and EVS [5],

To detect signal activity based on SNR, VAD methods often rely on adaptively changing thresholds based on long-term noise estimation techniques (as proposed in US 9,761 ,246 B2 and in EVS [5]). In some cases, closed loop designs are also used (as shown in Fig. 10, which is taken from EP 2 162 881 B1), where the noise estimate is itself influenced by the past VAD decisions. Because of their adaptive nature based on long-term statistics, these methods usually require an initial start-up time to adjust to the noise environment. This means that for a certain period the VAD decision is not reliable. These methods can further be less effective in rapidly changing noise environments due to the inherent latency of the adaptive algorithm. When dependent on past VAD decisions, these methods can also lead to error-propagation.

One of the commonly used methods for noise estimation is also the 'minimum statistics' approach [1], This method works by measuring the lowest power levels in the noise spectrum, thereby establishing a baseline for noise levels and does not use any long-term statistics. This method can be effective in providing an effective noise estimate over a relatively short period of time. While using this method for SNR computation can be advantageous in a low delay real-time system, it may not be reliable when used as the sole criterion for signal activity detection, particularly in complex acoustic environments. It results in low accuracy and does not work well for low energy voice segments in low SNR conditions [7, 8, 9], However, a combination of acoustic features alongside SNR can enhance the effectiveness of the VAD.

VADs based on adaptive mechanisms often require manual tuning efforts for optimization and do not allow automatic global optimization of the algorithm. This is also true for VADs where the final decision is a combination of multiple VAD decisions. The VAD proposed in US 9,773, 511 B2, the VAD proposed in EVS [5] and the one shown in Fig. 9 [4] are a few examples for such a VAD. This approach may simplify the manual optimization effort as single VADs can be optimized manually for dedicated content. However, it makes the global optimization difficult which is straightforward in case of simple machine learning models using automatic training procedures. These automatic optimization methods can prove to be very beneficial when the SAD needs to be adapted for various applications with unfamiliar noise environments. This makes the SAD algorithm more versatile to be used across a wide range of applications and acoustic settings.

A task of the present invention is to provide an improved apparatus and an improved method for Signal Activity Detection (SAD), in particular which requires negligible preamble time, remains stable under fast-changing conditions, allows automatic global optimization while having low-complexity. This task is solved by an apparatus for processing a pre-processed audio input signal to obtain an activity output signal according to claims 1 , by a method of processing a pre- processed audio signal to obtain an activity output signal according to claim 17, and a computer program according to claim 21.

According to a first aspect of the present invention, an apparatus for processing a pre-processed audio input signal to obtain an activity output signal is disclosed. The apparatus comprises a feature extractor adapted to receive the pre-processed audio input signal, wherein the pre-processed audio input signal comprises a spectral energy and/or a noise- free spectral energy in at least one frequency band of a plurality of frequency bands. The feature extractor is adapted to extract at least one feature and/or at least one further feature, each depending on the spectral energy and/or depending on the noise-free spectral energy. The apparatus comprises a temporal context extractor, adapted to receive the at least one extracted feature, wherein the temporal context extractor is adapted to extract at least one temporal context feature based on the at least one extracted feature, or the temporal context extractor, adapted to receive at least one further extracted feature, wherein the temporal context extractor is adapted to extract at least one temporal context feature based on the at least one further extracted feature. The apparatus comprises a feature combiner coupled with the feature extractor and coupled with the temporal context extractor, and the feature combiner adapted to receive the at least one extracted feature and/or the at least one further extracted feature and the at least one temporal context feature and a plurality of static combination weights for combining the at least one extracted feature and/or the at least one further extracted feature and the at least one temporal context feature using the plurality of static combination weights to obtain the activity output signal. In particular, the feature extractor is adapted to extract the at least one feature depending on the spectral energy and to extract the at least one further feature depending on the noise- free spectral energy. Further in particular, the feature extractor is adapted to extract the at least one further feature depending in the spectral energy and to extract the at least one feature depending on the noise-free spectral energy.

In particular, the pre-processed audio input signal comprises a power spectrum, and/or a pitch estimate. According to the present disclosure, the feature extractor is adapted to extract the at least one feature depending on the spectral energy and/or depending on the noise-free spectral energy. The at least one feature is provided to the temporal context extractor and is provided to the feature combiner. The at least one temporal context feature extracted from the at least one feature by the temporal context extractor is provided to the feature combiner as well. Alternatively or additionally, the feature extractor is adapted to extract the at least one further feature depending on the spectral energy and/or depending on the noise- free spectral energy and the at least one feature depending on the spectral energy and/or depending on the noise-free spectral energy. In this case, the at least one further feature is provided to the temporal context extractor and the at least one feature is provided directly to the feature combiner. The at least one feature and the at least one further feature are different features. A temporal context feature represents time-dependent context of the at least one feature or the at least one further feature, which can be the estimated mean value or variance, for example.

Stated differently, in both cases, the feature combiner is adapted to receive the at least one extracted feature and/or or the at least one further extracted feature, the at least one temporal context feature and the combination weights. The proposed apparatus, i.e. the proposed SAD (signal activity detector), is using a feature extractor, a temporal context feature extractor and a feature combiner to obtain an improved activity output signal using an open loop design and a feedforward structure. Using an open loop design means that a current SAD decision or the activity output signal does not depend on previous SAD decisions or previous activity output signals. The system is also independent of any adaptive mechanisms like adaptively changing thresholds based on changing signal conditions as the system parameters, for example the combination weights remain static. The feature combiner has a feedforward structure which means that the information flows in one direction from input to the output which is usually the case for fully connected artificial neural networks, for example. The proposed solution allows automatic global optimization using training methods like in the case of simple neural networks. Due to the usage of said extractor and the combiner and thus due the open loop design, no long-term statistics are required to stabilize the decision. According to the present invention the SAD or VAD decision becomes active after negligible preamble time, for example after an initialization period of around 0.1 second.

According to the present invention, an apparatus is proposed allowing an automatic global optimization of the SAD. Presently, “global” means that the entire algorithm is optimized as a whole rather than optimizing multiple VADs separately. The optimization process is automatic because it does not require manual effort because the training procedure for machine learning model is in place and the training happens automatically once the desired training dataset is available. According to a second aspect of the present invention, a method of processing a pre-processed audio input signal to obtain an activity output signal is disclosed. The method comprises receiving, by a feature extractor, the pre-processed audio input signal comprising a spectral energy and/or a noise-free energy in at least one frequency band of a plurality of frequency bands. The method further comprises extracting, by the feature extractor, at least one feature and/or at least one further feature, each depending on a spectral energy and/or at least one feature depending on a noise-free spectral energy in at least one frequency band of a plurality of frequency bands. The method further comprises receiving, by a temporal context extractor, the at least one extracted feature, wherein the temporal context extractor extracts at least one temporal context feature based on the at least one extracted feature, or receiving, by the temporal context extractor, at the least one further extracted feature, wherein the temporal context extractor extracts at least one temporal context feature based on the at least one further extracted feature. The method comprises receiving, by a feature combiner coupled with the feature extractor and coupled with the temporal context extractor, the at least one extracted feature and/or the at least one further extracted feature and the at least one temporal context feature and a plurality of static combination weights for finally combining the at least one extracted feature and/or the at least one further extracted feature and the at least one temporal context feature using the plurality of static combination weights to obtain the activity output signal.

