Voice endpoint detection method for low signal-to-noise ratio noise environment
By introducing the LM cepstral distance product and adaptive dynamic threshold method into the traditional speech endpoint detection technology, the problem of insufficient accuracy of speech endpoint detection in low signal-to-noise ratio environments is solved, and more efficient and accurate speech endpoint detection is achieved.
Patent Information
- Application Number
- CN202511129243.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-13
- Publication Date
- 2025-09-19
AI Technical Summary
Traditional speech endpoint detection technology has difficulty in accurately identifying speech signals in low signal-to-noise ratio noise environments, resulting in a significant reduction in detection performance.
An adaptive dynamic threshold endpoint detection algorithm based on LM cepstral distance product is adopted. LPCC and MFCC parameters are used to extract speech features. Combined with the improved adaptive dynamic threshold method, the traditional cepstral distance endpoint detection algorithm is improved.
In low signal-to-noise ratio noise environments, the accuracy of speech endpoint detection is significantly improved, the probability of noise frames being misidentified as speech frames is reduced, and the efficiency of the speech recognition process is improved.
Smart Images
Figure CN120673789A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of speech recognition technology, and in particular is a method for detecting speech endpoints in a low signal-to-noise ratio noise environment. Background Art With the rapid development of technology and the internet, communication between humans and machines is becoming increasingly commonplace, making it crucial to achieve better communication between humans and machines. As a high-tech technology, speech recognition plays a crucial role in this field. Speech recognition research involves many scientific fields, including acoustics, digital signal processing, and pattern recognition, enabling human-machine interaction. Currently, speech recognition technology is often susceptible to interference from ambient noise, which can affect recognition accuracy. This has a significant impact on people's quality of life and living standards in noisy environments.
[0002] Voice Activity Detection (VAD) technology typically distinguishes speech from non-speech signals amidst complex noise interference and determines the start and end points of speech signals. VAD is widely used in communication systems, speech coding, and other fields, and is an essential component of speech recognition. Feature extraction and speech recognition are only effective when accurate endpoint detection technology is available.
[0003] Traditional speech endpoint detection technologies include short-term speech features, cepstrum, and spectral entropy. These algorithms perform well in high signal-to-noise ratio environments, but their performance deteriorates significantly in poor speech environments with low signal-to-noise ratios, making them difficult to meet practical needs. Summary of the Invention
[0004] The purpose of the present invention is to overcome the deficiency of traditional speech endpoint detection technology that it is difficult to accurately identify speech signals in a low signal-to-noise ratio noise environment. A speech endpoint detection method for a low signal-to-noise ratio noise environment is provided. By extracting speech features using linear prediction cepstral coefficients and Mel-frequency cepstral coefficients on the basis of traditional algorithms, and then combining the improved adaptive dynamic threshold method to improve the traditional cepstral distance endpoint detection algorithm, effective speech signals can be accurately identified in a low signal-to-noise ratio environment.
[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows: The first aspect of the present invention provides a method for speech endpoint detection in a low signal-to-noise ratio (SNR) noisy environment. The method uses an adaptive dynamic threshold endpoint detection algorithm based on the LM cepstral distance product. The algorithm uses new parameter features calculated from LPCC and MFCC parameters to better distinguish endpoints of noisy speech. The method specifically includes the following steps: Step 1: Collect noisy speech signals and perform preprocessing to obtain the speech signals to be tested; Step 2: Take the first n frames of the speech signal to be measured within the initial silent period, 5≤n≤30, and calculate the average value of the cepstral coefficients of these n frames of signal as the cepstral coefficient estimate of the background noise; Step 3: Calculate the cepstral distance between the cepstral coefficient of each frame of the speech signal to be tested and the estimated value of the background noise cepstral coefficient calculated in step 2. The cepstral distance includes the LPCC cepstral distance. and MFCC cepstral distance ; Step 4: LPCC cepstrum distance calculated according to step 3 and MFCC cepstral distance , calculate the LM cepstral distance product of each frame signal in the speech signal to be tested :
[0006] Step 5: LM cepstral distance product calculated according to step 4 , take the NIS frame segment The maximum value is used as the adaptive threshold parameter , then determine the double thresholds T1 and T2; Step 6: Based on the T1 and T2 values, a double-threshold decision method is used to detect the speech signal to be tested and identify the speech segment and the noise segment.
