Intelligent terminal speaker accurate identification method based on speech recognition and intelligent terminal

Abstract

The application relates to the technical field of speech recognition, in particular to a speaker accurate identification method of an intelligent terminal based on speech recognition and an intelligent terminal, which comprises the following steps: decomposing a current speech signal into a plurality of speech component signals, determining the speaker voice covering degree and the speaker voice feature difficult-to-distinguish degree of each speech component signal, combining the voice feature change condition with time of each speech component signal, determining the environmental noise interference factor of each speech component signal, obtaining the reconstruction weight of each speech component signal, weighting and reconstructing all the speech component signals, obtaining an updated speech signal, inputting the updated speech signal into a pre-trained speaker identification model, and outputting the identified speaker. The application decomposes the current speech signal, weights and reconstructs the speech signal, removes the environmental noise interference while protecting the voice features in the speech signal, and thus guarantees the accuracy of speaker identification.

Description

Technical Field

[0001] This invention relates to the field of speech recognition technology, and more specifically to a method for accurate speaker recognition in smart terminals based on speech recognition, and a smart terminal thereof. Background Technology

[0002] Accurate speaker recognition on intelligent terminals based on speech recognition is a multidisciplinary technical field involving speech signal processing, pattern recognition, machine learning, and other areas. Its core lies in extracting the speaker's unique features from the speech signal and then using advanced models for recognition and verification. These unique speaker features extracted from the speech signal can reflect the speaker's physiological and behavioral characteristics, such as vocal cavity structure and pronunciation habits.

[0003] Existing problems: In practical applications, the user's voice signal collected by the smart terminal may be interfered with by external environmental noise, such as traffic noise and conversations among people. This noise may mask the user's voice. In addition, in environments with echoes or reverberation, such as large conference rooms or empty rooms, the voice signal may be reflected multiple times, making it difficult to extract effective speaker voice features during the speech recognition process, which may affect the accuracy of speaker recognition. Summary of the Invention

[0004] This invention provides a method for accurate speaker recognition in a smart terminal based on speech recognition, as well as a smart terminal, to solve existing problems.

[0005] The present invention relates to a method for accurate speaker recognition in a smart terminal based on speech recognition, and the smart terminal adopts the following technical solution:

[0006] One embodiment of the present invention provides a method for accurate speaker identification in a smart terminal based on speech recognition, the method comprising the following steps:

[0007] Acquire the current speech signal and decompose it into several speech component signals;

[0008] Based on the amplitude changes at different times in each speech component signal, the degree to which the speaker's voice is masked in each speech component signal is determined.

[0009] Based on the energy distribution in each speech component signal, determine the speaker speech feature indistinguishability of each speech component signal;

[0010] Based on the degree of speaker voice obscuration and the difficulty in distinguishing speaker voice features in each speech component signal, and combined with the changes in speech features over time in each speech component signal, the environmental noise interference factor of each speech component signal is determined.

[0011] Based on the magnitude of the environmental noise interference factor of each speech component signal, the reconstruction weight of each speech component signal is obtained. All speech component signals are weighted and reconstructed to obtain the updated speech signal, which is then input into the pre-trained speaker recognition model to output the recognized speaker.

[0012] Furthermore, the specific steps for determining the degree to which the speaker's voice is masked in each speech component signal are as follows:

[0013] In each speech component signal, obtain the first... The amplitude at time and the first The normalized value of the absolute value of the difference in amplitude at time t is used as the first... Waveform smoothness at any given moment;

[0014] The moment when the waveform non-smoothness is greater than the preset non-smoothness threshold is recorded as the interference moment, and the time period formed by adjacent interference moments is recorded as the interference time period.

[0015] The severity of interference coverage for each interference period is determined based on the duration of each interference period and the waveform non-smoothness at all times within each interference period.

[0016] Obtain the sum of the interference severity for all interference periods, and record it as the second sum.

[0017] The ratio of the sum of the time intervals between all adjacent interference periods to the duration of the speech component signal is recorded as the second ratio. The normalized value of the product of the inverse proportional value of the second ratio and the second sum is recorded as the degree to which the speaker's voice is covered for each speech component signal.