The proposed method uses a feature extractor, a temporal context feature extractor and a feature combiner to obtain an improved activity output signal using an open loop design and a feedforward structure. Using an open loop design means that a current SAD decision or the activity output signal does not depend on previous SAD decisions or previous activity output signals. The system is also independent of any adaptive mechanisms like adaptively changing thresholds based on changing signal conditions as the system parameters, for example the combination weights remain static. The feature combiner has a feedforward structure which means that the information flows in one direction from input to the output which is usually the case for fully connected artificial neural networks, for example. The proposed solution allows automatic global optimization using training methods like in the case of simple neural networks. Due to the usage of said extractor and the combiner and thus due the open loop design, no long-term statistics are required to stabilize the decision. According to the present invention the proposed method allows the SAD or VAD decision to become active after negligible preamble time, for example after an initialization period of around 0.1 second.

With respect to the proposed apparatus and the proposed method, it is noted that the audio input signal is the input to the codec pre-processor and consists of a sequence of short-time frames, for example, each of length 20 milliseconds. The audio input signal is pre-processed by the codec pre-processor on a frame-by-frame basis to get the pre-processed audio input signal also as a sequence of frames. Subsequently, there is an activity output decision for each frame. The audio input signal may comprise speech and/or noise and/or music.

The pre-processed audio input signal refers to the output of the codec-preprocessor and comprises a spectral energy and/or a noise-free spectral energy in at least one frequency band of a plurality of frequency bands. The pre-processed audio input signal may also comprise additional outputs from the codec-preprocessor, for example, power spectrum, and/or a pitch estimate which will be explained in more detail in the description of the figures. Presently, the term “noise-free” implies that noise energy has been removed from the signal energy. The removal can be done by noise subtraction methods or by other noise-removal methods.

The feature extractor receives the pre-processed audio input signal from the codec-processor, but the pre-processed audio input signal is also used for other functions of the codec and not solely for feature extraction. In other words, the SAD algorithm uses significant amount of existing information from the codec pre-processor for feature extraction, and therefore its implementation does not significantly add to the overall complexity of the codec system.

According to a third aspect of the present invention, a computer program for performing, when running on a computer, the method described herein is proposed.

Because an automated training framework for the apparatus/method of processing a pre- processed audio signal to obtain an activity output signal according to the present invention is available, the optimization process requires minimal manual effort as opposed to frequent tuning that is required by the adaptive mechanisms described in the prior art. For example, optimization over new stationary or non-stationary noise types can easily be done by a simple training procedure.

It noted that all details described with respect to the apparatus are applicable to the method and vice versa. Thus, a detailed description of the apparatus can be understood as and applied to the method and vice versa. Preferred embodiments of the present invention are explained in detail below with reference to the accompanying drawings. Showing:

Fig. 1 an apparatus for processing a pre-processed audio input signal to obtain an activity output signal;

Fig. 2 an alternative apparatus for processing a pre-processed audio input signal to obtain an activity output signal;

Fig. 3 a detailed representation of the apparatus for processing a pre-processed audio input signal to obtain an activity output signal as shown in Fig. 1;

Fig. 4 a detailed representation of the alternative apparatus for processing a pre- processed audio input signal to obtain an activity output signal as shown in Fig. 2;

Fig. 5a, b a comparison of the activity output signal of the EVS (Enhanced Voice Services) decision known by the prior art compared to the SAD decision according to the present invention, when the initialization period is 3 sec (Fig. 5a) or when the initialization period is 0.1 sec (Fig. 5b);

Fig. 6a, b a comparison of the activity output signal of the EVS decision known by the prior art compared to the SAD decision according to the present invention, when the signal conditions change after a considerable break or silence in between (Fig. 5a) or when the signal conditions change suddenly (Fig. 5b);

Fig. 7 overview of the performance for the SAD of the present invention compared with the EVS VAD known in the prior art;

Fig. 8 a flow chart of the method of processing a pre-processed audio signal to obtain an activity output signal;

Fig. 9 a VAD known in the prior art; and

Fig. 10 another VAD known in the prior art. In the following description, embodiments are discussed in further detail using the figures, wherein in the figures and the specification identical elements and elements having the same functionality and/or the same technical or physical effect are provided with the same reference numbers or are identified with the same name. Thus, the description of these elements and of the functionality thereof as illustrated in the different embodiments are mutually exchangeable or may be applied to one another in the different embodiments.

Individual aspects of the invention described herein are described below in Figures 1 to 8. The specific embodiments discussed are merely illustrative of specific ways to make and use the present concept, and do not limit the scope of the embodiments. In the following description of embodiments, the same or similar elements having the same function have associated therewith the same reference signs or the same name, and a description of such elements will not be repeated for every embodiment. Moreover, features of the different embodiments described hereinafter may be combined with each other, unless specifically noted otherwise.

It is understood that when an element is referred to as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or intermediate elements may be present. Conversely, when an element is referred to as being "directly" connected to another element, "connected" or "coupled," there are no intermediate elements. Other terms used to describe the relationship between elements should be construed in a similar fashion (e.g., "between" versus "directly between", "adjacent" versus "directly adjacent", and “on” versus “directly on”, etc.).

The present invention proposes an open loop design and a feedforward structure for the SAD algorithm which has low complexity, requires negligible preamble time, remains stable under fast-changing signal conditions and allows automatic global optimization. Using an open loop design means that a current SAD decision or the activity output signal does not depend on previous SAD decisions or previous activity output signals. The system is also independent of any adaptive mechanisms like adaptively changing thresholds based on changing signal conditions as the system parameters, for example the combination weights, remain static. The feature combiner has a feedforward structure, which means that the information flows in one direction from input to the output which is usually the case for fully connected artificial neural networks. As a result, no long-term statistics are required to stabilize the decision and the algorithm works in a stable way under fastchanging signal conditions and with negligible preamble time. The proposed solution also allows automatic global optimization using training methods like in the case of simple neural networks. The algorithm also does not add to the overall codec complexity as several computations required for extracting the features are already performed for the general codec functions and not specifically for feature extraction for SAD.

The core of the proposed apparatus 100 comprises the feature extractor 20, the temporal context extractor 70, the feature combiner 30, and uses the combinations weights.