[0007] The above-mentioned method designed by the present invention greatly reduces the probability of noise frames being misjudged as speech frames by utilizing the advantages of LPCC cepstrum's sensitivity to vowels and MFCC cepstrum's selective perception characteristics and noise resistance to signal spectrum in low signal-to-noise ratio noise environments, thereby improving the accuracy of the speech endpoint detection method.
[0008] As a possible implementation of the first aspect of the present invention, in step 2, the calculated average cepstral coefficients of the first n frames of signal include the average LPCC cepstral coefficients and the average MFCC cepstral coefficients, that is, the LPCC cepstral coefficient estimate of the background noise is calculated and recorded as , the MFCC cepstral coefficient estimate of the background noise is recorded as .
[0009] As a possible implementation of the first aspect of the present invention, in step 3, the calculated cepstral distance includes the LPCC cepstral distance and the MFCC cepstral distance, wherein:
[0010]
[0011] Where, Indicates the first i LPCC cepstral coefficients of frame signal and the estimated value of background noise cepstral coefficient The cepstral distance between Indicates the first i MFCC cepstral coefficients of frame signal and the estimated value of background noise cepstral coefficient The cepstral distance between .
[0012] As a possible implementation of the first aspect of the present invention, in step 5, the calculated LM cepstral distance product is first Smoothing is performed, and then NIS frame segments are selected The maximum value is used as the adaptive threshold parameter .
[0013] As a possible implementation of the first aspect of the present invention, in step 5, the formula for determining the dual thresholds T1 and T2 is as follows:
[0014] Where, i Take 1 or 2, see k 1 and k 2 are the two multiplication coefficients of the decision threshold, k 1< k 2, λ is the threshold increment coefficient. Specifically, λ =0.3, k 1=1.7, k 2=2.2, which can effectively balance the noise robustness and speech detection sensitivity.
[0015] As a possible implementation of the first aspect of the present invention, a method for preprocessing a noisy speech signal includes: performing pre-emphasis and frame windowing processing on the speech signal.
[0016] A second aspect of the present invention provides a processor, which is used to run a program, wherein the program executes the above-mentioned voice endpoint detection method when running.
[0017] A third aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the above-mentioned voice endpoint method.
[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention is a speech endpoint detection method for low signal-to-noise ratio noise environments. An adaptive dynamic threshold endpoint detection algorithm based on LM cepstral distance product is designed. In low signal-to-noise ratio noise environments, the advantages of LPCC cepstral sensitivity to vowels and MFCC cepstral selective perception characteristics and noise resistance to signal spectrum are utilized to greatly reduce the probability of noise frames being misjudged as speech frames.
[0019] (2) The present invention is a speech endpoint detection method for low signal-to-noise ratio noise environment. In the speech recognition process, it requires one LPCC cepstral distance calculation, one MFCC cepstral distance calculation and one LM cepstral distance product calculation, and then confirms the threshold parameter d th , decision threshold coefficient k i and the threshold increment coefficient λ. Compared with existing work, this method improves the efficiency of the speech recognition process and reduces computational overhead while ensuring the accuracy of speech endpoint detection under low signal-to-noise ratio conditions.