[0018] Furthermore, the specific steps for determining the severity of interference coverage during each interference period are as follows:

[0019] In each speech component signal, the ratio of the duration of each interference period to the duration of the speech component signal is obtained and recorded as the first ratio. The sum of waveform non-smoothness at all times within each interference period is obtained and recorded as the first sum. The product of the first ratio and the first sum is recorded as the interference masking severity of each interference period.

[0020] Furthermore, the specific steps for determining the speaker speech feature indistinguishability of each speech component signal are as follows:

[0021] Obtain the energy value of each speech component signal at different frequencies;

[0022] For each speech component signal, the mean of the energy values ​​at all frequencies is obtained and recorded as the energy threshold; all frequencies with energy values ​​greater than the energy threshold are recorded as high-energy frequencies; the normalized value of the difference between the maximum high-energy frequency and the minimum high-energy frequency is recorded as the size of the high-energy frequency distribution range.

[0023] Using the absolute value of the difference between any two high-energy frequencies as the clustering distance, all high-energy frequencies are clustered to obtain several clusters;

[0024] The total frequency interval distance is obtained based on the difference in high-energy frequencies between clusters;

[0025] The energy distribution discreteness is obtained by considering the number of clusters and the total frequency interval distance.

[0026] The mean of the range of high-energy frequency distribution and the energy distribution dispersion is obtained as the speaker speech feature indistinguishability of each speech component signal.

[0027] Furthermore, the specific steps for obtaining the total frequency interval distance are as follows:

[0028] Obtain the highest high-energy frequency among all high-energy frequencies in each cluster, and denote it as the label frequency of each cluster. Sort all clusters in ascending order of label frequency to obtain the cluster sequence.

[0029] In the cluster sequence, obtain the first... The minimum high-energy frequency among all high-energy frequencies in the cluster minus the i-th The difference between the highest high-energy frequencies among all high-energy frequencies in the cluster is used as the ... The and the first The frequency interval distance between each cluster is obtained by summing the frequency interval distances between all adjacent clusters and normalizing the sum, which is denoted as the total frequency interval distance.

[0030] Furthermore, the specific steps for obtaining the energy distribution discreteness are as follows:

[0031] The normalized value of the product of the number of clusters and the total frequency interval distance is denoted as the energy distribution discreteness.

[0032] Furthermore, the specific steps for determining the environmental noise interference factor of each speech component signal are as follows:

[0033] Based on the degree to which the speaker's voice is obscured and the indistinguishability of the speaker's voice features in each speech component signal, the degree of suspected noise interference in each speech component signal is obtained.

[0034] With a preset threshold M, each speech component signal is divided into M speech component signal segments. The Mel frequency cepstral coefficients of each speech component signal segment are obtained. The mean of the cosine similarity of the Mel frequency cepstral coefficients of any two speech component signal segments is obtained and denoted as the first similarity.

[0035] Each speech component signal is filtered to obtain a filtered speech component signal. The filtered speech component signal is then divided into M new speech component signal segments. The Mel frequency cepstral coefficients of each new speech component signal segment are obtained. The mean of the cosine similarity of the Mel frequency cepstral coefficients of any two new speech component signal segments is obtained and denoted as the second similarity.

[0036] Obtain the normalized value of the difference between the second similarity and the first similarity, denoted as the first difference value; obtain the normalized value of the product of the first difference value and the second similarity, denoted as the speech feature consistency factor of each speech component signal.

[0037] The average of the speech feature consistency factor and the suspected noise interference level of each speech component signal is obtained and used as the environmental noise interference factor of each speech component signal.

[0038] Furthermore, the specific steps for obtaining the suspected noise interference level of each speech component signal are as follows:

[0039] The average of the degree to which the speaker's voice is obscured and the indistinguishability of the speaker's speech features for each speech component signal is used as the suspected noise interference level for each speech component signal.

[0040] Furthermore, the specific steps for obtaining the reconstruction weights of each speech component signal are as follows:

[0041] Obtain the feature values ​​of each speech component signal;

[0042] The normalized value of the product of the inverse proportional value of the environmental noise interference factor and the feature value of each speech component signal is obtained and used as the reconstruction weight of each speech component signal.

[0043] The present invention also proposes a speech recognition-based intelligent terminal speaker accurate identification intelligent terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program stored in the memory to implement the steps of the aforementioned speech recognition-based intelligent terminal speaker accurate identification method.

