A method, apparatus, device, and storage medium for speech processing

Abstract

This application discloses a method, apparatus, device, and storage medium for speech processing. The method includes: acquiring a first speech signal to be processed; filtering speech signals other than the target speaker's speech signal from the first speech signal based on a first voiceprint feature of a target speaker to obtain a second speech signal; and updating the first voiceprint feature of the target speaker based on the second speech signal to obtain a second voiceprint feature of the target speaker. This enables adaptive updating of the target speaker's voiceprint feature, avoiding inaccurate extraction of the target speaker's speech signal due to changes in the target speaker's voice features over time and in the environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing, and particularly relates to a speech processing method and device, equipment and a storage medium. BACKGROUND

[0002] In daily life, people will talk to others in some noisy environment, and the pure speaker's speech signal will inevitably be contaminated by noise. How to enhance the speech of the speaker in a complex environment is a current research hotspot.

[0003] When the mixed speech signal includes the voices of multiple speakers and environmental noise, the current mainstream method is to use the pre-generated voiceprint feature of the target speaker as prior information, so as to find the part with the highest similarity to the voiceprint feature of the target speaker in the mixed speech signal, identify the speech of the target speaker, and then extract a more pure speech signal of the target speaker, so as to achieve the effect of enhancing the speech of the target speaker.

[0004] However, the voice features of the speaker may change over time and environment, such as voice change of children, voice change caused by poor health due to cold, etc. At this time, if the pre-generated voiceprint feature of the speaker is still used as prior information, the accuracy of subsequent voice recognition of the speaker may be reduced, and the extracted speech signal of the target speaker may be inaccurate. SUMMARY

[0005] The embodiments of the present application at least provide a speech processing method, device, equipment and storage medium.

[0006] The technical scheme of the embodiments of the present application is as follows:

[0007] In a first aspect, the embodiments of the present application provide a speech processing method, which comprises: obtaining a first speech signal to be processed; filtering speech signals other than the speech signal of a target speaker in the first speech signal based on a first voiceprint feature of the target speaker to obtain a second speech signal; updating the first voiceprint feature of the target speaker based on the second speech signal to obtain a second voiceprint feature of the target speaker.

[0008] In a second aspect, the embodiments of the present application provide a speech processing device, which comprises: a first obtaining module configured to obtain a first speech signal to be processed; a filtering module configured to filter speech signals other than the speech signal of a target speaker in the first speech signal based on a first voiceprint feature of the target speaker to obtain a second speech signal; and an updating module configured to update the first voiceprint feature of the target speaker based on the second speech signal to obtain a second voiceprint feature of the target speaker.

[0009] In a third aspect, an embodiment of the present application provides a speech processing device, comprising a memory and a processor; wherein the memory is configured to store at least one program or at least one instruction; and the processor is connected to the memory and is configured to implement the method according to the first aspect by executing the program or the instruction.

[0010] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program, when executed by at least one processor, implements the method according to the first aspect.

[0011] The method, device, equipment and computer readable storage medium for speech processing provided in the embodiments of the present application can obtain a first speech signal to be processed, filter speech signals other than speech signals of a target speaker in the first speech signal based on a first voiceprint feature of the target speaker to obtain a second speech signal, and then update the first voiceprint feature of the target speaker based on the second speech signal to obtain a second voiceprint feature of the target speaker. In this way, adaptive updating of the voiceprint feature of the target speaker can be implemented to avoid the problem of inaccurate extraction of speech signals of the target speaker due to changes in the voice feature of the target speaker over time and environment. BRIEF DESCRIPTION OF DRAWINGS

[0012] The accompanying drawings, which are incorporated herein and constitute part of the specification, illustrate embodiments consistent with the present application and, together with the specification, serve to explain the technical solutions of the present application.

[0013] Figure 1 A flowchart of a method for speech processing provided in an embodiment of the present application is shown in FIG. 1.

[0014] Figure 2 A possible implementation flowchart of the speech processing method provided in an embodiment of the present application is shown in FIG. 2.

[0015] Figure 3 A principle block diagram of the speech processing method of an embodiment of the present application is shown in FIG. 3.

[0016] Figure 4 A flowchart of processing of a mixed speech signal using a speaker speech signal extraction model provided in an embodiment of the present application is shown in FIG. 4.

[0017] Figure 5 A composition structure diagram of a device for speech processing provided in an embodiment of the present application is shown in FIG. 5.