Fig. 1 and Fig. 2, each shows an apparatus 100 for processing a pre-processed audio input signal 10 to obtain an activity output signal 50. The apparatus 100 comprises a feature extractor 20 adapted to receive the pre-processed audio input signal 10, wherein the pre- processed audio input signal 10 comprises a spectral energy and/or a noise-free spectral energy in at least one frequency band of a plurality of frequency bands. The feature extractor 20 is adapted to extract at least one feature (in particular, and/or at least one further feature, each) depending on the spectral energy and/or depending on the noise-free spectral energy. The apparatus 100 comprises a temporal context extractor 70, adapted to receive the at least one extracted feature, wherein the temporal context extractor 70 is adapted to extract at least one temporal context feature based on the at least one extracted feature. Alternatively or additionally, the temporal context extractor 70 is adapted to receive at least one further extracted feature, wherein the temporal context extractor 70 is adapted to extract at least one temporal context feature based on the at least one further extracted feature. The at least one extracted feature and the at least one further extracted feature are, for example, two different features. Both of them, are extracted by the feature extractor 20. According to both Figs. 1 and 2, the at least one extracted feature is, in particular directly, coupled with or transmitted to a feature combiner 30. According to Fig. 1 , the at least one extracted feature is also, in particular directly, coupled with or transmitted to the temporal context extractor 70. According to Fig. 2, the at least one further extracted feature is, in particular directly, coupled or transmitted to the temporal context extractor 70 instead of the at least one extracted feature.

The apparatus 100 comprises the feature combiner 30 coupled with the feature extractor 20 and coupled with the temporal context extractor 70. The feature combiner 30 is adapted to receive the at least one extracted feature and/or the at least one further extracted feature and the at least one temporal context feature and a plurality of static combination weights for combining the at least one extracted feature and/or the at least one further extracted feature and the at least one temporal context feature using the plurality of static combination weights to obtain the activity output signal 50.

The apparatus comprises a feature extractor 20 adapted to receive the pre-processed audio input signal 10, wherein the pre-processed audio input signal 10 comprises a spectral energy and/or a noise-free spectral energy in at least one frequency band of a plurality of frequency bands. In particular, the pre-processed audio input signal 10 comprises a power spectrum, and/or a pitch estimate. The feature extractor 20 is adapted to extract at least one feature depending on the spectral energy and/or depending on the noise-free spectral energy. The apparatus 100 comprises a temporal context extractor 70, adapted to receive the at least one extracted feature, wherein the temporal context extractor is adapted to extract at least one temporal context feature based on the at least one extracted feature. Alternatively, the temporal context extractor 70 is adapted to receive at least one further extracted feature, wherein the temporal context extractor is adapted to extract at least one temporal context feature based on the at least one further extracted feature. The apparatus 100 comprises furthermore a feature combiner 30 coupled with the feature extractor 20 and coupled with the temporal context extractor, wherein the feature combiner 30 is adapted to receive the at least one extracted feature or the at least one further extracted feature and the at least one temporal context feature and a plurality of static combination weights for combining the at least one extracted feature or the at least one further extracted feature and the at least one temporal context feature using the plurality of static combination weights to obtain the activity output signal 50.

The at least one temporal context feature is extracted using the temporal context extractor 70. The temporal context extractor 70 is positioned directly between the feature extractor 20 and the feature combiner 30 (see Fig. 1 , 2, 3 or 4). Thus, the feature extractor 20 is directly coupled with the feature combiner 20 and is directly coupled with the temporal context extractor 70.

Fig. 3 shows the apparatus 100 of Fig. 1 in detail. Fig. 4 shows the apparatus 100 of Fig. 2 in detail. Figs. 3 and 4 show for example, which features of the at least one feature or of the at least one further feature are extracted by the feature extractor 20. Figs. 3 and 4 also show which information may be included in the pre-processed audio input signal 10.

The audio input signal 10’ is the input to the codec pre-processor 60 and consists of a sequence of short-time frames, for example, each of length 20 milliseconds. The audio input signal 10’ is pre-processed by the codec pre-processor 60 on a frame-by-frame basis to get the pre-processed audio input signal 10 also as a sequence of frames. The pre-processed audio input signal 10 refers to the output of the codec-preprocessor 60 and comprises of a spectral energy and/or a noise-free spectral energy in at least one frequency band of a plurality of frequency bands. The pre-processed audio input signal 10 may also comprise a power spectrum and/or a pitch estimate (see Figs. 3 and 4). The feature extractor 20 is configured to extract at least one feature and/or to extract the at least one further feature, each depending on spectral energy and/or noise- free spectral energy. The feature combiner 30 receives the extracted feature or the further extracted feature from the feature extractor 20, the temporal context feature from a temporal context extractor 70 and combination weights from a memory 40.

As can be seen in Figs. 1 to 4, preferably, the feature combiner 30 is directly coupled to the memory 40 storing the plurality of static combination weights which are to be received by the feature combiner 30. The combination weights are stored in the memory 40, wherein the combination weights are obtained, for example, from an automatic training procedure.

The feature extractor 20 is adapted to receive the pre-processed audio input signal to extract at least one feature depending on the spectral energy and/or depending on the noise-free spectral energy, for example signal-to-noise ratio (SNR). The feature extractor may also extract additional useful features to enhance the effectiveness of the SAD algorithm.. These can include other spectral energy-based features, spectral shape-based features, and harmonicity-based features. Additionally, temporal context features which are computed as higher order statistics of the extracted features can also be used in giving a stable output activity decision as they capture time-dependent characteristics of the audio signal. This multi-feature approach not only addresses the limitations of relying solely on SNR as a feature but also significantly improves the accuracy and reliability of SAD/VAD systems in diverse and challenging noise conditions.

As can be seen in Figs. 1 to 4, preferably, the feature extractor 20 is, in particular directly, coupled with the codec pre-processor 60, wherein the codec pre-processor 60 is operative to pre-process an audio input signal 10’ to obtain the pre-processed audio input signal 10. The codec pre-processor 60 is operative to transmit the pre-processed audio input signal 10 to the feature extractor 20. The pre-processed audio input signal 10 comprises information characteristic of the audio input signal 10’, e.g. a power spectrum that represents the spectral energy distribution across various frequencies of the audio input signal, and/or band energies which refer to spectral energies in specific frequency bands, and/or noise-free band energies which refer to noise-free spectral energies in specific frequency bands and/or a pitch estimate which gives information about the fundamental frequency and harmonicity of the audio signal.

In the context of the present invention, spectral energy in a frequency band which can also be referred to as band energy is the amount of signal energy present within a specific frequency range of an audio signal. It is computed using a 'power spectrum,' which gives the distribution of spectral energy across different frequencies of the audio signal. Note that the terms “power” and “energy” can be used interchangeably in the text.