[0020] (3) The present invention is a method for detecting speech endpoints in a low signal-to-noise ratio noise environment. In order to improve the quality of speech endpoint detection, a new adaptive threshold is designed. Since the decision threshold of the speech endpoint is related to the strength of the noise, an adaptive threshold is particularly important for the speech endpoint. The original fixed decision threshold is multiplied by the coefficient k i A threshold increment coefficient λ is added on the basis of λ, which can change with SNR and reasonably distinguish the decision threshold of cepstral distance to improve the quality of speech endpoint detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 Schematic diagram of a flow chart of a method for voice endpoint detection in a low signal-to-noise ratio noise environment according to the present invention; Figure 2 This is a waveform diagram of a clean speech signal used in an embodiment of the present invention; Figure 3 This is a waveform diagram of a speech signal with white noise superimposed on pure speech in Example 1 of the present invention (SNR 5dB); Figure 4 This is a waveform diagram of the speech signal of Example 2 of the present invention with the hfchannel voice superimposed on the pure speech (SNR 5dB); Figure 5 This is a waveform diagram of a speech signal with leopard noise superimposed on pure speech in Example 3 of the present invention (SNR 5dB); Figure 6 The results of detecting the noisy speech in Example 1 using the method of the present invention (the signal-to-noise ratio is 5dB); Figure 7The results of detecting the noisy speech in Example 2 using the method of the present invention (the signal-to-noise ratio is 5dB); Figure 8 The results of the present invention's method for detecting the noisy speech in Example 3 (the signal-to-noise ratio is 5dB); Figure 9 The results of using LPCC cepstral distance method to detect noisy speech (white noise, signal-to-noise ratio is 0dB); Figure 10 This is the result of using the MFCC cepstral distance method to detect noisy speech (white noise, signal-to-noise ratio is 0dB); Figure 11 This is the result of detecting noisy speech using the method of the present invention (white noise, signal-to-noise ratio is 0dB). DETAILED DESCRIPTION
[0022] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the embodiments described are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0023] Speech endpoint detection algorithms can be categorized into two types: feature-based and model-based. The latter are complex and difficult to implement, so in practical applications, we prefer feature-based methods, which are relatively simple, require minimal computation, and are easy to implement. Feature-based methods primarily involve the following steps: preprocessing, feature extraction, threshold estimation, and endpoint detection. Feature extraction converts speech signal features into characteristic parameters, which are then used to distinguish speech from non-speech segments. This is the key to endpoint detection and, ultimately, the entire speech recognition process.
[0024] However, in speech endpoint detection, the signal-to-noise ratio (SNR) directly determines the separability of speech signals from noise. Compared to high-SNR environments, low SNR (typically defined as SNR < 10dB, meaning noise energy approaches or exceeds speech energy) significantly amplifies the "masking" and "interference" of speech features. This blurs the distinction between speech and non-speech, significantly increasing detection difficulty and significantly reducing accuracy.
[0025] In order to solve the above problems, the endpoint detection of speech can be realized under low signal-to-noise ratio conditions. Figure 1 The present invention provides a method for detecting speech endpoints in a low signal-to-noise ratio (SNR) noise environment, which specifically includes the following steps: Step 1: Collect noisy speech signals and perform preprocessing to obtain the speech signals to be tested; The preprocessing of the collected noisy speech signal includes pre-emphasis and frame windowing processing, where: Pre-emphasis processing uses a high-pass filter to enhance the high-frequency components of the voice signal, balance the spectral characteristics of the signal, and make subsequent processing (such as feature extraction and judgment) more accurate.
[0026] Frame windowing can convert continuous speech signals into "short-term stationary signals," providing a basis for extracting "short-term features." Windowing can suppress spectral leakage and improve the accuracy of subsequent spectrum analysis. In this embodiment, the frame length is 20-30ms, the frame shift is 50%, and a Hamming window or a Haining window is used to reduce spectral leakage.
[0027] Step 2: Take the first n frames of the speech signal to be measured within the initial silent period, 5≤n≤30, and calculate the average value of the cepstral coefficients of these n frames of signal as the cepstral coefficient estimate of the background noise; It is worth noting that the cepstral coefficients described in this step include LPCC cepstral coefficients (i.e. linear prediction cepstral coefficients) and MFCC cepstral coefficients (i.e. Mel-frequency cepstral coefficients). The cepstral coefficients are calculated as follows: Assume a speech signal is , obtained after preprocessing The frame speech signal is . The LPCC cepstral coefficient is recorded as , The MFCC cepstral coefficient is recorded as .