[0018] Figure 6 A hardware entity diagram of a speech processing device in an embodiment of the present application is shown in FIG. 6. DETAILED DESCRIPTION

[0019] In order to gain a more detailed understanding of the features and technical content of the embodiments of this application, the implementation of the embodiments of this application will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for reference and illustration only and are not intended to limit the embodiments of this application.

[0020] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application.

[0021] In the following description, references to "some embodiments" refer to a subset of all possible embodiments. It is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. It should also be noted that the terms "first, second, third" used in the embodiments of this application are merely for distinguishing similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0022] It should be understood that the term "and / or" in the embodiments of this application is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. In addition, the character " / " in this document generally indicates that the related objects before and after it have an "or" relationship.

[0023] In daily life, people often converse in noisy environments, inevitably contaminating pure speech signals with noise. Enhancing speech in complex environments is a current research hotspot. The principles of speech enhancement include speech denoising, such as suppressing noise components in noisy speech signals to improve speech quality and intelligibility. Traditional speech enhancement methods include spectral subtraction, Wiener filtering, minimum mean square error estimation, and wavelet transform-based enhancement methods. One existing method, to adaptively enhance a speaker's speech, calculates the power of the current speech signal frame based on the received signal and compares it with the noise power of the previous frame. Based on this comparison and the noise power of the previous frame, a noise estimate for the current frame is obtained, leading to the enhanced speech signal. While this method is effective at denoising noisy speech signals, it struggles to accurately distinguish the target speaker's speech signal from other interference signals when multiple speakers are present. This inaccurate noise estimation results in less than ideal enhancement of the target speaker's speech signal.

[0024] When a mixed speech signal includes the voices of multiple speakers and environmental noise, the current mainstream approach is to use pre-generated voiceprint features of the target speaker as prior information. This involves identifying the target speaker's speech by finding the portion of the mixed speech signal with the highest similarity to the target speaker's voiceprint features, thereby extracting a cleaner speech signal and enhancing the target speaker's voice. For example, deep learning methods can be used for speech recognition. In this method, high-level speech features to be recognized are extracted from the acquired mixed speech signal and the pre-acquired clean speech signal of the target speaker, thus obtaining the identity feature vector to be recognized. The identity feature vector with the highest similarity to the identity vector to be recognized is then searched from a pre-established speaker database. The identity information corresponding to the found identity feature vector is used as the speaker recognition result. Finally, the recognized target speaker's speech signal is individually enhanced to output a cleaner target speech signal.

[0025] However, a speaker's voice characteristics may change over time and with varying environments, such as a child's voice changing or a person's voice changing due to a cold. In such cases, continuing to use the speaker's previously generated voiceprint features as prior information may reduce the accuracy of subsequent voice recognition, leading to inaccurate extraction of the target speaker's speech signal.

[0026] In view of this, embodiments of this application provide a method, apparatus, device, and storage medium for speech processing. In this method, a speech signal to be processed (e.g., denoted as a first speech signal) can be acquired, and speech signals other than the target speaker's speech signal can be filtered from the speech signal based on the target speaker's voiceprint features (e.g., denoted as first voiceprint features) to extract the target speaker's speech signal, thereby enhancing the target speaker's speech signal. Furthermore, the target speaker's voiceprint features can be updated based on the extracted target speaker's speech signal (e.g., denoted as a second speech signal). Thus, adaptive updating of the target speaker's voiceprint features can be achieved, avoiding the problem of inaccurate extraction of the target speaker's speech signal due to changes in the target speaker's voice features over time and in the environment.

[0027] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0028] Figure 1 This application illustrates a voice processing method according to an embodiment of the present application. This method can be executed, for example, by a voice processing device, or by a chip, chip system, or circuit configured in a voice processing device; this application embodiment does not limit this. The voice processing device can refer to a device with data processing capabilities, such as a server, laptop, tablet, desktop computer, or mobile phone. Figure 1 As shown, the method may include:

[0029] S101, acquire the first speech signal to be processed.

[0030] The first speech signal to be processed may include the speech signal of the target speaker. For example, the first speech signal may be a mixed speech signal including the speech signal of the target speaker and interference signals, wherein the interference signals may include the speech signals of other speakers and / or ambient noise.

[0031] In some embodiments, the first speech signal may be preprocessed as follows: 11) the first speech signal is framed and windowed; 12) the first speech signal after framed and windowed processing is subjected to feature extraction processing.