The power spectrum is obtained using a Discrete Fourier Transform (DFT) which converts the time-domain representation of the audio signal 10’ into frequency-domain representation. This conversion results in a series of complex numbers which describe the magnitude and phase of each frequency component within the signal. The power spectrum is computed as the squared magnitude of the DFT spectrum:

P(fk) = m I² = W)² + Im(X[k])² where [/c] is the DFT output for the discrete frequency bin k, wherein k is a natural number, Re( [/c]) and /m( [/c]) are the real and imaginary parts of [/c] respectively. The power spectrum P f_k) is then used to compute band energies E_t for specific frequency bands. The energy Et for the i -th frequency band is obtained by summing the power spectrum values within the specified band limits and Ui as follows wherein i is a natural number. The limits of these frequency bands can depend on a predefined scale, such as the Bark scale, in which case these bands are called critical bands. According to the proposed solution the “band energies” may refer to critical band energies. In yet another example, L; may be equal to U_t, in which case the band energies correspond to the log of the power spectrum. Here, the logarithm is applied as it can help with scaling and is common in audio processing for better representation and handling of a wide range of power values. The frequency band energies can, alternatively, also be computed in the linear scale instead of the log scale. Note that the “frequency band energies” are also referred to as “band energies” or “spectral energies in frequency bands" in the text.

The noise-free band energies E-, are also obtained for each frequency band and computed by means of noise removal techniques, for example spectral subtraction. Presently, the term “noise-free” implies that noise energy has been removed from the signal energy, for example through spectral subtraction which can be expressed as E- = E_t - N_t, where is the noise band energy or noise spectral energy in the j-th frequency band expressed in the log domain and obtained through a noise estimation method, for example, minimum statistics which will be explained in detail further. The noise band energies can, alternatively, also be computed in the linear scale instead of the log scale. When computed in the log domain, the choice of the log base can be variable. However, both the spectral energy and the noise-free spectral energy should be computed using the same log base.

Preferably, the feature extractor 20 is operative to receive the spectral energies (also called herein band energies) and/or the noise-free spectral energies (also called herein- band energies) and to combine the spectral energies and/or the noise-free spectral energies into spectral energies and/or noise-free spectral energies in at least one larger frequency band of a plurality of larger frequency bands. For example, frequency band energies that exhibit high linear correlation are combined through an averaging operation. Since these groups of highly correlated bands are typically adjacent frequency bands, combining these into a single value represents spectral energy over a broader band. These features can be calculated respectively by combining either band energies, or noise-free band energies, or band energies and noise-free band energies. See Figs. 3 and 4. The following equation expresses how adjacent frequency band energies can be combined to get E_avg to get average energy over a larger frequency band.

Preferably the feature extractor 20 is operative to extract a spectral flux and/or at least one feature based on the spectral flux, wherein spectral flux represents a difference of energy in successive frames of the audio input signal. An example for a flux-based feature is the spectral flux which represents a change in spectral energy. This feature can be computed using either band energies, or noise-free band energies. As shown in Figs. 3 and 4, the spectral flux block receives the noise-free band energies. However, it is possible, that spectral flux block receives the band energies instead or in addition to the noise- free band energies also the band energies. These two possibilities are not shown in the Figs. 3 and 4. The following equation shows how to compute the spectral flux using noise- free band energies for given N frequency bands, wherein E- and E-_ denote the noise- free energies in j-th frequency band of successive frames of the audio signal.

Flux event is another flux-based feature based on spectral flux and marks instances when the spectral flux exceeds a certain threshold, indicating significant energy transitions. By identifying instances when the energy change surpasses a predefined level, this feature helps to differentiate between relevant (e.g., speech onset) and irrelevant fluctuations in energy, reducing false positives that might occur from minor or insignificant energy changes in the audio signal.

Preferably, the feature extractor 20 is operative to receive the spectral energies and/or the noise free spectral energies, wherein the received noise-free spectral energies or the received noise-free spectral energies and the received spectral energies are used to calculate an SNR estimate.

Preferably the feature extractor 20 is operative to receive the noise-free band energies computed using a spectral subtraction method based on a noise estimation method like minimum statistics, wherein the received noise-free spectral energies E- are used to calculate an SNR estimate over the total available frequency range by summing up the noise-free energies over all the frequency bands as shown in the equation below. Note that this method can only be used if the energies E- are given in the log domain.

Preferably the feature extractor 20 is operative to receive a power spectrum to extract a spectral shape-based feature, for example a spectral flatness measure (SFM) which helps measure the ‘tonality’ of a signal. This helps distinguish between speech, which typically has a more tonal or harmonic structure, and noise, which tends to have a more uniform or flat spectral distribution. The spectral flatness measure (SFM) represents a spectral shape of the audio input signal. SFM is computed over M frequency bins of the power spectrum using the following equation where P f_k) refers to the power spectrum value for the k -th discrete frequency bin.

The spectral flatness measure can be enhanced by dividing it with a crest factor which measures the ‘peakiness’ of the signal spectrum and is computed as a ratio between the highest peak in the spectrum divided by the root mean square (RMS) value of the spectrum.

Preferably the feature extractor 20 is operative to receive a pitch estimate being a result of a pitch search algorithm for determining a harmonic stability-based feature like pitch delta. Thus, the feature extractor 20 is operative to determine a harmonic stability feature depending on a received pitch estimate being a result of a pitch search algorithm The pitch search algorithm can, for example, output multiple pitch estimates for sub frame segments which can be used to compute a delta pitch value. Alternatively, two pitch deltas between consecutive subframes can be averaged to get an average pitch delta value for a frame.

Up to now, all features extracted by the feature extractor 20 are disclosed see Figs. 3 and 4.

As shown in Figs. 1 to 4, the codec pre-processor 60 is operative to transmit the pre-processed audio input signal 10 to the feature extractor 20. The pre-processed audio input signal 10 comprises of a spectral energy and/or a noise-free spectral energy in at least one frequency band of a plurality of frequency bands. The term “noise-free” here implies that noise energy has been removed from the signal energy for example through a spectral subtraction method based on a noise estimation method like minimum statistics, (see Figs 3 and 4). Preferably the codec pre-processor 60 is operative to determine spectral energy in at least one frequency band of a plurality of frequency bands and transmit the determined band energies (also called spectral energies herein) to a spectral subtractor 75 to compute noise-free spectral energies in at least one frequency band of a plurality of frequency bands using a spectral subtraction method, wherein the spectral subtractor 75 is operative to transmit the noise-free band energies to the feature extractor 20. The spectral subtractor 75 is operative to receive the band energies and then using a spectral subtraction method based on a noise estimation method to determine the ‘noise-free’ band energies. Using these the SNR value is then calculated during feature extraction by the feature extractor 20 as explained above. The subtraction method is based on a noise estimation method, for example the minimum statistics method.

