Noise robust acoustic feature extraction method based on gamma channel scaled basis vector

Abstract

The application discloses a noise-robust acoustic feature extraction method based on a gamma pass scaling basis vector. The noise-robust acoustic feature extraction method based on the gamma pass scaling basis vector comprises the following steps: performing pre-emphasis processing on a speech signal; performing frame processing on the speech signal after the pre-emphasis processing; performing Fourier transform on the speech signal after the frame processing; calculating a scaling coefficient according to the spectral distribution characteristics of a gamma pass filter bank, and optimizing a basis vector based on the scaling coefficient; and performing discrete cosine transform on the speech signal after the Fourier transform and the optimized basis vector to extract noise-robust acoustic features from the speech signal. According to the application, the scaling coefficient is calculated based on the frequency domain distribution characteristics of the gamma pass filter bank, and is directly applied to the basis vector for generating acoustic features, so that the original details of the speech signal are retained to the greatest extent, the information carrying capacity of the acoustic features under the interference of noise signals is ensured, and the speech recognition effect can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of speech recognition technology, and in particular to a noise-robust acoustic feature extraction method based on gamma-pass scaling basis vectors. Background Technology

[0002] Speech recognition technology has been widely used in today's society, such as intelligent voice search services in mobile terminals, smart home control based on voice commands, and intelligent transcription functions in conference systems. However, how to maintain the high recognition accuracy of intelligent speech recognition systems in noisy environments has always been a major concern for researchers.

[0003] In speech recognition systems, speech signals are typically converted into acoustic features before subsequent model training. The choice of acoustic features directly determines whether the key information contained in the speech signal can be extracted using a feature vector of finite dimensions. Currently, the most widely used acoustic feature is the Mel Frequency Cepstral Coefficient (MFCC). The MFCC uses a Mel filter bank to simulate the sensitivity of human hearing to sound signals of different frequencies, applying nonlinear gain to the speech signal in the frequency domain, amplifying frequency bands with high auditory correlation and reducing those with low correlation. Compared to the MFCC, the Gammatone filter bank better simulates the impulse response of the physiological filter bank in the human cochlea to sound signals in the peripheral auditory nerve, and can better contain the amplitude and phase information of the speech signal. Some studies have applied Gammatone filter banks to acoustic feature extraction, such as Gammatone Frequency Cepstral Coefficients (GFCC).

[0004] Although the application of gamma-pass filter banks enables acoustic features to better simulate human hearing characteristics and realize the nonlinear characteristics of speech signal distribution in the frequency domain, which helps to improve the accuracy of speech recognition systems to some extent, the frequency domain resolution of the speech signal is reduced after passing through the filter bank, and the information content becomes sparser than the original signal. This causes the loss of some useful details in the speech signal, which affects the density of the final extracted acoustic feature information and thus affects the speech recognition effect. Summary of the Invention

[0005] This invention provides a noise-robust acoustic feature extraction method based on gamma-ton scaling basis vectors, which solves the problems of poor robustness of acoustic features in noisy environments and low speech recognition accuracy in the prior art.

[0006] The noise-robust acoustic feature extraction method based on gamma-pass scaling basis vectors according to embodiments of the present invention includes:

[0007] Pre-emphasis processing of speech signals;

[0008] The pre-emphasis processed speech signal is then framed.

[0009] Perform Fourier transform on the speech signal after frame segmentation;

[0010] The scaling factor is calculated based on the spectral distribution characteristics of the gamma-pass filter bank, and the basis vector is optimized based on the scaling factor.

[0011] Discrete cosine transform is performed on the speech signal after Fourier transform and the optimized basis vectors to extract noise-robust acoustic features from the speech signal.

[0012] According to some embodiments of the present invention, the step of performing framing processing on the pre-emphasis processed speech signal includes:

[0013] Set the frame segmentation parameters as follows: frame interval is 2 milliseconds, and unit frame length is 50 milliseconds;

[0014] Based on the set framing parameters, the pre-emphasis processed speech signal is framing.

[0015] According to some embodiments of the present invention, the calculation of the scaling factor based on the spectral distribution characteristics of the gamma-pass filter bank includes:

[0016] Based on the nonlinear characteristics of the gamma-pass filter bank, the scaling factor A(k) is calculated:

[0017]

[0018] Where c represents the scaling factor, m represents the m-th order filter, M represents the total order of the filter, ! represents the factorial, b represents the time attenuation factor, i represents the i-th frame, k represents the k-th FFT point, k0(m) represents the center frequency of the m-th order filter, and φ represents the filter phase.

