CN113327634B - A voice activity detection method and system for low power consumption circuit - Google Patents

Abstract

The invention discloses a voice activity detection method and a voice activity detection system applied to a low-power-consumption circuit, wherein the method comprises the following steps: receiving input voice, extracting voice characteristics, adopting sub-band energy characteristics, and reducing the number of sub-bands; extracting the characteristics to obtain characteristic values, and entering the subsequent classification; and selecting a linear classifier of the support vector machine to carry out training classification, outputting a voice classification result, and completing voice activity detection. The invention provides a simpler voice activity detection method, which adopts lower implementation complexity to obtain better classification accuracy under low signal-to-noise ratio; meanwhile, the voice characteristic extraction part is realized by adopting a full analog circuit, so that the possibility of realizing an actual circuit is considered in design, and the low-power consumption circuit requirement can be additionally met.

Description

A voice activity detection method and system for low power consumption circuit

Technical Field

The present invention relates to the field of speech algorithm design, and in particular to a speech silence detection method and system based on linear SVM (support vector machine).

Background technique

Voice Activity Detection (VAD), also known as voice endpoint detection, is to determine the starting and ending positions of speech. It was first used in communication fields such as telephone transmission and detection. Today, it is also widely used in speech recognition and speech compression. It is an important speech preprocessing technology.

The existing speech endpoint detection algorithms can be divided into time domain feature VAD algorithms and frequency domain feature VAD algorithms according to the speech features used for classification. The time domain features include: short-time energy, short-time zero-crossing rate, short-time autocorrelation, etc.; the frequency domain features include: pitch period, Mel cepstrum distance, etc. At the same time, speech endpoint detection is actually a binary classification problem, that is, the speech signal is classified according to the difference in the characteristics of the speech segment and the noise segment in the time domain or frequency domain. Therefore, different VAD algorithms may use different classifiers. For example, the most classic dual threshold detection algorithm uses threshold judgment, and there are also algorithms based on more complex classifiers such as decision trees, finite state machines or neural networks for classification.

Speech features need to be able to better reflect the difference between speech and noise; time domain features have better results in high signal-to-noise ratio conditions, but in a noisy environment, the noise will drown out the speech signal, resulting in errors in judgments based on energy or zero-crossing rate features. Frequency domain features are less affected by noise than time domain features to a certain extent, but the computational complexity of frequency domain features is higher than that of time domain features.

Summary of the invention

In order to compromise the contradiction between the computational complexity and accuracy of the prior art and achieve better classification accuracy with lower algorithm complexity, the present invention proposes a voice activity detection method and system applied to a low power consumption circuit.

The technical problem of the present invention is solved by the following technical solutions:

The present invention proposes a voice activity detection method applied to a low-power circuit, which is characterized in that it includes the following steps: S1: receiving input voice, performing voice feature extraction, using sub-band energy features, and reducing the number of sub-bands; after feature extraction, obtaining feature values for subsequent classification; S2: selecting a linear classifier of a support vector machine for training and classification, outputting voice classification results, and completing voice activity detection.

In some embodiments, in step S1, the speech is divided into frames according to the frame length, and the frame shift is equal to the frame length; a rectangular window is used to perform a windowing operation on the speech signal to implement the framing and windowing operation, and the window length is the number of data points corresponding to the frame length.

In some embodiments, after the frame splitting and windowing operations, band-pass filtering is performed and short-time energy is calculated.

In some embodiments, in step S1, the speech feature extraction is implemented using a full analog circuit.

In some embodiments, in step S1, a background noise feature based on recursive average estimation is added as a new feature.

In some embodiments, the background noise based on the recursive average estimation is calculated as follows:

In the formula, β ₁ and β ₂ are between 0 and 1; NL(i) and E(i) are the values of the i-th

for_E(i)<NL(i-1):

NL(i)＝β ₂ NL(i-1)+(1-β ₂ )E(i)

The background noise and short-time energy of the frame signal; the value of the smoothing factor β is determined by the threshold method, and different β values are selected according to the size relationship between NL(i-1) and E(i).

In some embodiments, the analog domain characteristics are quantized and converted into 8-bit digital domain characteristics.

In some embodiments, the weight values of the linear classifier of the support vector machine are limited in number of bits to reduce complexity.

The present invention also proposes a voice activity detection system applied to a low-power circuit, characterized in that it includes: a feature extraction module and a classification module; the feature extraction module receives input voice, performs voice feature extraction, and uses sub-band energy and background noise based on recursive average estimation as classification features; the classification module uses a linear classifier of a support vector machine for training and classification; after passing through the feature extraction module, the characteristic values obtained enter the subsequent classification module for training and classification, output the voice classification results, and complete the voice activity detection.

The present invention further proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and is characterized in that when the computer program is executed by a processor, the steps of any of the above methods are implemented.