The minimum statistics (MS) algorithm according to the present invention does not make use of any voice activity detector as is done in some noise estimation methods of the prior art [2], Rather, it identifies the minimum spectral values in each frequency band without differentiating between active speech and inactive parts. The main motivation behind the algorithm is that speech and noise are usually uncorrelated or statistically independent and that the signal power approaches the noise power level in short pauses between speech segments or syllables. Even though the MS algorithm tracks the running minimum and applies a time-varying smoothing factor to the minimum noise floor, the duration required to achieve a stable value is significantly less than adaptive algorithms for noise estimation mentioned in the prior art as it is not influenced by past SAD decisions and greatly depends on the minimum noise floor of the signal spectrum. For the MS algorithm used in the proposed invention, the minimum noise floor value is available for each frame and the minimum noise tracking is done over a window of 1.4 seconds. However, it is observed that despite the used window length for the minimum noise tracking, the SAD decision is stable after a preamble of 0.1 second.

The feature extractor 20 is coupled with a temporal context extractor 70. At least one of the features extracted by the feature extractor 20 is used to extract at least one temporal context feature. The temporal context extractor 70 extracts time-dependent context from at least one feature or the at least one further feature received from the feature extractor 20 and outputs meaningful context features or temporal context features. These context features include, but are not limited to, running mean and variance of the data, which help in understanding the temporal dynamics of the received features. Additionally, these context features significantly help in improving stability of the SAD decision. In particular, the feature extractor 20 is operative to transmit at least one of the extracted features and/or at least one of the further extracted features extracted by the feature extractor 20 to the temporal context extractor 70, wherein the temporal context extractor 70 is operative to extract a temporal context feature representing time-dependent context of the audio input signal

Preferably the feature extractor 20 is operative to transmit at least one of the features extracted by the feature extractor 20 (, i.e. at least one of the extracted features and/or at least one of the further extracted features) to the temporal context extractor 70, wherein the temporal context extractor 70 is operative to receive and to smooth the at least one extracted feature and/or the at least one further extracted feature and/or estimate the variance of the extracted at least one feature and/or of the at least one further extracted feature for obtaining the at least one temporal context feature representing time-dependent context of the audio input signal. The temporal context extractor 70 is operative to receive and to smooth or compute the running mean of the extracted at least one feature and/or estimate the variance of the extracted at least one feature using a recursive method, for example an HR filter, or an FIR filter for obtaining the at least one temporal context feature. The smoothed feature value and/or the estimate of the variance being a temporal context feature is then transmitted to the feature combiner 30 (see Figs 1 to 4).

Preferably the temporal context extractor 70 is operative to extract the temporal context feature of the at least one extracted feature or of the at least one further extracted feature using a digital filter, in particular an 11 R-filter or an Fl R-filter, to obtain a smoothed feature value and/or a corresponding estimate of a feature variance. The temporal context extractor 70 extracts a temporal context feature, where the ‘temporal context’ means timedependent context or information, for example the running mean and/or the estimated variance. For example, a first-order HR filter can be used to stabilize the neural network decision or the machine learning device decision without the use of additional memory. The smoothed running mean value for a frame x[n] can be calculated using the following equation where n is the frame index, f [n] is the feature value for the n-th frame and a is a constant, with value between 0 and 1 , that controls the “memory” of the filter by weighting the most recent feature value compared to the past mean value. x[n] = a * f[n] + (1 — a) * x[n — 1] When a is closer to 1, the filter gives more weight to the most recent input signal, making the output more responsive to recent changes in the input. This results in a less smoothed output signal, as the running mean is more influenced by recent fluctuations. For example, the value of a here is chosen as 0.142857. However, a different value of a can also be used.

The running variance v[n] can be computed in a similar way by using an HR filter through the following equations which compute the running mean of the squared differences of the feature value x[n] from the mean value f [n], e[n] is the squared difference of the feature value %[n] from the mean value f[n] and v[n] is the variance estimated by computing running mean of the squared difference computed in similar way as the running mean of the feature value. e[n] = | [n] — x[n]|² v[n] = a * e[n] + (1 — a) * v[n — 1]

Preferably the temporal context extractor 70 is operative to transmit the at least one temporal context feature representing at least one smoothed feature value and/or the corresponding estimate of the feature variance to the feature combiner 30 (see Figs. 1 to 4). The feature combiner 30 then uses the received at least one temporal context feature together with the at least one received extracted feature or the at least one further extracted feature and the combination weights received from the memory 40, to combine the features in order to output the activity output signal 50, also called SAD decision 50.

Preferably the feature combiner 30 is an artificial neural network with a feedforward structure, for example a Multi-Layer Perceptron (MLP). Alternatively, the feature combiner 30 might be another machine learning device or model. The feature combiner is adapted to receive static combination weights which are determined by an automatic training procedure using a desired training set. The combination weights are referred to as “static” as they remain constant after the training procedure is complete as opposed to the adaptive algorithms in the prior art where the parameters are adapted with changing signal conditions. These weights would remain static during the application of the model. The weights change only when the training process is repeated for a different application, with a different training set for example. The SAD algorithm described within the present invention uses a feature combiner which is modelled using a simple neural network, for example an MLP model, having one hidden layer of 20 neurons and a single output for binary classification. The number of hidden layers used, and the size of the layers can also be different. The above computed features along with the temporal context features are given as an input to the feature combiner or the neural network. Before providing the features as input to the feature combiner, each feature can be normalized to ensure a mean of zero and a standard deviation of one. This normalization process is beneficial for optimizing the performance of the machine learning model, as it helps in accelerating the convergence of the training algorithm and achieving a more stable learning process. The neural network outputs probability values, for example between 0 and 1, indicating the probability of presence of an active signal on which a static threshold, for example 0.5, can be applied for binary classification. The model described here is trained to detect signal activity and inactivity in four types of data, namely, clean speech, noisy speech, speech over background music and music in general as shown in Fig. 7, where speech and music are considered as ‘active’ signals and the rest, for example noise, are considered ’inactive’. The definition of ‘active’ and ‘inactive’ can also be slightly changed depending upon the goal of the application and the training procedure can be performed again with a new dataset. As can be seen in Fig. 7, the mean error rate for SAD is generally comparable to that of EVS VAD (known in the prior art/ EVS short for Enhanced Voice Services [5]) over most categories of signals and is even smaller in case of noisy speech. Fig. 7 shows an overview of the performance for the SAD of the present invention compared with the EVS VAD known in the prior art, wherein the focus should be laid on the mean error rate, showing that the overall performance of the SAD of the present invention is comparable to that of the EVS VAD while having the added advantage of fast reactivity and requiring negligible preamble time. The noisy speech dataset consists of different noisy environments including but not limited to car, street and office noise with different SNR levels for speech. Due to the automatic training procedure, the SAD has the capability to adapt to new noise environments which potentially opens doors for new applications. The overall performance for the current SAD according to the present invention as compared with the EVS VAD is shown in Fig. 7.