[0019] According to some embodiments of the present invention, performing discrete cosine transform on the speech signal after Fourier transform and the optimized basis vectors to extract noise-robust acoustic features from the speech signal includes:

[0020] Calculate the noise-robust acoustic feature F(i,k) extracted from the k-th FFT point of the i-th frame of the speech signal after Fourier transform:

[0021]

[0022] Where E(i,k) represents the speech signal energy at the k-th FFT point of the i-th frame of the speech signal after Fourier transform, and N represents the total number of FFT points. This represents the semicosine basis vector.

[0023] According to some embodiments of the present invention, the center frequency k0(m) of the m-th order filter is calculated using the following formula:

[0024]

[0025] Where freqhigh represents the upper limit of the Fourier transform frequency domain, and freqlow represents the lower limit of the Fourier transform frequency domain.

[0026] According to some embodiments of the present invention, the value of c is 4, the value of b satisfies: b = 24.7(4.37k0(m) / 1000+1), and the value of φ is 1.

[0027] According to some embodiments of the present invention, the pre-emphasis processing of the speech signal includes:

[0028] The speech signal is pre-emphasized using an FIR high-pass digital filter.

[0029] The speech recognition method according to an embodiment of the present invention includes:

[0030] Noise-robust acoustic features are extracted from speech signals using the noise-robust acoustic feature extraction method based on gamma-ton scaling basis vectors as described above.

[0031] The speech signal is identified based on the noise-robust acoustic characteristics.

[0032] A speech recognition system according to an embodiment of the present invention includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the speech recognition method as described above.

[0033] According to an embodiment of the present invention, a computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the speech recognition method as described above.

[0034] By employing the embodiments of the present invention, scaling factors are obtained based on the frequency domain distribution characteristics of the gamma-pass filter bank and directly applied to generate the basis vectors of acoustic features. This preserves the original details of the speech signal to the greatest extent, ensures the amount of information carried by the acoustic features under noise interference, and effectively improves the speech recognition effect.

[0035] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0036] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of the embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings:

[0037] Figure 1 This is a flowchart of a noise robust acoustic feature extraction method based on gamma-pass scaling basis vectors in an embodiment of the present invention. Detailed Implementation

[0038] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0039] like Figure 1 As shown, the noise-robust acoustic feature extraction method based on gamma-pass scaling basis vectors in this embodiment of the invention includes:

[0040] Pre-emphasis processing of the speech signal can make the energy distribution of the speech signal more even in the frequency domain, thereby increasing the amplitude of the high-frequency band of the speech signal and suppressing the amplitude of the low-frequency band of the speech signal.

[0041] The pre-emphasis processed speech signal is then framed; this framed processing is based on the fact that the speech signal can be approximated as a steady-state time-invariant signal within a short time range.

[0042] Perform a Fourier transform on the speech signal that has completed frame-segmentation; thereby realizing the conversion of the speech signal from the time domain to the frequency domain.

[0043] The scaling factor is calculated based on the spectral distribution characteristics of the gamma-pass filter bank, and the basis vector is optimized based on the scaling factor. By realizing the nonlinear characteristics of the frequency domain distribution of the speech signal in the process of generating the half-cosine basis vector, the problem of low information frequency domain resolution caused by the filter bank is effectively improved.

[0044] Discrete cosine transform is performed on the speech signal after Fourier transform and the optimized basis vectors to extract noise-robust acoustic features from the speech signal.

[0045] By employing the embodiments of the present invention, scaling factors are obtained based on the frequency domain distribution characteristics of the gamma-pass filter bank and directly applied to generate the basis vectors of acoustic features. This preserves the original details of the speech signal to the greatest extent, ensures the amount of information carried by the acoustic features under noise interference, and effectively improves the speech recognition effect.

[0046] Based on the above embodiments, further variant embodiments are proposed. It should be noted that, in order to keep the description brief, only the differences from the above embodiments are described in each variant embodiment.

[0047] According to some embodiments of the present invention, the step of performing framing processing on the pre-emphasis processed speech signal includes:

[0048] Set the frame segmentation parameters as follows: frame interval is 2 milliseconds, and unit frame length is 50 milliseconds;

[0049] Based on the set framing parameters, the pre-emphasis processed speech signal is framing.