[0044] The beneficial effects of the technical solution of the present invention are:

[0045] In this embodiment of the invention, the current speech signal is decomposed into several speech component signals. Based on the amplitude changes of each speech component signal at different times, the degree of speaker voice obscuration in each speech component signal is determined. Based on the energy distribution in each speech component signal, the speaker's speech feature indistinguishability in each speech component signal is determined. Thus, through analysis of the speech signal in the time and frequency domains, the degree of suspected noise interference is determined, ensuring the denoising effect of subsequent speech signals. Then, combining the changes of speech features in each speech component signal over time, the environmental noise interference factor of each speech component signal is determined, thereby distinguishing pseudo-noise interference caused by the user's physical condition and emotional state, further ensuring the denoising effect of subsequent speech signals, and thus ensuring the accuracy of speaker recognition. The reconstruction weight of each speech component signal is obtained, and all speech component signals are weighted and reconstructed to obtain an updated speech signal, which is input into a pre-trained speaker recognition model, outputting the recognized speaker. Thus, this invention decomposes and reconstructs the current speech signal with weights, thereby protecting the speech features in the speech signal while removing environmental noise interference, obtaining an updated speech signal, and using the updated speech signal for speaker recognition, thereby ensuring the accuracy of speaker recognition. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the steps of the method for accurate speaker recognition in a smart terminal based on speech recognition, as described in this invention.

[0048] Figure 2 This is a schematic diagram of the user's voice signal. Detailed Implementation

[0049] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the speech recognition-based intelligent terminal speaker accurate identification method and intelligent terminal proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0050] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0051] The following description, in conjunction with the accompanying drawings, details the method for accurate speaker recognition in a smart terminal based on speech recognition, and the specific solution of the smart terminal provided by this invention.

[0052] Please see Figure 1 The diagram illustrates a flowchart of a method for accurate speaker recognition in a smart terminal based on speech recognition, according to an embodiment of the present invention. The method includes the following steps:

[0053] Step S001: Obtain the current speech signal and decompose the current speech signal into several speech component signals.

[0054] In this embodiment, the speaker accurate identification process of the intelligent terminal based on speech recognition is as follows: (1) Speech signal acquisition: The intelligent terminal acquires the user's speech signal through the microphone. A schematic diagram of the user's speech signal is shown below. Figure 2 As shown, Figure 2 The horizontal axis represents time, and the vertical axis represents amplitude. (2) Preprocessing: The collected speech signal is subjected to preprocessing operations such as noise reduction and echo cancellation to improve the accuracy of speech recognition. (3) Feature extraction: Features are extracted from the processed speech signal, such as Mel frequency cepstral coefficients (MFCC), which can represent the speaker's speech characteristics. (4) Speaker model training: The extracted features are used to train the speaker recognition model, such as Gaussian mixture model (GMM) or deep learning model (such as CNN, RNN). (5) Speaker verification or identification: When the user uses the smart terminal again, the system will extract the features of the current speech and compare them with the trained model to verify the user's identity or identify the speaker. (6) Decision and feedback: Based on the speaker recognition results, the smart terminal makes corresponding decisions, such as unlocking the device, executing voice commands, etc., and provides feedback to the user.

[0055] This allows us to obtain a pre-trained speaker recognition model, as well as the current speech signal.

[0056] It should be noted that in the process of obtaining the pre-trained speaker recognition model, the features extracted from the user's historical speech signal are Mel frequency cepstral coefficients, and the speaker recognition model used is a Gaussian mixture model, which is a well-known technique. The specific method will not be introduced here.

[0057] It should be further noted that: due to the potential interference of external environmental noise in the current speech signal, it is difficult to extract effective speaker speech features during speech recognition. Therefore, it is necessary to filter the current speech signal to enhance the speaker speech features and eliminate environmental noise interference, thereby ensuring the accuracy of subsequent speech recognition. In practical applications of speech recognition-based intelligent terminal speaker identification, speech signals are easily affected by various noise interference sources. For example, when using an intelligent terminal outdoors, surrounding traffic noise, crowd noise, or natural wind noise can all interfere with the clarity of the speech signal. Therefore, in this embodiment, the current speech signal is first decomposed. This allows for the extraction of key feature components from the complex speech signal. These feature components are crucial for speech recognition and understanding and can also help identify and separate noise components, thereby selectively removing or reducing the impact of these noises in subsequent processing.