[0032] For step 12), as one implementation, a short-time Fourier transform (STFT) can be performed on each frame of the first speech signal after framing and windowing to convert it into a frequency domain signal, so that the spectral features can be used as the features of the extracted first speech signal; as another implementation, the spectrum of each frame of the speech signal can be further converted into a power spectrum, and a logarithmic transform can be performed on the power spectrum to obtain the corresponding logarithmic power spectrum, so that the logarithmic power spectrum or the result of the logarithmic power spectrum after discrete cosine transform processing can be used as the features of the first speech signal.

[0033] It should be understood that the operations performed on the first speech signal in subsequent steps can also be performed on the first speech signal after feature extraction processing, or in other words, the operations can be performed on the features of the extracted first speech signal.

[0034] S102, based on the first voiceprint features of the target speaker, filter out the speech signals in the first speech signal other than the speech signal of the target speaker to obtain the second speech signal.

[0035] The first voiceprint feature can be used to characterize the voice features of the target speaker. This first voiceprint feature can be pre-stored in a voiceprint database, and can be read or retrieved from the database when needed.

[0036] In one possible approach, the first voiceprint feature could be obtained, for example, by feature extraction from a clean speech signal of the target speaker. As an example, the signal-to-noise ratio of the clean speech signal could be greater than or equal to a second threshold; in other words, the ratio of the target speaker's speech signal to the noise signal in the clean speech signal could be greater than or equal to the second threshold.

[0037] For example, the step of extracting features from the clean speech signal of the target speaker to obtain the first voiceprint feature may include:

[0038] 21) Obtain the clean speech signal of the target speaker. As an example, the speech signal generated by the user in a quiet environment can be acquired as the clean speech signal of the target speaker; as another example, the mixed speech signal including the speech signal of the target speaker can be denoised or the speech of the target speaker can be enhanced to obtain the clean speech signal of the target speaker.

[0039] 22) Use a neural network to extract features from the clean speech signal of the target speaker (e.g., extract d-vector features from the clean speech signal of the target speaker), and use the output of the neural network as the first voiceprint feature of the target speaker.

[0040] It should be noted that the above-described method for feature extraction from the clean speech signal of the target speaker is merely exemplary. For example, in another possible approach, the d-vector features can be replaced with x-vector features; furthermore, spectral features, power spectrum features, etc., of the clean speech signal of the target speaker can be extracted based on time-frequency transformation, and this application embodiment does not limit this approach.

[0041] In S102, based on the first voiceprint characteristics of the target speaker, speech signals other than the target speaker's speech signal in the first speech signal can be filtered. For example, when the first speech signal includes the target speaker's speech signal and interference signals, the interference signals in the first speech signal can be filtered to extract the target speaker's speech signal, thereby enhancing the target speaker's speech signal.

[0042] As one implementation method, based on the first voiceprint feature of the target speaker, the speech signals other than the speech signal of the target speaker in the first speech signal are filtered to obtain the second speech signal. This can be achieved by: performing similarity matching between the first voiceprint feature of the target speaker and each speech signal component in the first speech signal to obtain the probability value of each speech signal component being the speech signal of the target speaker; multiplying each probability value by the corresponding speech signal component to obtain the second speech signal.

[0043] For example, embodiments of this application may use a time-frequency diagram to represent each speech signal component in the first speech signal. For instance, the horizontal axis of the time-frequency diagram may represent the time domain, and the vertical axis may represent the frequency domain. Each time-frequency point in the time-frequency diagram may correspond to a speech signal component in the first speech signal. Further, the first voiceprint feature of the target speaker can be matched with the speech signal component corresponding to each time-frequency point to obtain a probability value that the speech signal component corresponding to each time-frequency point is the speech signal of the target speaker. This probability value can be represented as a value between 0 and 1. Subsequently, each probability value can be multiplied by the corresponding speech signal component to obtain the second speech signal.

[0044] For example, if the probability value of the speech signal component corresponding to a certain time-frequency point being the target speaker is 0.8, then the speech signal component corresponding to that time-frequency point can be multiplied by 0.8 to obtain the filtered speech signal component corresponding to that time-frequency point. By performing the above operation on the speech signal component corresponding to each time-frequency point, the speech signal after filtering the first speech signal, i.e., the second speech signal, can be obtained.