[0050] Compared to the rapid changes in the time domain of speech signals, most noise signals change relatively slowly in the time domain. When processing noisy speech in frames, using a longer frame length can better model the noise, making it easier to design acoustic features in noisy environments. However, the high computational complexity also limits the application of these methods in practical engineering.

[0051] Therefore, this embodiment of the invention optimizes the frame length selection during the speech signal framing process. Based on experimental experience, a frame length of 50 milliseconds is selected for processing noisy speech signals. This framing duration is both higher than the optimal framing length for speech signals in an ideal environment (25 milliseconds) and lower than the framing length used in previous acoustic feature extraction methods in noisy environments (≥100 milliseconds), thus balancing computational complexity and the impact of frame length on noise modeling.

[0052] Furthermore, selecting an appropriate framing interval for the speech signal is crucial, considering the characteristics of most noise signals. In this embodiment of the invention, based on experimental experience, the framing interval length is set to 2 milliseconds. Using a smaller framing interval is more helpful in capturing the differences in the rate of change between the speech signal and the noise signal between frames.

[0053] According to some embodiments of the present invention, the calculation of the scaling factor based on the spectral distribution characteristics of the gamma-pass filter bank includes:

[0054] Based on the nonlinear characteristics of the gamma-pass filter bank, the scaling factor A(k) is calculated:

[0055]

[0056] Where c represents the scaling factor, m represents the m-th order filter, M represents the total order of the filter, ! represents the factorial, b represents the time attenuation factor, i represents the i-th frame, k represents the k-th FFT point, k0(m) represents the center frequency of the m-th order filter, and φ represents the filter phase.

[0057] According to some embodiments of the present invention, performing discrete cosine transform on the speech signal after Fourier transform and the optimized basis vectors to extract noise-robust acoustic features from the speech signal includes:

[0058] Calculate the noise-robust acoustic feature F(i,k) extracted from the k-th FFT point of the i-th frame of the speech signal after Fourier transform:

[0059]

[0060] Where E(i,k) represents the speech signal energy at the k-th FFT point of the i-th frame of the speech signal after Fourier transform, and N represents the total number of FFT points. This represents the semicosine basis vector.

[0061] According to some embodiments of the present invention, the center frequency k0(m) of the m-th order filter is calculated using the following formula:

[0062]

[0063] Where freqhigh represents the upper limit of the Fourier transform frequency domain, and freqlow represents the lower limit of the Fourier transform frequency domain.

[0064] According to some embodiments of the present invention, the value of c is 4, the value of b satisfies: b = 24.7(4.37k0(m) / 1000+1), and the value of φ is 1.

[0065] According to some embodiments of the present invention, the pre-emphasis processing of the speech signal includes:

[0066] The speech signal is pre-emphasized using an FIR high-pass digital filter.

[0067] The following reference Figure 1 A noise-robust acoustic feature extraction method based on gamma-pass scaling basis vectors according to an embodiment of the present invention is described in detail below. It is to be understood that the following description is merely illustrative and not intended to limit the scope of the invention. Any similar structures or variations thereof employing the present invention should be included within the scope of protection of the present invention.

[0068] The flowchart of the acoustic feature extraction method proposed in this invention is as follows: Figure 1 As shown, it includes the following steps:

[0069] First, considering the uneven distribution of frequency domain energy in speech signals, with high-frequency signal energy lower than low-frequency signal energy, pre-emphasis processing is performed on the speech signal to make its frequency domain energy distribution more even, thereby increasing the amplitude of high-frequency signals and suppressing the amplitude of low-frequency signals.

[0070] Secondly, since speech signals can be approximated as steady-state time-invariant signals within a short time range, they need to be processed by framing. Therefore, considering the characteristics of noise signals, the frame length and interval are optimized during the framing process.

[0071] Compared to the rapid changes in speech signals in the time domain, most noise signals change relatively slowly. When processing noisy speech signals by frame, using a longer frame length allows for better noise modeling and facilitates the design of acoustic features in noisy environments. For example, in the Power function-based Distribution Normalization (PNDD) method, the frame length is 100 milliseconds; in the Long-Term Log-Spectral Subtraction (LTLSS) method, the frame length is over 1 second. Although using a longer frame length plays a positive role in extracting speech signal information from acoustic features in noisy environments, the high computational complexity also limits the application of these methods in practical engineering.