Compared with the prior art, the beneficial effects of the present invention include: compared with the traditional algorithm using sub-band energy features, the number of sub-bands in the present invention is greatly reduced, thereby reducing the complexity of the algorithm; in terms of the classifier, SVM, which compromises accuracy and complexity, is used as the classifier; at the same time, the introduction of NL features ensures that the accuracy does not drop significantly compared with the traditional algorithm; the present invention uses a lower implementation complexity to achieve better classification accuracy under low signal-to-noise ratio.

In some embodiments, the beneficial effects of the present invention compared with the prior art include: since the speech feature extraction part of the method is implemented using a full analog circuit, the possibility of actual circuit implementation is taken into consideration during the design, so that the method can also meet the low-power circuit requirements at the same time.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of a processing flow of an embodiment of the present invention;

FIG2 is a speech schematic diagram of a feature extraction module according to an embodiment of the present invention;

FIG. 3 is a speech schematic diagram of a classification module according to an embodiment of the present invention.

Detailed ways

The present invention is further described below with reference to the accompanying drawings and in combination with preferred embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

It should be noted that the directional terms such as left, right, up, down, top, and bottom in this embodiment are merely relative concepts or are based on the normal use status of the product, and should not be considered as restrictive.

The purpose of the present invention is to propose a relatively simple voice activity detection method and system, which can still have a good classification accuracy under low signal-to-noise ratio conditions; at the same time, because the voice feature extraction part of the method can be implemented using a full analog circuit, the method can also additionally meet the low-power circuit requirements.

The processing flow of the voice activity detection method of the embodiment of the present invention is shown in Figure 1, which consists of a feature extraction module and a classification module, wherein the feature extraction module uses subband energy and background noise NL based on recursive average estimation as classification features; the classification module uses linear SVM as a classifier.

A voice activity detection method applied to a low-power circuit according to an embodiment of the present invention comprises the following steps: S1: receiving input voice, performing voice feature extraction, using sub-band energy features, and reducing the number of sub-bands; after feature extraction, obtaining feature values for subsequent classification; S2: selecting a linear classifier of a support vector machine for training and classification, outputting voice classification results, and completing voice activity detection.

In step S1, the speech is divided into frames according to the frame length, and the frame shift is equal to the frame length; a rectangular window is used to perform a windowing operation on the speech signal to implement the framing and windowing operation, and the window length is the number of data points corresponding to the frame length.

In the frame-by-frame windowing operation, bandpass filtering is performed and short-time energy is calculated.

In step S1, the speech feature extraction is implemented using a full analog circuit.

In step S1, the background noise feature based on recursive average estimation is added as a new feature.

The processing flow of this method is as follows:

The first stage: divide the speech into frames with a frame length of 25ms. At the same time, if the frame shift is less than the frame length, the circuit must be implemented by adding a digital circuit to the register, which is not convenient for low-power analog circuit implementation. Therefore, in this design, the frame shift is equal to the frame length, that is, there is no overlap between frames. When implementing the method, a rectangular window is used to perform a windowing operation on the speech signal to realize the framing operation. The window length is the number of data points corresponding to 25ms. According to the sampling rate of 16k for general speech signals, the window length N is 400.

In the speech feature selection part, we use subband energy features, but greatly reduce the number of subbands, and only select 4 frequency bands between 100-5khz in the sound spectrum to calculate short-time energy. This reduces the classification accuracy to a certain extent. As compensation, we add background noise estimation NL features as new features. The background noise calculation method based on recursive average estimation is as follows: The special

for_E(i)<NL(i-1):

NL(i)＝β ₂ NL(i-1)+(1-β ₂ )E(i)

The feature is obtained by recursive averaging based on time constants. Recursive averaging is a commonly used method in the field of speech enhancement, and the calculation object is mostly the signal-to-noise ratio SNR. Recursive averaging can reduce mutations and make the calculation target change more smoothly. The above formula is to calculate the background noise NL by recursive averaging. In the formula, β ₁ and β ₂ are between 0 and 1; NL(i) and E(i) are the background noise and short-time energy of the i-th frame signal, and β is the smoothing factor. In the recursive averaging calculation, the value of the smoothing factor adopts the threshold method, that is, different β values are selected according to the size relationship between NL(i-1) and E(i). Due to the characteristics of the recursive averaging mentioned above, this feature can reduce the fluctuation of the feature value in the transition zone within the speech segment (that is, the part where the signal amplitude suddenly drops in Figure 2, but it is still the speech part) to a certain extent, thereby reducing misjudgment.

After passing through the feature extraction module, the 5 sets of feature values obtained are quantized and then sent to the subsequent classification module for training and classification.