[0028] The above method can be used to calculate the LPCC cepstral coefficients and MFCC cepstral coefficients of each frame signal in the first n frames of the signal during the initial silence period of the speech signal to be tested (the speech during the initial silence period is selected, and the noise characteristics are stable and representative). Then, the average value of these n LPCC cepstral coefficients is calculated to obtain the LPCC cepstral coefficient estimate of the background noise, which is recorded as , calculate the average value of these n MFCC cepstral coefficients and get the estimated value of the MFCC cepstral coefficient of the background noise, recorded as .
[0029] Step 3: Calculate the cepstral distance between the cepstral coefficient of each frame of the speech signal to be tested and the estimated value of the cepstral coefficient of the background noise, and the cepstral distance includes the LPCC cepstral distance. and MFCC cepstral distance ; The calculation method of the cepstral distance in this step is as follows: Suppose there are two sets of signals and According to the calculation method of the cepstral coefficients in step 2, their LPCC cepstral coefficients are calculated as follows: and , then the cepstral distance between them is for:
[0030] Under the same setting, their MFCC cepstral coefficients are and , then the cepstral distance between them is for:
[0031] Correspondingly, the above method and formula can be used to calculate the LPCC cepstral distance and MFCC cepstral distance between each frame of the speech signal and the background noise, as follows:
[0032]
[0033] Where, Represents the LPCC cepstral coefficient of the i-th frame signal in the speech signal to be tested and the estimated value of background noise cepstral coefficient The cepstral distance between Represents the MFCC cepstral coefficient of the i-th frame signal in the speech signal to be tested and the estimated value of background noise cepstral coefficient The cepstral distance between .
[0034] Step 4: LPCC cepstrum distance calculated according to step 3 and MFCC cepstral distance , calculate the LM cepstral distance product of each frame signal in the speech signal to be tested :
[0035] It should be noted that the applicant has found through research that when performing speech endpoint detection, LPCC is suitable for scenarios with high signal-to-noise ratio and mainly voiced sounds, but its ability to distinguish between clear sounds and noise is insufficient, and its ability to resist noise is insufficient, especially in low signal-to-noise ratio environments, and it is prone to distortion. The scope of use of MFCC is wider than that of LPCC, and its ability to resist noise is stronger, but its false detection rate is high under complex noise, the calculation is time-consuming, and it may also fail under specific noise. Through research, the applicant found that by drawing on the product form of the energy-entropy product (the product of energy and information entropy), the complementary characteristics of energy and entropy are integrated through the product form, which significantly improves the detection robustness and accuracy in low signal-to-noise ratio environments, and the calculated LPCC cepstral distance and MFCC cepstral distance are coupled, and the calculation formula of the present invention is used to calculate the LM cepstral distance product. , which can combine the advantages of LPCC and MFCC detection and overcome their respective shortcomings.
[0036] Step 5: LM cepstral distance product calculated according to step 4 , first Perform smoothing and then take the NIS frame segment The maximum value is used as the adaptive threshold parameter ,Finally, combined with the adaptive threshold parameter The method for determining the dual thresholds T1 and T2 is as follows:
[0037] Where, i Take 1 or 2, see k 1 and k 2 are the two multiplication coefficients of the decision threshold, k 1< k 2, λ is the threshold increment coefficient.
[0038] The applicant has verified based on a large amount of experimental data that when the threshold increment coefficient If the threshold is lower than 0.3, it cannot effectively distinguish between sudden noise or weak speech segments. If the threshold is higher than 0.3, it will be too sensitive to the signal-to-noise ratio fluctuation. In order to reduce the detection performance of the system under steady noise, Taking 0.3 can better balance noise robustness and speech detection sensitivity.
[0039] In addition, the two multiplication coefficients of the decision threshold and The setting is based on its core role in the double threshold mechanism. A value of 1.7 can reliably distinguish between stationary background noise and frames containing speech or burst noise, balancing noise robustness and weak speech detection sensitivity. When the value is 2.2, it ensures that only speech frames that are significantly higher than the noise (especially voiced sounds) can trigger the high threshold detection, avoiding sudden noise from being misjudged as speech, and is consistent with the threshold increment coefficient ( =0.3), ensuring that the system can maintain good detection performance and anti-interference ability under both stationary noise and low signal-to-noise ratio conditions.