[0058] Principal component analysis is used to decompose the current speech signal into several speech component signals and obtain the feature values ​​of each speech component signal.

[0059] Principal component analysis is a well-known technique, and its specific method will not be described here. The larger the eigenvalue, the more information the corresponding speech component signal contains.

[0060] Step S002: Determine the degree to which the speaker's voice is masked in each speech component signal based on the amplitude changes at different times in each speech component signal.

[0061] It should be noted that: because users expect smart terminals to accurately understand and respond to their commands, they will normally speak as clearly and steadily as possible to avoid recognition errors due to excessively loud or soft volume. Therefore, noise-free speech signals will appear relatively clear and regular in the time domain, with waveforms typically exhibiting smooth amplitude changes, clearly reflecting the rhythm and intensity of the speech signal, such as the start and end of syllables and the continuity of speech. In contrast, speech signals with noise interference typically exhibit chaotic waveforms and large amplitude fluctuations in the time domain. The larger the amplitude of the noise interference superimposed on the speech signal, the more it can mask the original speech signal, making the speaker's voice difficult to hear. Therefore, it is necessary to determine the degree to which the speaker's voice is masked for each speech component based on its waveform performance in the time domain.

[0062] Preferably, in one embodiment of the present invention, the method for obtaining the degree of speaker voice obscuration for each speech component signal includes:

[0063] In each speech component signal, obtain the first... The amplitude at time and the first The normalized value of the absolute value of the difference in amplitude at time t is used as the first... Waveform smoothness at any given moment.

[0064] In this embodiment, the min-max normalization method is used to normalize the absolute value of the amplitude difference between all adjacent moments in each speech component signal to a range of 0 to 1. Furthermore, the waveform unsmoothness at the last moment in each speech component signal is set to the waveform unsmoothness at the second-to-last moment. The min-max normalization method is a well-known technique, and its specific implementation will not be described here.

[0065] The default non-smoothing threshold is 0.5, and we will use this as an example for explanation.

[0066] In each speech component signal, the moment when the waveform non-smoothness is greater than the preset non-smoothness threshold is recorded as the interference moment, and the time period formed by adjacent interference moments is recorded as the interference time period.

[0067] In each speech component signal, the ratio of the duration of each interference period to the duration of the speech component signal is obtained and recorded as the first ratio. The sum of waveform non-smoothness at all times within each interference period is obtained and recorded as the first sum. The product of the first ratio and the first sum is recorded as the interference masking severity of each interference period.

[0068] In each speech component signal, the sum of the interference severity for all interference periods is obtained and denoted as the second sum. Following chronological order, the ratio of the sum of the time intervals between all adjacent interference periods to the duration of the speech component signal is obtained and denoted as the second ratio. The product of the inverse proportional value of the second ratio and the second sum is then calculated. The normalized value is denoted as the degree to which the speaker's voice is masked for each speech component signal.

[0069] It should be noted that the longer the duration of environmental noise interference, and the less smooth the waveform of the speech signal during the interference period, the more severe the noise interference masking during that period. In this embodiment, the difference between 1 and the second ratio is used as the inverse proportional value of the second ratio. If there is only one interference period in the speech component signal, then the inverse proportional value of the second ratio is set to 1. As The normalized value, where, It is a linear normalization function. Therefore, the greater the severity of interference masking in all interference periods of the speech component signal, and the smaller the time interval between all adjacent interference periods, that is, the better the continuity of noise interference, the greater the degree to which the speaker's voice is masked in each speech component signal.

[0070] Step S003: Determine the speaker speech feature indistinguishability of each speech component signal based on the energy distribution in each speech component signal.

[0071] It should be noted that: the above analysis discussed the superposition interference of environmental noise on the amplitude of speech component signals. Further analysis is needed to examine the frequency characteristics of the interference noise in the speech signal. The more different frequencies of interference noise in the speech signal, the more complex the spectral structure of the noise, which significantly increases the difficulty of speech recognition. Since the energy in a normal speech signal is mainly concentrated within a narrow frequency range, while noise components may cause energy dispersion, resulting in a wider energy distribution, the speaker's speech feature indistinguishability for each speech component signal can be determined based on its energy distribution in the frequency domain.