[0072] Therefore, this invention optimizes the frame length selection in the speech signal framing process. Based on experimental experience, a frame length of 50 milliseconds is selected for processing noisy speech signals. This framing time is both higher than the optimal framing length for speech signals in an ideal environment (25 milliseconds) and lower than the framing length used in previous acoustic feature extraction methods in noisy environments (≥100 milliseconds), thus balancing computational complexity and the impact of frame length on noise modeling.

[0073] Furthermore, selecting an appropriate framing interval for the speech signal is crucial, considering the characteristics of most noise signals. In this invention, based on experimental experience, the framing interval length is set to 2 milliseconds. Using a smaller framing interval is more helpful in capturing the differences in the rate of change between the speech signal and the noise signal between frames.

[0074] Then, Fourier transform is performed on the frame-segmented speech signal to realize its conversion from the time domain to the frequency domain; at the same time, scaling factors are calculated based on the spectral distribution characteristics of the gamma-pass filter bank to optimize the basis vector.

[0075] This invention eliminates the use of filter banks and realizes the nonlinear characteristics of the frequency domain distribution of the speech signal during the generation of the half-cosine basis vector, effectively improving the problem of low information frequency domain resolution caused by filter banks. Specifically, based on the nonlinear characteristics of the gamma-ton filter bank, a set of scaling coefficients A(k) with the same nonlinear characteristics as the gamma-ton filter bank is generated through mathematical transformation:

[0076]

[0077] Where c is the proportionality coefficient; m is the filter order, of which there are M, and M is taken as 32 in this material experiment; b is the time attenuation coefficient, the larger b is, the shorter the filtering time; k0(m) is the center frequency of the m-order filter; φ is the filter phase.

[0078] The above parameters can be adjusted according to requirements. In the corresponding experiment of this invention, the following values ​​can be set: c = 4, b = 24.7 (4.37k0(m) / 1000+1), φ = 1.

[0079] The formula for calculating the center frequency k0(m) is:

[0080]

[0081] Finally, the acoustic feature calculation proposed in this invention is as follows:

[0082]

[0083] The acoustic features proposed in this invention maintain the original frequency domain resolution N (FFT points) in the frequency domain, while the frequency domain resolution of traditional GFCC is only the filter order M of the filter bank under the action of the filter bank.

[0084] Finally, the acoustic features are calculated by performing Discrete Cosine Transform (DCT) on the frequency domain data after Fourier transform and the optimized basis vectors.

[0085] The acoustic feature extraction method disclosed in this invention can be applied to speech recognition systems. By optimizing the selection of speech framing parameters and combining the optimization of basis vectors using gamma-pass filter banks, the robustness of acoustic features in noisy environments can be improved, thereby increasing the recognition accuracy of the speech recognition system.

[0086] This invention optimizes the selection of frame length and frame interval, taking into account both the computational complexity of the algorithm and the impact of frame parameters on noise modeling. This not only helps to capture the difference in the rate of change of speech and noise signals between frames, but also makes it easier to apply in practical engineering.

[0087] This invention uses scaling coefficients based on gamma-pass filter banks to process basis vectors, eliminating the problem of reduced information density of acoustic features in the frequency domain caused by gamma-pass filter banks, ensuring the frequency domain resolution of acoustic features, and effectively improving speech recognition performance.

[0088] The following is a specific example of a robust acoustic feature extraction method applied to a speech recognition system.

[0089] Using the TIMIT speech dataset as the experimental dataset, with 3696 speech data points used as training data and 1344 speech data points used as validation data, and MFCC and GFCC as contrast features, the experimental results under different noise environments are shown in Table 1:

[0090] Table 1. Speech recognition accuracy (%) in noisy environments

[0091] SNR MFCC GFCC Method of the present invention No noise 93.1 93.3 95.0 20dB 92.2 92.4 94.6 15dB 86.3 87.7 93.7 10dB 78.7 79.1 90.6 5dB 30.6 54.4 83.3 0dB 15.5 36.2 65.5

[0092] Experimental results show that, compared with classic MFCC and GFCC features, the acoustic features proposed in this invention achieve higher speech recognition accuracy in different environments, especially in noisy environments. The experimental results verify the robustness of the acoustic features proposed in this invention in noisy environments.