[0045] According to the method of this embodiment, when a certain speech signal component in the first speech signal is a speech signal other than the speech signal of the target speaker (such as an interference signal), the similarity (matching degree) between this speech signal component and the first voiceprint feature of the target speaker is usually low, and therefore the obtained probability value is low. By multiplying this speech signal component by a low probability value, the effect of filtering the speech signal component can be achieved. Thus, the method of this embodiment can filter the speech signals other than the speech signal of the target speaker in the first speech signal.

[0046] In some embodiments, if an STFT is performed on the first speech signal in S101, the method may further include: performing an inverse short-time Fourier transform (ISTFT) on the second speech signal, and using the result of the ISTFT transform as the enhanced speech signal of the target speaker.

[0047] S103, update the first voiceprint feature of the target speaker based on the second speech signal to obtain the second voiceprint feature of the target speaker.

[0048] In this step, the first voiceprint feature of the target speaker can be updated based on the second speech signal obtained in S102 to obtain the second voiceprint feature of the target speaker. In this way, adaptive updating of the voiceprint feature of the target speaker can be achieved, so as to avoid the problem of inaccurate extraction of the target speaker's speech signal due to changes in the target speaker's voice features over time and in the environment.

[0049] In one possible approach, the method further includes: determining the similarity between the second speech signal and the first voiceprint feature of the target speaker. In this case, updating the first voiceprint feature of the target speaker based on the second speech signal to obtain the second voiceprint feature of the target speaker may include: updating the first voiceprint feature of the target speaker based on the second speech signal if the similarity between the second speech signal and the first voiceprint feature of the target speaker is greater than or equal to a first threshold.

[0050] On the one hand, since the first voiceprint feature is obtained by feature extraction from the clean speech signal of the target speaker, when the similarity between the second speech signal and the first voiceprint feature of the target speaker is greater than or equal to the first threshold, the second speech signal can be considered to be close to the clean speech signal of the target speaker, and thus its speech quality can be considered high. Therefore, the second speech signal can be used to update the first voiceprint feature of the target speaker. On the other hand, when the similarity between the second speech signal and the first voiceprint feature of the target speaker is less than the first threshold, the speech quality of the second speech signal can be considered low. In this case, the second speech signal may not be used to update the first voiceprint feature. For example, if the first threshold is 10%, when the similarity between the second speech signal and the first voiceprint feature is less than 10%, it can be inferred that the following situations may have occurred: the first speech signal does not include the speech signal of the target speaker, or, the speech signal of another speaker is mistakenly identified as the speech signal of the target speaker. In this case, it is not necessary to use the second speech signal to update the first voiceprint feature to avoid affecting the accuracy of the target speaker's voiceprint feature.

[0051] It should be understood that the first threshold in the embodiments of this application can be a numerical value or a range of values, and is not limited thereto. The first threshold can be a pre-configured fixed value or a dynamically configured value. The device executing the speech processing method of this application can save the first threshold in advance, or it can receive the first threshold from other devices.

[0052] For example, updating the first voiceprint feature of the target speaker based on the second speech signal to obtain the second voiceprint feature of the target speaker can be achieved through the following steps: 31) weighting the second speech signal and the first voiceprint feature of the target speaker to obtain a weighted speech signal; 32) determining the weighted speech signal as the second voiceprint feature of the target speaker.

[0053] In some embodiments, after obtaining the second speech signal, a neural network can be used to extract features from the second speech signal (e.g., extract d-vector features from the second speech signal). Thus, the operations performed on the second speech signal in subsequent steps can also be performed on the second speech signal after feature extraction, or in other words, the extracted features of the second speech signal can also be performed.

[0054] It should be noted that the above-described method for feature extraction of the second speech signal is merely exemplary. For example, in another possible approach, the d-vector features can be replaced with x-vector features; furthermore, spectral features, power spectrum features, etc., can be extracted from the second speech signal based on time-frequency transformation, which is not limited in this embodiment.

[0055] In some embodiments, the method further includes: determining the weight of the second speech signal and the weight of the first voiceprint feature of the target speaker; accordingly, the weighting of the second speech signal and the first voiceprint feature of the target speaker can be performed, for example, by weighting the second speech signal and the first voiceprint feature of the target speaker based on the weight of the second speech signal and the weight of the first voiceprint feature of the target speaker.

[0056] For example, the methods for determining the weights of the second speech signal and the first voiceprint feature of the target speaker may include the following methods 1 and 2.