[0072] Preferably, in one embodiment of the present invention, the method for obtaining the speaker speech feature indistinguishability of each speech component signal includes:

[0073] Short-time Fourier transform is used to obtain the energy value of each speech component signal at different frequencies.

[0074] The short-time Fourier transform is a well-known technique, and its specific method will not be described here.

[0075] For each speech component signal, the difference between the maximum frequency and the minimum frequency is obtained as the size of the total frequency distribution range.

[0076] For each speech component signal, the average energy value across all frequencies is calculated and recorded as the energy threshold. All frequencies with energy values ​​greater than the energy threshold are designated as high-energy frequencies. The difference between the maximum and minimum high-energy frequencies is then calculated. The normalized value is denoted as the size of the high-energy frequency distribution range.

[0077] Among them, the difference The ratio of the total frequency distribution range to the total frequency distribution range is used as the difference. The normalized value.

[0078] For each speech component signal, the absolute value of the difference between any two high-energy frequencies is used as the clustering distance. A density-based clustering algorithm (DBSCAN) is used to cluster all high-energy frequencies to obtain several clusters.

[0079] Density-based clustering algorithms are well-known techniques, and their specific methods will not be described here. This clustering operation is equivalent to dividing a high-energy-frequency distribution range into several smaller areas (clusters) where high-energy frequencies cluster together.

[0080] For each speech component signal, the highest high-energy frequency among all high-energy frequencies in each cluster is obtained and denoted as the labeled frequency of each cluster. All clusters are then sorted in ascending order of their labeled frequencies to obtain a cluster sequence.

[0081] For each speech component signal, obtain the first [number] in the cluster sequence. The minimum high-energy frequency among all high-energy frequencies in the cluster minus the i-th The difference between the highest high-energy frequencies among all high-energy frequencies in the cluster is used as the ... The and the first The frequency interval distance between each cluster is calculated as the sum of the frequency interval distances between all adjacent clusters. The normalized value is denoted as the total frequency interval distance.

[0082] Among them, the sum value The ratio of the sum to the total frequency distribution range is used as the sum value. The normalized value.

[0083] For each speech component signal, obtain the product of the number of clusters and the total frequency interval distance. The normalized value of is denoted as the energy distribution discreteness.

[0084] Among them, with As The normalized value. If the number of clusters corresponding to each speech component signal is 1, it means that the energy in the speech component signal is mainly concentrated in a narrow frequency range, and the noise component does not cause energy dispersion, so the energy distribution discreteness is set to 0.

[0085] For each speech component signal, the mean of the range of high-energy frequency distribution and the energy distribution dispersion is obtained as the speaker speech feature indistinguishability of each speech component signal.

[0086] It should be noted that the larger the range of high-energy frequency distribution and the greater the energy distribution dispersion, the more the noise components cause energy dispersion, resulting in a wider energy distribution. In other words, the more complex the spectral structure of the noise is, which will significantly increase the difficulty of speech recognition.

[0087] Step S004: Based on the degree of speaker voice obscuration and the indistinguishability of speaker voice features in each speech component signal, and combined with the changes in speech features over time in each speech component signal, determine the environmental noise interference factor of each speech component signal.

[0088] It should be noted that a user's physical condition and emotional state can affect the waveform of the speech signal in the time domain and the energy distribution in the frequency domain. For example, a cold or sore throat may cause a user's voice to become hoarse or hoarse, while excitement or anger may cause the user to speak with a cracked voice or a faster speech rate. This results in larger data values ​​corresponding to the degree of speaker voice masking and the difficulty in distinguishing speaker voice features for each speech component signal. However, changes caused by a user's physical condition and emotional state are part of the speech signal itself, not external noise, and therefore cannot be filtered out. This is because in practical applications, smart terminals need to infer the user's physical or emotional state by monitoring subtle changes in speech in order to respond appropriately. For example, "I'm tired" may simply express physical fatigue and a need for rest when the user is tired, while it may express deeper disappointment or helplessness and a need for more care and support when the user is frustrated. Therefore, for "I'm tired," smart terminals need to respond differently based on different emotional expressions.

[0089] It should be further noted that when the data values ​​corresponding to the degree of speaker voice obscuration and the difficulty in distinguishing speaker voice features are both large for each speech component signal, it is necessary to analyze whether it is caused by the user's physical condition and emotional state or by environmental noise interference, and then focus on filtering the environmental noise interference.