[0093] It should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0094] The speech recognition method according to an embodiment of the present invention includes:

[0095] Noise-robust acoustic features are extracted from speech signals using the noise-robust acoustic feature extraction method based on gamma-ton scaling basis vectors as described above.

[0096] The speech signal is identified based on the noise-robust acoustic characteristics.

[0097] By employing the embodiments of the present invention, scaling factors are obtained based on the frequency domain distribution characteristics of the gamma-pass filter bank and directly applied to generate the basis vectors of acoustic features. This preserves the original details of the speech signal to the greatest extent, ensures the amount of information carried by the acoustic features under noise interference, and effectively improves the speech recognition effect.

[0098] A speech recognition system according to an embodiment of the present invention includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the speech recognition method as described above.

[0099] By employing the embodiments of the present invention, scaling factors are obtained based on the frequency domain distribution characteristics of the gamma-pass filter bank and directly applied to generate the basis vectors of acoustic features. This preserves the original details of the speech signal to the greatest extent, ensures the amount of information carried by the acoustic features under noise interference, and effectively improves the speech recognition effect.

[0100] According to an embodiment of the present invention, a computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the speech recognition method as described above.

[0101] By employing the embodiments of the present invention, scaling factors are obtained based on the frequency domain distribution characteristics of the gamma-pass filter bank and directly applied to generate the basis vectors of acoustic features. This preserves the original details of the speech signal to the greatest extent, ensures the amount of information carried by the acoustic features under noise interference, and effectively improves the speech recognition effect.

[0102] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0103] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0104] Furthermore, numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0105] It should also be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. Any reference signs enclosed in parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.

Claims

1. A noise-robust acoustic feature extraction method based on gamma-pass scaling basis vectors, characterized in that, include: Pre-emphasis processing of speech signals; The pre-emphasis processed speech signal is then framed. Perform Fourier transform on the speech signal after frame segmentation; calculate the scaling factor based on the spectral distribution characteristics of the gamma-pass filter bank, and optimize the basis vector based on the scaling factor; Discrete cosine transform is performed on the speech signal after Fourier transform and the optimized basis vectors to extract noise-robust acoustic features from the speech signal. The calculation of the scaling factor based on the spectral distribution characteristics of the gamma-pass filter bank includes: Calculate the scaling factor based on the nonlinear characteristics of the gamma-ton filter bank. : ， in, Represents the proportionality coefficient. Indicates the first First-order filter, The '!' indicates the total order of the filter, and the '!' indicates the factorial. Indicates the time decay coefficient. Indicates the first frame, Indicates the first FFT points, Indicates the first The center frequency of the first-order filter, Indicates the filter phase; The step of performing a discrete cosine transform on the speech signal after Fourier transform and the optimized basis vectors to extract noise-robust acoustic features from the speech signal includes: The speech signal whose Fourier transform has been completed is calculated. Frame number Noise robust acoustic features extracted from FFT points : ， in, This represents the first speech signal that has undergone Fourier transform. Frame number Speech signal energy at each FFT point Represents the total number of points in the FFT. This represents the semicosine basis vector.

2. The method as described in claim 1, characterized in that, The step of performing frame segmentation on the pre-emphasis processed speech signal includes: Set the frame segmentation parameters as follows: frame interval is 2 milliseconds, and unit frame length is 50 milliseconds; Based on the set framing parameters, the pre-emphasis processed speech signal is framing.

3. The method as described in claim 1, characterized in that, The first Center frequency of the first-order filter The calculation formula is: ， Where freqhigh represents the upper limit of the frequency domain of the speech signal during Fourier transform, and freqlow represents the lower limit of the frequency domain of the speech signal during Fourier transform.

4. The method as described in claim 1, characterized in that, The The value of is 4, the satisfy: =24.7(4.37 / 1000+1), the The value is 1.

5. The method as described in claim 1, characterized in that, The pre-emphasis processing of the speech signal includes: The speech signal is pre-emphasized using an FIR high-pass digital filter.

6. A speech recognition method, characterized in that, include: The noise robust acoustic feature extraction method based on gamma-pass scaling basis vectors as described in any one of claims 1-5 is used to extract noise robust acoustic features from speech signals; The speech signal is identified based on the noise-robust acoustic characteristics.

7. A speech recognition system, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the speech recognition method as described in claim 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the speech recognition method as described in claim 6.

Images