[0040] Step 6: Based on the T1 and T2 values, a double-threshold decision method is used to detect the speech signal to be tested and identify the speech segment and the noise segment.
[0041] The method of the present invention is further described below with reference to specific examples.
[0042] Example 1 The speech used in this embodiment is a pure speech signal of "blue sky, white clouds, and green sea" recorded by professional recording equipment in a quiet environment. The signal sample is a monophonic, 8kHz sampling frequency, and 16-bit quantized .wav file, as shown in FIG. Figure 2 shown.
[0043] The noise selected in this embodiment is Gaussian white noise from the NOISEX-92 database. The signal-to-noise ratio is set to 5dB. The noise is superimposed on the pure speech. The result is as follows: Figure 3 shown.
[0044] The speech endpoint detection method of the present invention is used for detection, and the detection results are as follows: Figure 6 shown.
[0045] Example 2 The speech used in this embodiment is a pure speech signal of "blue sky, white clouds, and green sea" recorded by professional recording equipment in a quiet environment. The signal sample is a monophonic, 8kHz sampling frequency, and 16-bit quantized .wav file, as shown in FIG. Figure 2 shown.
[0046] The noise selected in this embodiment is the hfchannel noise from the TV. The signal-to-noise ratio is set to 5dB. The noise is superimposed on the pure speech. The result is as follows: Figure 4 shown.
[0047] The speech endpoint detection method of the present invention is used for detection, and the detection results are as follows: Figure 7 shown.
[0048] Example 3 The speech used in this embodiment is a pure speech signal of "blue sky, white clouds, and green sea" recorded by professional recording equipment in a quiet environment. The signal sample is a monophonic, 8kHz sampling frequency, and 16-bit quantized .wav file, as shown in FIG. Figure 2 shown.
[0049] The noise selected in this embodiment is the noise from the Leopard tank operation. The signal-to-noise ratio is set to 5dB. The noise is superimposed on the pure speech. The result is as follows: Figure 5 shown.
[0050] The speech endpoint detection method of the present invention is used for detection, and the detection results are as follows: Figure 8 shown.
[0051] Combine Figure 2-8 It can be seen that the detection method of the present invention performs endpoint detection on speech signals under different stationary noise interferences and accurately detects the start and end of speech.
[0052] Example 4 The speech used in this embodiment is a pure speech signal of "blue sky, white clouds, and green sea" recorded by professional recording equipment in a quiet environment. The signal sample is a monophonic, 8kHz sampling frequency, and 16-bit quantized .wav file, as shown in FIG. Figure 2 shown.
[0053] The noise selected in this embodiment is Gaussian white noise from the NOISEX-92 database, the signal-to-noise ratio is set to 0dB, the noise is superimposed on the pure speech, and the speech endpoint detection method of the present invention is used for detection. The detection results are as follows: Figure 11 shown.
[0054] Comparative Example 1 This comparative example uses the speech signal in Example 4 for detection, and adopts the LPCC cepstral distance method for detection. The detection results are as follows: Figure 9 shown.
[0055] Comparative Example 2 This comparative example uses the speech signal in Example 4 for detection, and adopts MFCC cepstral distance method for detection. The detection results are as follows Figure 10 shown.
[0056] Combine Figure 9-11 It can be seen that compared with the LPCC cepstral distance method and the MFCC cepstral distance method, the speech endpoint detection method of the present invention can accurately detect the start and end of speech, and has higher detection accuracy.
[0057] To prevent accidental errors, the applicant conducted extensive data verification and provides a comparative table of endpoint detection accuracy using various algorithms under varying noise and signal-to-noise ratio environments (Table 1). The accuracy is calculated by subtracting the percentage of error frames from the total number of frames (the data is manually tested, allowing for a 2-3% error).