[0090] Preferably, in one embodiment of the present invention, the method for obtaining the environmental noise interference factor of each speech component signal includes:

[0091] The average of the degree to which the speaker's voice is obscured and the indistinguishability of the speaker's voice features in each speech component signal is obtained as the suspected noise interference level of each speech component signal.

[0092] The preset quantity threshold M is 5, and this will be used as an example for explanation.

[0093] Each speech component signal is divided into M speech component signal segments. The Mel frequency cepstral coefficients of each speech component signal segment are obtained. The mean of the cosine similarity of the Mel frequency cepstral coefficients of any two speech component signal segments is obtained and denoted as the first similarity.

[0094] Wiener filtering is used to filter each speech component signal to obtain the filtered speech component signal. The filtered speech component signal is then divided into M new speech component signal segments. The Mel frequency cepstral coefficients of each new speech component signal segment are obtained. The mean of the cosine similarity of the Mel frequency cepstral coefficients of any two new speech component signal segments is obtained and denoted as the second similarity.

[0095] Wiener filtering is a well-known technique, and its specific methods will not be described here.

[0096] Calculate the difference between the second similarity and the first similarity. The normalized value is denoted as the first difference value, and the product of the first difference value and the second similarity is obtained. The normalized value of is denoted as the speech feature consistency factor for each speech component signal.

[0097] It should be noted that: with and As respectively and The normalized value. The acquisition of Mel-frequency cepstral coefficients (MFCCs) is a well-known technique, and the specific method will not be described here. MFCC features are one of the most commonly used features in speech recognition systems, effectively extracting information such as timbre and formants from speech signals. Furthermore, the Mel-frequency cepstral coefficients (MFCCs) are a set of data values ​​that together constitute a feature vector. Since the timbre, formants, and other information in a speech signal are determined by the physical characteristics of the speaker's vocal tract, external environmental noise mainly affects timbre and formants by interfering with the clarity and intelligibility of the speech signal. This does not change the physical characteristics of the speaker's vocal tract, which can be reduced through filtering. However, the user's physical condition and emotional state directly affect the speaker's physiological state and vocal mechanism, thereby altering timbre and formants. For example, a cold can cause inflammation of the vocal cords, throat, and nasal cavity, causing changes in the vibration of the vocal cords and the shape of the resonance cavity, which directly affects the position and intensity of timbre and formants. When emotionally agitated, the speaker's breathing rate increases and vocal cord tension increases, which also leads to changes in pitch, volume, and timbre, thus affecting formants. Therefore, under the influence of a cold and emotional excitement, the speech features in the speech signal will change continuously, resulting in dissimilar Mel-frequency cepstral coefficients of different speech segments, i.e., low cosine similarity.

[0098] To further clarify: if the suspected noise interference in the speech component signal is significant due to the user's physical condition and emotional state, then both the first and second similarities before and after filtering should be low, and their magnitudes should be similar. If the suspected noise interference in the speech component signal is significant due to environmental noise interference, then the first similarity before filtering should be low, while the second similarity after filtering should be high. Therefore, the larger the difference between the second and first similarities, and the higher the second similarity (i.e., the higher the speech feature consistency factor of each speech component signal), the more likely the significant suspected noise interference is caused by environmental noise interference.

[0099] The average of the speech feature consistency factor and the suspected noise interference level of each speech component signal is obtained and used as the environmental noise interference factor of each speech component signal.

[0100] Step S005: Based on the magnitude of the environmental noise interference factor of each speech component signal, obtain the reconstruction weight of each speech component signal, perform weighted reconstruction on all speech component signals, obtain the updated speech signal, input it into the pre-trained speaker recognition model, and output the recognized speaker.

[0101] Preferably, in one embodiment of the present invention, the speaker identification method includes:

[0102] The normalized value of the product of the inverse proportional value of the environmental noise interference factor of each speech component signal and the feature value of each speech component signal is obtained and used as the reconstruction weight of each speech component signal.