[0057] Method 1: Determine the weights of the second speech signal and the first voiceprint features of the target speaker based on the similarity between the second speech signal and the first voiceprint features of the target speaker; wherein, the similarity between the second speech signal and the first voiceprint features of the target speaker is negatively correlated with the weight of the second speech signal.

[0058] According to the method of this embodiment, when the similarity between the second speech signal and the first voiceprint feature of the target speaker is high, it can be considered that the voice feature of the target speaker has little change, so the second speech signal can be assigned a smaller weight. Correspondingly, when the similarity between the second speech signal and the first voiceprint feature of the target speaker is low, it can be considered that the voice feature of the target speaker has large change, so the second speech signal can be assigned a larger weight.

[0059] Method 2: Based on the time difference between the time of generating the first speech signal to be processed and the time of generating the first voiceprint feature of the target speaker, determine the weight of the second speech signal and the weight of the first voiceprint feature of the target speaker; wherein, the time difference is positively correlated with the weight of the second speech signal. In this case, the method may further include: determining the time difference between the time of generating the first speech signal to be processed and the time of generating the first voiceprint feature of the target speaker.

[0060] As one implementation, when generating the first speech signal to be processed, the time of generating the first speech signal to be processed can be recorded simultaneously; similarly, when generating the first voiceprint feature of the target speaker, the time of generating the first voiceprint feature of the target speaker can be recorded simultaneously, thereby determining the time difference between the time of generating the first speech signal to be processed and the time of generating the first voiceprint feature of the target speaker, and then determining the weight of the second speech signal and the weight of the first voiceprint feature of the target speaker based on the time difference.

[0061] Since the voice characteristics of the target speaker generally do not change significantly in a short period of time, when the time difference between generating the first speech signal to be processed and generating the first voiceprint feature of the target speaker is small, it can be assumed that the voice characteristics of the target speaker have not changed significantly, so the second speech signal can be given a smaller weight. Conversely, when the time difference between generating the first speech signal to be processed and generating the first voiceprint feature of the target speaker is large, it can be assumed that the voice characteristics of the target speaker may have changed significantly, so the second speech signal can be given a larger weight.

[0062] In some embodiments, the weight of the second speech signal is less than the weight of the target speaker's first voiceprint feature. That is, when weighting the second speech signal and the target speaker's first voiceprint feature, the target speaker's first voiceprint feature can constitute the main component, while the second speech signal can constitute the secondary component, thus achieving fine-tuning of the target speaker's first voiceprint feature. As one implementation, the weight of the target speaker's first voiceprint feature is greater than 80%, or in other words, the weight of the second speech signal is less than 20%.

[0063] The above text combined Figure 1 This application introduces a speech processing method provided by an embodiment of the present application. To facilitate understanding of the embodiments of this application, the following description is provided in conjunction with... Figure 2 This application describes one possible implementation flow of the speech processing method provided in the embodiments.

[0064] Currently, existing speech enhancement technologies mainly suffer from the following drawbacks:

[0065] 1) Traditional speech enhancement techniques are effective in processing noisy speech signals. However, when there are multiple speakers in the speaking environment, traditional speech enhancement techniques (such as spectral subtraction) are difficult to accurately distinguish the target speaker's speech from other interference signals. Therefore, they cannot accurately enhance the target speaker's speech signal, resulting in an unsatisfactory output speech signal.

[0066] 2) Existing deep learning-based speech enhancement methods can enhance the speech of a target speaker in mixed speech signals involving multiple speakers, achieving certain results. However, since the target speaker's voiceprint features (e.g., denoted as speaker ID) are generated based on the target speaker's clean speech signal, the speech signal characteristics will differ when the target speaker's voice features change, such as with age (children's voice change) or when the target speaker is unwell due to a cold. In such cases, the existing methods will become less effective at enhancing the target speaker's speech.