The SAD outputs a binary decision of ‘active’ or ‘inactive’ per frame of 20 milliseconds duration, for example. The neural network model, for example the MLP model works in a feedforward way, i.e., the process flow occurs in one direction between the feature inputs and the output and there is no past information stored or used by the model itself. Additionally, a static or dynamic hangover can be applied to the SAD output for a smoothed decision. A static hangover is also acceptable in this case as the requirements for a communication codec are to have as minimum mis-detections as possible.

The SAD system works on a frame-by-frame basis which makes it suitable to be used in real-time applications.

Figs. 5a, b, each shows a comparison of the activity output signal of a EVS decision 510 known by the prior art compared to the SAD active decision 520 according to the present invention, when an active signal is present and when the initialization period is 3 sec (Fig. 5a) or when the initialization period is 0.1 sec (Fig. 5b). The waveform 500 of the active signal is shown as normalized amplitude over time in seconds. Without a long preamble, the SAD algorithm in EVS (see reference sign 410) is not initialized properly and has many mis-detections on an active signal which is not the case with the activity decision output 520 of the SAD of the present invention. Figure 5 is comparing the EVS VAD and the proposed SAD behavior for the preamble length of 3 sec (Fig. 5a) and 0.1 sec (Fig. 5b). It can be seen that with preamble length of 0.1 sec the active decision output 520 of the neural network model shows nearly no mis-detection, i.e. the SAD algorithm according to the present invention has been improved. Note that here the SAD algorithm in EVS and EVS VAD refer to the same algorithm in the EVS codec.

Figs. 6a, b, each shows a comparison of the activity output signal of the EVS (enhanced voice services) VAD decision 510 known by the prior art compared to the SAD active decision 520 according to the present invention, when the signal conditions are changing fast (Fig. 6b) and when there is time between change of signal conditions which is indicated by a silence break between the signals (Fig. 6a). Figs. 6a, b are depicting the behaviour in fast changing signal conditions. Each of the Figs. 6a, b show two active signals combined without enough break in between in Fig. 6b and in Fig. 6a there is some silence added. So, the line 510 derived from using EVS is not able to adapt quickly when the signals conditions are changing suddenly and as a result it gives an unstable output while the line 620 derived from using the present invention stays stable (compare line 510 and 520 in Fig. 6b). The waveform 500 of the detected speech signal is also shown as normalized amplitude over time in seconds. The wave signal 500 comprises two different waveforms 500a, 500b which are considered as active signals. It can be seen that without enough time for adaptation (indicated by 550 in Fig 6a), the SAD algorithm in EVS (see reference sign 510 in Fig. 6b) is not initialized properly and has mis-detections on an active signal which is not the case with the output 510 in Fig. 6a where the signal characteristics change after a significant period of time. The decision output 520 of the MLP model in both figures is stable irrespective of the time after which the signal changes., i.e. the SAD algorithm according to the present invention has been improved.

Because an automated training framework for the apparatus /method of processing a pre- processed audio signal 10 according to the present invention is available, the optimization process requires minimal manual effort as opposed to frequent tuning that is required by the adaptive mechanisms described in the prior art. The model can also be easily optimized over new stationary or non-stationary noise types by a simple training procedure.

According to a second aspect of the present invention, a method 800 of processing a pre- processed audio signal 10 to obtain an activity output signal 50 is proposed. A flow chart of method 800 is illustrated in Fig. 8. The method 800 comprises in step 810 receiving, by a feature extractor 20, the pre-processed audio input signal 10 comprising a spectral energy and/or a noise-free energy in at least one frequency band of a plurality of frequency bands. The method 800 comprises in step 820 extracting, by the feature extractor 20, at least one feature and/or at least one further feature, each depending on the spectral energy and/or depending on the noise-free spectral energy in the at least one frequency band of the plurality of frequency bands. The method 800 comprises in step 830 receiving, by a temporal context extractor 70, the at least one extracted feature, wherein the temporal context extractor 70 extracts at least one temporal context feature based on the at least one extracted feature. Alternatively, the method 800 comprises in step 830 receiving, by the temporal context extractor 70, the at least one further extracted feature, wherein the temporal context extractor 70 extracts at least one temporal context feature based on the at least one further extracted feature. The method 800 comprises in step 840 receiving, by a feature combiner 30 coupled with the feature extractor 20 and coupled with the temporal context extractor 70, the at least one extracted feature or the at least one further extracted feature and the at least one temporal context feature and a plurality of static combination weights for finally performing step 850. Step 850 comprises combining the at least one extracted feature or the at least one further extracted feature and the at least one temporal context feature using the plurality of static combination weights to obtain the activity output signal 50. The step 810, 820, 830, 840, and 850 should be performed in this order. Preferably, the method 800 is performed using the apparatus described herein. The feature extractor 20 is receiving in step 810 the pre-processed audio input signal 10, wherein the pre-processed audio input signal 10 comprises a spectral energy and/or a noise-free spectral energy in at least one frequency band of a plurality of frequency bands, wherein the feature extractor 20 is adapted to extract at least one feature depending on the spectral energy and/or depending on the noise-free spectral energy. The input to the feature extractor 20 is the pre-processed audio input signal 10. The pre-processed audio input signal 10 is the output of the codec pre-processor 60. The pre-processed audio input signal 10 refers to the output of the codec-preprocessor 60. The temporal context extractor 70 is receiving in step 830 the at least one extracted feature, wherein the temporal context extractor extracts at least one temporal context feature based on the at least one extracted feature, or the temporal context extractor 70 is receiving the at least one further extracted feature, wherein the temporal context extractor extracts at least one temporal context feature based on the at least one further extracted feature, The feature combiner 30, coupled with the feature extractor 20 and coupled with the temporal context extractor, is receiving in step 840 the at least one extracted feature or the at least one further extracted feature and the at least one temporal context feature and a plurality of static combination weights for combining in step 850 the at least one extracted feature or the at least one further extracted feature and the at least one temporal context feature using the plurality of static combination weights to obtain the activity output signal 50

Preferably the method 800 comprises coupling the feature extractor 20 with the codec pre-processor 60, and pre-processing, by the codec pre-processor 60, an audio input signal 10’ to obtain the pre-processed audio input signal 10. The pre-processed audio input signal 10 comprises information characteristic of the audio input signal, e.g. a power spectrum that represents the spectral energy distribution across various frequencies of the audio input signal, and/or band energies which refer to spectral energies in specific frequency bands, and/or noise-free band energies which refer to noise-free spectral energies in specific frequency bands and/or a pitch estimate which gives information about the fundamental frequency and harmonicity of the audio signal.