[0103] It should be noted that in this embodiment, the difference between 1 and the environmental noise interference factor of each speech component signal is used as the inverse proportional value of the environmental noise interference factor of each speech component signal. The product of the inverse proportional value of the environmental noise interference factor of each speech component signal and the eigenvalue is used as the first product. The sum of the first products of all speech component signals is recorded as the third sum. The ratio of the first product of each speech component signal to the third sum is used as the normalized value of the product of the inverse proportional value of the environmental noise interference factor of each speech component signal and the eigenvalue of each speech component signal. The larger the eigenvalue, the more information the speech component signal contains, and the smaller the environmental noise interference. Therefore, a larger weight is assigned during reconstruction. Conversely, when the eigenvalue is small and the environmental noise interference is large, a smaller reconstruction weight is assigned to ensure that noise interference is filtered out while preserving the speech features in the speech signal.

[0104] Based on the reconstruction weight of each speech component signal, all speech component signals are reconstructed using weighted average to obtain the updated speech signal.

[0105] The weighted reconstruction method after decomposing the speech signal using principal component analysis is a well-known technique, and the specific method will not be introduced here.

[0106] The updated speech signal is input into a pre-trained speaker recognition model, and the recognized speaker is output.

[0107] Specifically, features such as Mel-frequency cepstral coefficients (MFCCs) are extracted from the updated speech signal and then compared with a pre-trained speaker recognition model to verify the user's identity or identify the speaker.

[0108] The present invention also provides a speech recognition-based intelligent terminal speaker accurate identification intelligent terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program stored in the memory to implement the steps of the aforementioned speech recognition-based intelligent terminal speaker accurate identification method.

[0109] This invention is now complete.

[0110] In summary, in this embodiment of the invention, the current speech signal is decomposed into several speech component signals. Based on the amplitude changes of each speech component signal at different times, the degree of speaker voice obscuration in each speech component signal is determined. Based on the energy distribution in each speech component signal, the speaker's speech feature indistinguishability in each speech component signal is determined. Combining the changes of speech features in each speech component signal over time, the environmental noise interference factor of each speech component signal is determined, thereby obtaining the reconstruction weight of each speech component signal. All speech component signals are then weighted and reconstructed to obtain an updated speech signal, which is input into a pre-trained speaker recognition model, outputting the identified speaker. This invention, by decomposing and weighting the current speech signal, removes environmental noise interference while preserving the speech features in the speech signal, thus ensuring the accuracy of speaker recognition.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for accurate speaker identification in intelligent terminals based on speech recognition, characterized in that, The method includes the following steps: Acquire the current speech signal and decompose it into several speech component signals; Based on the amplitude changes at different times in each speech component signal, the degree to which the speaker's voice is masked in each speech component signal is determined. Based on the energy distribution in each speech component signal, determine the speaker speech feature indistinguishability of each speech component signal; Based on the degree of speaker voice obscuration and the difficulty in distinguishing speaker voice features in each speech component signal, and combined with the changes in speech features over time in each speech component signal, the environmental noise interference factor of each speech component signal is determined. Based on the magnitude of the environmental noise interference factor of each speech component signal, the reconstruction weight of each speech component signal is obtained. All speech component signals are weighted and reconstructed to obtain the updated speech signal, which is then input into the pre-trained speaker recognition model and outputs the recognized speaker. The specific steps involved in determining the speaker speech feature indistinguishability of each speech component signal are as follows: Obtain the energy value of each speech component signal at different frequencies; For each speech component signal, the mean of the energy values ​​at all frequencies is obtained and recorded as the energy threshold; all frequencies with energy values ​​greater than the energy threshold are recorded as high-energy frequencies; the normalized value of the difference between the maximum high-energy frequency and the minimum high-energy frequency is recorded as the size of the high-energy frequency distribution range. Using the absolute value of the difference between any two high-energy frequencies as the clustering distance, all high-energy frequencies are clustered to obtain several clusters; Obtain the highest high-energy frequency among all high-energy frequencies in each cluster, and denote it as the label frequency of each cluster. Sort all clusters in ascending order of label frequency to obtain the cluster sequence. In the cluster sequence, obtain the first... The minimum high-energy frequency among all high-energy frequencies in the cluster minus the i-th The difference between the highest high-energy frequencies among all high-energy frequencies in the cluster is used as the ... The and the first The frequency interval distance between each cluster is obtained by summing the frequency interval distances between all adjacent clusters and normalizing the sum of the frequency interval distances between each cluster. This normalized value is denoted as the total frequency interval distance. The energy distribution discreteness is obtained by considering the number of clusters and the total frequency interval distance. The mean of the range of high-energy frequency distribution and the energy distribution dispersion is obtained as the speaker speech feature indistinguishability of each speech component signal.