[0058] Table 1 Comparison of voice endpoint detection accuracy in different environments (%)
[0059] Combine Figure 9-11 As shown in Table 1, the endpoint detection method of the present invention has better detection effect and noise immunity than LPCC cepstral distance endpoint detection and MFCC cepstral distance endpoint detection. Even in a low signal-to-noise ratio (0dB) environment, the accuracy rate is higher than 85%, which is 15 percentage points and 10 percentage points higher than the LPCC cepstral distance method and MFCC cepstral distance method, respectively. This verifies the superiority and detection accuracy of the new algorithm.
[0060] The above is a schematic description of the present invention and its embodiments, which is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs a structure and embodiment similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.
Claims
1. A method for speech endpoint detection in a low signal-to-noise ratio (SNR) noise environment, characterized by: The following steps are involved: Step 1: Collect noisy speech signals and perform preprocessing to obtain the speech signals to be tested; Step 2: Take the first n frames of the speech signal to be measured within the initial silent period, 5≤n≤30, and calculate the average value of the cepstral coefficients of these n frames of signal as the cepstral coefficient estimate of the background noise; Step 3: Calculate the cepstral distance between the cepstral coefficient of each frame of the speech signal to be tested and the estimated cepstral coefficient of the background noise calculated in step 2. The cepstral distance includes the LPCC cepstral distance. and MFCC cepstral distance ; Step 4: LPCC cepstrum distance calculated according to step 3 and MFCC cepstral distance , use the following formula to calculate the LM cepstral distance product of each frame signal in the speech signal to be tested : Step 5: LM cepstral distance product calculated according to step 4 , take the NIS frame segment The maximum value is used as the adaptive threshold parameter , then determine the double thresholds T1 and T2; Step 6: Based on the T1 and T2 values, a double-threshold decision method is used to detect the speech signal to be tested and identify the speech segment and the noise segment.
2. The method for speech endpoint detection in a low signal-to-noise ratio noise environment according to claim 1, characterized in that: In step 2, the calculated average cepstral coefficients of the first n frames of signal include the average LPCC cepstral coefficients and the average MFCC cepstral coefficients, that is, the LPCC cepstral coefficient estimate of the background noise is calculated and recorded as , the MFCC cepstral coefficient estimate of the background noise is recorded as .
3. The method for speech endpoint detection in a low signal-to-noise ratio noise environment according to claim 2, wherein: In step 3, the calculated cepstral distance includes the LPCC cepstral distance and the MFCC cepstral distance, where: Where, Indicates the first i LPCC cepstral coefficients of frame signal and the estimated value of background noise cepstral coefficient The cepstral distance between Indicates the first i MFCC cepstral coefficients of frame signal and the estimated value of background noise cepstral coefficient The cepstral distance between .
4. A method for speech endpoint detection in a low signal-to-noise ratio noise environment according to any one of claims 1 to 3, characterized in that: In step 5, first calculate the LM cepstral distance product Smoothing is performed, and then NIS frame segments are selected The maximum value is used as the adaptive threshold parameter .
5. The method for speech endpoint detection in a low signal-to-noise ratio noise environment according to claim 4, characterized in that: In step 5, the formula for determining the dual thresholds T1 and T2 is as follows: Where, i Take 1 or 2, see k 1 and k 2 are the two multiplication coefficients of the decision threshold, k 1< k 2, λ is the threshold increment coefficient.
6. The method for speech endpoint detection in a low signal-to-noise ratio noise environment according to claim 5, characterized in that: λ =0.3， k 1=1.7， k 2=2.2。 7. The method for speech endpoint detection in a low signal-to-noise ratio noise environment according to claim 4, characterized in that: In step 1, the preprocessing method of the noisy speech signal includes: performing pre-emphasis and frame windowing processing on the speech signal.
8. A processor, characterized in that: The processor is configured to run a program, wherein the program, when running, executes the voice endpoint method according to any one of claims 1 to 7.
9. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the voice endpoint method according to any one of claims 1 to 7.