Preferably the method 800 comprises determining, by the codec pre-processor 60, the spectral energies (also called herein frequency band energies) in the at least one frequency band of the plurality of frequency bands, transmitting the determined spectral energies to a spectral subtractor 75; computing, by the spectral subtractor 75, the noise-free energies in the at least one frequency band of the plurality of frequency bands using a spectral subtraction method, and transmitting, by the spectral subtractor 75, the noise-free energies to the feature extractor 20 (see Figs. 3 and 4). The spectral subtractor 75, which is operative to receive the frequency band energies and then using a spectral subtraction method to determine the “noise-free” band energies. Using these the SNR value is then calculated during feature extraction by the feature extractor 20. The spectral subtraction method is based on a noise estimation method, for example a minimum statistics method.

Preferably the method 800 comprises coupling the feature extractor 20 with a temporal context extractor 70, transmitting, by the feature extractor 20, the at least one of the features extracted by the feature extractor 20 or the at least one of the further features extracted by the feature extractor 20 to the temporal context extractor 70, receiving and smoothing, by the temporal context extractor 70, the extracted at least one feature or the extracted at least one further feature and/or estimate the variance of the extracted at least one feature or of the extracted at least one further feature for obtaining at least one temporal context feature. Finally, the method 800 comprises transmitting, by the temporal context extractor 70, the at least one temporal context feature representing at least one smoothed feature value and/or the corresponding estimate of the variance of the extracted at least one feature to the feature combiner 30. It is referred to the details of the temporal context extractor 70 disclosed above, which will not be repeated here.

According to a third aspect of the present invention, a computer program for performing, when running on a computer, a method described herein is proposed.

The disclosed technical teaching of the present application has a low complexity. According to the disclosed technical teaching, the features are extracted through pre-processed mechanisms already implemented for the general codec. A simple neural network (MLP) with one hidden layer can be used to perform the proposed technical teaching. The proposed technical teaching requires only a short an automated training procedure which makes it easier to train on new noise conditions or music. Furthermore, no long-term adaptive mechanism is used for feature extraction. Instead, simple feed-forward neural network architecture with no dependance on past SAD decisions is used and an open loop system design without adaptive mechanism is used. An advantage of the present invention is its negligible start-up time. The present invention is reactive and accurate under fast changing conditions, which is shown in Figs. 5, 6 and 7. The feature combination of the present invention is unique and brings a clear benefit compared to the state-of-the-art.

The present invention has the following advantages: A simple MLP-based neural network allows the use of static weights obtained after training. The model outputs a probability value indicating the presence of an active signal on which a static threshold can be applied for classification. As the window length requirement for MS algorithm for SNR estimation is significantly lower and there is no dependence on past output decision of the model, the major advantage of using this SAD model is that it works well even with a minimum preamble of 0.1 second.

Although some aspects have been described as features in the context of an apparatus it is clear that such a description may also be regarded as a description of corresponding features of a method. Although some aspects have been described as features in the context of a method, it is clear that such a description may also be regarded as a description of corresponding features concerning the functionality of an apparatus. For reasons of redundancy, the present invention is not described in full in the form of process steps an apparatus features.

Depending on certain implementation requirements, embodiments of the proposed apparatus can be implemented in hardware or in software or at least partially in hardware or at least partially in software. Generally, embodiments of the apparatus can be implemented as a computer program product with a program code, the program code being operative for performing one of the methods when the computer program product runs on a computer. The program code may for example be stored on a machine readable carrier.

In the preceding detailed description, different features have sometimes been grouped together in examples in order to rationalize the disclosure. This type of disclosure should not be interpreted as an intention that the claimed examples have more features than are expressly recited in each claim. Rather, as the following claims reflect, subject matter may be found in fewer than all of the features of a single disclosed example. Consequently, the following claims are hereby incorporated into the detailed description, and each claim may stand as its own separate example. While each claim may stand as its own separate example, it should be noted that although dependent claims in the claims refer back to a specific combination with one or more other claims, other examples also include a combination of dependent claims with the subject matter of any other dependent claim or a combination of any feature with other dependent or independent claims. Such combinations are encompassed unless it is stated that a specific combination is not intended. It is further intended that a combination of features of a claim with any other independent claim is also encompassed, even if that claim is not directly dependent on the independent claim. References

[1] Martin, Rainer. "Noise power spectral density estimation based on optimal smoothing and minimum statistics." IEEE Transactions on speech and audio processing 9.5 (2001): 504-512.

[2] Cornu, Etienne, et al. "ETSI AMR-2 VAD: evaluation and ultra low-resource implementation." 2003 IEEE International Conference on Acoustics, Speech, and Signal Processing, 2003. Proceedings. (ICASSP'03).. Vol. 2. IEEE, 2003.

[3] ETSI EN 301 708: Voice Activity Detector (VAD) for Adaptive Multi-Rate (AMR), 1998

[4] Malenovsky, Vladimir, and Milan Jelinek. "Improving the detection efficiency of the VMR-WB VAD algorithm on music signals." 2008 16^th European Signal Processing Conference. IEEE, 2008.

[5] Bruhn, Stefan, et al. "Standardization of the new 3GPP EVS codec." 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2015.

[6] Benyassine, A., et al. "A silence compression scheme for use with G. 729 optimized for V. 70 digital simultaneous voice and data applications (recommendation G. 729 annex B)." IEEE Commun. Mag 35.9 (1997): 64-73.

[7] Kara, Fatih, Tuba Islam, and Hasan Palaz. "Voice activity detection based on minimum statistics." Proceedings of the IEEE 12th Signal Processing and Communications Applications Conference, 2004.. IEEE, 2004.

[8] Davis, Alan, Sven Nordholm, and Roberto Togneri. "Statistical voice activity detection using low-variance spectrum estimation and an adaptive threshold." IEEE Transactions on Audio, Speech, and Language Processing 14.2 (2006): 412-424.

[9] Freudenberger, Jurgen, and Sebastian Stenzel. "Time-frequency dependent voice activity detection based on a simple threshold test." 2011 IEEE Statistical Signal Processing Workshop (SSP). IEEE, 2011.

Claims

1. Apparatus (100) for processing a pre-processed audio input signal (10) to obtain an activity output signal (50), the apparatus comprising: a feature extractor (20) adapted to receive the pre-processed audio input signal (10), wherein the pre-processed audio input signal (10) comprises a spectral energy and/or a noise-free spectral energy in at least one frequency band of a plurality of frequency bands, wherein the feature extractor (20) is adapted to extract at least one feature and/or at least one further feature, each depending on the spectral energy and/or depending on the noise-free spectral energy, a temporal context extractor (70), adapted to receive the at least one extracted feature, wherein the temporal context extractor is adapted to extract at least one temporal context feature based on the at least one extracted feature, or the temporal context extractor (70), adapted to receive the at least one further extracted feature, wherein the temporal context extractor (70) is adapted to extract at least one temporal context feature based on the at least one further extracted feature, and a feature combiner (30) coupled with the feature extractor (20) and coupled with the temporal context extractor (70), and the feature combiner (30) adapted to receive the at least one extracted feature and/or the at least one further extracted feature and the at least one temporal context feature and a plurality of static combination weights for combining the at least one extracted feature and/or the at least one further extracted feature and the at least one temporal context feature using the plurality of static combination weights to obtain the activity output signal (50).