2. The method for accurate speaker recognition in intelligent terminals based on speech recognition according to claim 1, characterized in that, The specific steps for determining the degree to which the speaker's voice is masked in each speech component signal are as follows: In each speech component signal, obtain the first... The amplitude at time and the first The normalized value of the absolute value of the difference in amplitude at time t is used as the first... Waveform smoothness at any given moment; The moment when the waveform non-smoothness is greater than the preset non-smoothness threshold is recorded as the interference moment, and the time period formed by adjacent interference moments is recorded as the interference time period. The severity of interference coverage for each interference period is determined based on the duration of each interference period and the waveform non-smoothness at all times within each interference period. Obtain the sum of the interference severity for all interference periods, and record it as the second sum. The ratio of the sum of the time intervals between all adjacent interference periods to the duration of the speech component signal is recorded as the second ratio. The normalized value of the product of the inverse proportional value of the second ratio and the second sum is recorded as the degree to which the speaker's voice is covered for each speech component signal.

3. The method for accurate speaker recognition in intelligent terminals based on speech recognition according to claim 2, characterized in that, The specific steps for determining the severity of interference coverage during each interference period are as follows: In each speech component signal, the ratio of the duration of each interference period to the duration of the speech component signal is obtained and recorded as the first ratio. The sum of waveform non-smoothness at all times within each interference period is obtained and recorded as the first sum. The product of the first ratio and the first sum is recorded as the interference masking severity of each interference period.

4. The method for accurate speaker recognition in intelligent terminals based on speech recognition according to claim 1, characterized in that, The specific steps involved in obtaining the energy distribution discreteness are as follows: The normalized value of the product of the number of clusters and the total frequency interval distance is denoted as the energy distribution discreteness.

5. The method for accurate speaker recognition in intelligent terminals based on speech recognition according to claim 1, characterized in that, The specific steps for determining the environmental noise interference factor of each speech component signal are as follows: Based on the degree to which the speaker's voice is obscured and the indistinguishability of the speaker's voice features in each speech component signal, the degree of suspected noise interference in each speech component signal is obtained. With a preset threshold M, each speech component signal is divided into M speech component signal segments. The Mel frequency cepstral coefficients of each speech component signal segment are obtained. The mean of the cosine similarity of the Mel frequency cepstral coefficients of any two speech component signal segments is obtained and denoted as the first similarity. Each speech component signal is filtered to obtain a filtered speech component signal. The filtered speech component signal is then divided into M new speech component signal segments. The Mel frequency cepstral coefficients of each new speech component signal segment are obtained. The mean of the cosine similarity of the Mel frequency cepstral coefficients of any two new speech component signal segments is obtained and denoted as the second similarity. Obtain the normalized value of the difference between the second similarity and the first similarity, denoted as the first difference value; obtain the normalized value of the product of the first difference value and the second similarity, denoted as the speech feature consistency factor of each speech component signal. The average of the speech feature consistency factor and the suspected noise interference level of each speech component signal is obtained and used as the environmental noise interference factor of each speech component signal.

6. The method for accurate speaker recognition in intelligent terminals based on speech recognition according to claim 5, characterized in that, The specific steps for obtaining the suspected noise interference level of each speech component signal are as follows: The average of the degree to which the speaker's voice is obscured and the indistinguishability of the speaker's speech features for each speech component signal is used as the suspected noise interference level for each speech component signal.

7. The method for accurate speaker recognition in intelligent terminals based on speech recognition according to claim 1, characterized in that, The specific steps for obtaining the reconstruction weights of each speech component signal are as follows: Obtain the feature values ​​of each speech component signal; The normalized value of the product of the inverse proportional value of the environmental noise interference factor and the feature value of each speech component signal is obtained and used as the reconstruction weight of each speech component signal.

8. A speech recognition-based intelligent terminal for accurate speaker identification, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is executed by the processor, it implements the steps of the speech recognition-based intelligent terminal speaker accurate recognition method as described in any one of claims 1-7.

Images