Phase II: In terms of classifiers, we choose the linear classifier SVM of support vector machines. Compared with other types of classifiers, the linear classifier SVM of support vector machines has higher accuracy than domain value judgment and decision tree; compared with DNN deep neural network, the linear classifier SVM of support vector machines has smaller complexity. Among SVM classifiers, there are SVM with kernel function and linear SVM. After algorithm simulation and comparison, for the selected speech features, we found that the effect of linear SVM is very similar to that of kernel function, but the circuit implementation is simpler and the power consumption is lower, so the linear SVM classifier was finally selected.

At the same time, considering the circuit implementation part, we quantize the analog domain features and convert them into 8-bit digital domain features. At the same time, the weight value w of the linear SVM classifier is limited to the number of bits to reduce the complexity. If the classification result of each frame is judged to be correct according to the same points between the labels provided by the data set, as shown in Figure 3, the proportion of correct frames to the total number of frames is the classification accuracy. If the noise segment and the speech segment are not considered to be calculated separately, the algorithm can finally achieve a classification accuracy of 90% under a 10db signal-to-noise ratio; under a 5db signal-to-noise ratio, it also has a classification accuracy of 85%. If the classification accuracy is considered separately from the noise segment and the speech segment, the accuracy of the noise part is 89.87% and the accuracy of the speech part is 91.55% under a 10db signal-to-noise ratio.

A voice activity detection system applied to a low-power circuit according to an embodiment of the present invention comprises: a feature extraction module and a classification module; the feature extraction module receives input voice, performs voice feature extraction, and uses subband energy and background noise based on recursive average estimation as classification features; the classification module uses a linear classifier of a support vector machine for training and classification; after passing through the feature extraction module, feature values are obtained and entered into a subsequent classification module for training and classification, and a voice classification result is output to complete voice activity detection.

A computer-readable storage medium according to an embodiment of the present invention stores a computer program, wherein the computer program implements the steps of any of the above methods when executed by a processor.

The voice activity detection method proposed in the present invention has achieved good classification accuracy under low signal-to-noise ratio with relatively low implementation complexity; at the same time, the possibility of actual circuit implementation is taken into consideration during the design, so that the feature extraction part of the method can be implemented with a full analog circuit to meet low power consumption requirements.

The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims (9)

1. A voice activity detection method for a low power circuit, comprising the steps of:

S1: receiving input voice, extracting voice characteristics, adopting sub-band energy characteristics, reducing the number of sub-bands, adding background noise characteristics based on recursive average estimation as new characteristics, wherein 4 frequency bands are selected only between 100-5khz in a voice frequency spectrum to calculate short-time energy, and the background noise characteristics based on the recursive average estimation are used as compensation to reduce fluctuation of characteristic values of a transition region in a voice section, namely a voice part with suddenly reduced signal amplitude; extracting the characteristics to obtain characteristic values, and entering the subsequent classification;

s2: and selecting a linear classifier of the support vector machine to carry out training classification, outputting a voice classification result, and completing voice activity detection.

2. The voice activity detection method for low power consumption circuits according to claim 1, wherein: in the S1 step, the voice is divided into frames according to the frame length, and the frame length is equal to the frame length; and a rectangular window is adopted to carry out windowing operation on the voice signal to realize framing windowing operation, and the window length is the number of data points corresponding to the frame length.

3. The voice activity detection method applied to a low power consumption circuit as claimed in claim 2, wherein: after framing and windowing operations, bandpass filtering is performed to calculate short-time energy.

4. The voice activity detection method for low power consumption circuits according to claim 1, wherein: in step S1, the speech feature extraction is implemented using a full analog circuit.

5. The voice activity detection method for low power consumption circuits according to claim 1, wherein: the background noise based on the recursive average estimation is calculated as follows: The value of beta ₁,β₂ in the formula is between 0 and 1; NL (i), E (i) is the background noise and short-time energy of the i-th frame signal; the value of the smoothing factor beta adopts a threshold method, and different beta values are selected according to the magnitude relation between NL (i-1) and E (i).

6. The voice activity detection method for low power consumption circuits according to claim 1, wherein: and quantizing the analog domain features, and converting the analog domain features into 8-bit digital domain features.

7. The voice activity detection method for low power consumption circuits according to claim 1, wherein: and carrying out bit number limitation on the weight value of the linear classifier of the support vector machine so as to reduce complexity.

8. A voice activity detection system for use in a low power circuit, comprising:

the device comprises a feature extraction module and a classification module;

The feature extraction module receives input voice, performs voice feature extraction, and adopts subband energy and background noise estimated based on recursive average as classification features; wherein, only selecting 4 frequency bands between 100-5khz in the sound spectrum to calculate short-time energy, and using the background noise characteristic estimated based on recursive average as compensation to reduce the fluctuation of the characteristic value by the transition zone in the voice section, namely the voice part with suddenly reduced signal amplitude;

The classification module adopts a linear classifier of a support vector machine to carry out training classification;

After the characteristic extraction module, the obtained characteristic value enters a subsequent classification module to carry out training classification, and a voice classification result is output to finish voice activity detection.

9. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the method according to any of claims 1-8.