2. Apparatus (100) according to claim 1 , wherein the pre-processed audio input signal (10) comprises a power spectrum, and/or a pitch estimate.

3. Apparatus (100) according to claims 1 or 2, wherein the feature combiner (30) is coupled to a memory (40) storing the plurality of static combination weights which are to be received by the feature combiner (30).

4. Apparatus (100) according to one of the preceding claims, wherein the feature extractor (20) is coupled with a codec pre-processor (60), wherein the codec pre-processor (60) is operative to pre-process an audio input signal (10) to obtain the pre- processed audio input signal (10).

5. Apparatus (100) according to one of the preceding claims, wherein the feature extractor (20) is operative to receive the spectral energies and/or the noise-free spectral energies and to combine the spectral energies and/or the noise-free spectral energies into spectral energies and/or noise-free spectral energies in at least one larger frequency band of a plurality of larger frequency bands.

6. Apparatus (100) according to one of the preceding claims, wherein the feature extractor (20) is operative to extract a spectral flux and/or at least one feature based on the spectral flux, wherein spectral flux represents a difference of energy in successive frames of the audio input signal.

7. Apparatus (100) according to one of the preceding claims 2 to 6, wherein the feature extractor (20) is operative to receive the spectral energies and/or the noise free spectral energies, wherein the received noise-free spectral energies or the received noise-free spectral energies and the received spectral energies are used to calculate an SNR estimate.

8. Apparatus (100) according to one of the preceding claims, wherein the feature extractor (20) is operative to receive the power spectrum to compute a spectral flatness measure representing a spectral shape of the audio input signal.

9. Apparatus (100) according to one of the preceding claims, wherein the feature extractor (20) is operative to determine a harmonic stability feature depending on a received pitch estimate being a result of a pitch search algorithm.

10. Apparatus (100) according to one of the preceding claims 3 to 9, wherein the codec pre-processor (60) is operative to transmit the pre-processed audio input signal (10) to the feature extractor (20).

11. Apparatus (100) according to one of the preceding claims 4 to 10, wherein the codec pre-processor (60) is operative to determine the spectral energies in the at least one frequency band of the plurality of frequency bands and transmit the determined spectral energies to a spectral subtractor (75) to compute the noise-free spectral energies using a spectral subtraction method, wherein the spectral subtractor (75) is operative to transmit the noise-free spectral energies in the at least one frequency band of the plurality of frequency bands to the feature extractor (20).

12. Apparatus (100) according to one of the preceding claims, wherein the feature extractor (20) is operative to transmit the at least one extracted feature and/or the at least one further extracted feature extracted by the feature extractor (20) to the temporal context extractor (70), wherein the temporal context extractor (70) is operative to extract a temporal context feature representing time-dependent context of the audio input signal.

13. Apparatus (100) according to one of the preceding claims, wherein the feature extractor (20) is operative to transmit at least one of the extracted features and/or at least one of the further extracted features extracted by the feature extractor (20) to the temporal context extractor (70), wherein the temporal context extractor (70) is operative to receive and to smooth the at least one extracted feature and/or the at least one further extracted feature and/or estimate the variance of the extracted at least one feature and/or of the at least one further extracted feature for obtaining the at least one temporal context feature representing time-dependent context of the audio input signal.

14. Apparatus (100) according to one of the preceding claims, wherein the temporal context extractor (70) is operative to extract the temporal context feature of the at least one extracted feature and/or the at least one further extracted feature using a digital filter, in particular an 11 R-filter or an Fl R-filter, to obtain the smoothed feature value and/or the corresponding estimate of the feature variance.

15. Apparatus (100) according to one of the preceding claims 12 to 14, wherein the temporal context extractor (70) is operative to transmit the at least one temporal context feature representing the at least one smoothed feature value and/or the corresponding estimate of the feature variance to the feature combiner (30).

16. Apparatus (100) according to one of the preceding claims, wherein the feature combiner (30) is a neural network or a machine learning device.

17. Method of processing a pre-processed audio signal (10) to obtain an activity output signal (50), comprising: receiving, by a feature extractor (20), the pre-processed audio input signal (10) comprising a spectral energy and/or a noise-free spectral energy in at least one frequency band of a plurality of frequency bands, extracting, by the feature extractor (20), at least one feature and/or at least one further feature, each depending on the spectral energy and/or depending on the noise-free spectral energy in the at least one frequency band of the plurality of frequency bands, receiving, by a temporal context extractor (70), the at least one extracted feature, wherein the temporal context extractor extracts at least one temporal context feature based on the at least one extracted feature, or receiving, by the temporal context extractor (70), the at least one further extracted feature, wherein the temporal context extractor (70) extracts at least one temporal context feature based on the at least one further extracted feature, receiving, by a feature combiner (30) coupled with the feature extractor (20) and coupled with the temporal context extractor (70), the at least one extracted feature and/or the at least one further extracted feature and the at least one temporal context feature and a plurality of static combination weights for combining the at least one extracted feature or the at least one further extracted feature and the at least one temporal context feature using the plurality of static combination weights to obtain the activity output signal (50).

18. Method of claim 17, wherein the pre-processed audio input signal (10) comprises a power spectrum, and/or a pitch estimate.

19. Method of claims 17 or 18, comprising: coupling the feature extractor (20) with a codec pre-processor (60), pre-processing, by the codec pre-processor (60), an audio input signal (10a) to obtain the pre-processed audio input signal (10).

20. Method of one of the claims 17 to 19, comprising: determining, by the codec pre-processor (60), the spectral energies in the at least one frequency band of the plurality of frequency bands, transmitting the determined energies to a spectral subtractor (75); computing, by the spectral subtractor (75), the noise-free spectral energies in the at least one frequency band of the plurality of frequency bands using a spectral subtraction method, and transmitting, by the spectral subtractor (75), the noise-free spectral energies to the feature extractor (20).

21. Method of one of the claims 17 to 20, comprising: coupling the feature extractor (20) with a temporal context extractor (70), transmitting, by the feature extractor (20), the at least one of the extracted features and/or at least one of the further extracted features extracted by the feature extractor (20) to the temporal context extractor (70), receiving and smoothing, by the temporal context extractor (70), the at least one extracted feature and/or at least one of the further extracted feature and/or estimate the variance of the at least one extracted feature and/or of the at least one further extracted feature for obtaining at least one temporal context feature, and transmitting, by the temporal context extractor (70), the at least one temporal context feature representing at least one smoothed feature value and/or the corresponding estimate of the variance of the at least one extracted feature and/or of the at least one further extracted feature to the feature combiner (30).

22. Computer program product for performing, when running on a computer, a method in accordance with one of the claims 17 to 21.