[0067] To address the aforementioned problems, a speech processing method is proposed. In this method, the clean speech signal of the target speaker can be used as the reference input. For mixed speech signals containing interfering speech signals, since the characteristics of the target speech signal differ from those of the interfering speech signals, the target speaker's speech signal can be extracted after training with a neural network, thereby enhancing the target speaker's speech signal. This method considers that the characteristics of the target speaker's speech signal can change over time and with environmental factors, such as a child's voice changing or a person being unwell due to a cold. In these cases, the target speaker's speech characteristics will differ from those under normal circumstances. The method can update the speaker ID (voiceprint feature) generated by the network based on the speaker's speech characteristics in real time. For example, the steps of this method may include: acquiring the target speaker's reference speech signal, which is then processed by a model to generate a speaker ID. When processing mixed speech signals containing interference, the features of the mixed speech signal can be extracted. After training with a deep neural network model, the target speaker's speech signal can be identified by combining the previously established speaker ID prior information, and the target speaker's speech signal can be separated, achieving the effect of speech enhancement. Next, the separated speech signal can be input into a speech quality evaluation system to generate a new speaker ID. This new speaker ID is then compared with the previous speaker ID, and the similarity between the two IDs is used to determine whether to update the previous speaker ID. For example, if the similarity is greater than or equal to the first threshold, the previous speaker ID is replaced by a weighted average of the two IDs; otherwise, no update is performed. This method enables adaptive updating of the speaker ID, ultimately forming a speaker-adaptive speech enhancement system.

[0068] like Figure 2 As shown, a possible implementation flow of the speech processing method provided in this application embodiment may include:

[0069] S201, Acquire the mixed speech signal (an example of the first speech signal).

[0070] In complex speaking environments, mixed speech signals may include, for example, the speech signal of the target speaker and interference signals. Interference signals may include, for example, the speech signals of other speakers and / or ambient noise.

[0071] S202, performs STFT on the mixed speech signal.

[0072] By performing STFT on the mixed speech signal, the mixed speech signal can be converted into a frequency domain signal.

[0073] S203, extract features from the mixed speech signal after passing through STFT to obtain the features of the mixed speech signal.

[0074] One approach is to convert the mixed speech signal into a frequency domain signal, and then use the spectral features as the features of the extracted mixed speech signal. Another approach is to further convert the spectrum of each frame of the speech signal into a power spectrum, and then take a logarithmic transform on the power spectrum to obtain the corresponding logarithmic power spectrum. Thus, the logarithmic power spectrum or the result of performing a discrete cosine transform on the logarithmic power spectrum can be used as the feature of the mixed speech signal.

[0075] In this embodiment, S202 and S203 can be considered as preprocessing of the mixed speech signal. On the other hand, in order to extract the speech signal of the target speaker from the mixed speech signal, it is necessary to obtain the speaker ID of the target speaker (an example of the first voiceprint feature). Obtaining the speaker ID of the target speaker can be achieved through S205 to S207.

[0076] S205, acquire clean speech signal from the target speaker.

[0077] One example is that the speech signal generated by the user in a quiet environment can be acquired as a clean speech signal of the target speaker; another example is that the mixed speech signal including the speech signal of the target speaker can be denoised or the speech signal of the target speaker can be enhanced to obtain a clean speech signal of the target speaker.

[0078] S206, extract features from the clean speech signal of the target speaker.

[0079] For example, a neural network can be used to extract features from the clean speech signal of the target speaker. For instance, a neural network can be used to extract d-vector features from the clean speech signal of the target speaker.

[0080] S207, Generate the speaker ID of the target speaker.

[0081] In one possible approach, the features of the clean target speaker's speech signal can be stored in an established feature library (speaker library) of the target speaker using a voiceprint extractor, and the features of the clean target speaker's speech signal (such as d-vector features) can be used as the target speaker's speaker ID.

[0082] After obtaining the speaker ID of the target speaker, the speaker ID of the target speaker and the features of the mixed speech signal obtained in S203 can be input into a neural network model (e.g., called the speaker speech signal extraction model), which can be used to execute S204 and S208.

[0083] S204, Perform dimensional transformation on the features of the mixed speech signal obtained in step S203.

[0084] For example, the mixed speech signal can be input into a deep neural network (DNN) or a convolutional neural network (CNN) to perform dimensionality transformation on the features of the mixed speech signal through the DNN or CNN.

[0085] S208, based on the speaker ID of the target speaker obtained in step S207, perform speech extraction on the mixed speech signal to extract the speech signal of the target speaker.

[0086] In this step, based on the speaker ID of the target speaker, the features of the mixed speech signal that has undergone dimensional transformation in S204 can be extracted to obtain the speech signal of the target speaker; or, based on the speaker ID of the target speaker, the speech signals other than the speech signal of the target speaker in the mixed speech signal can be filtered to obtain the extracted speech signal of the target speaker (an example of the second speech signal).

[0087] One implementation approach is to perform similarity matching between the target speaker's speaker ID and each speech signal component in the mixed speech signal to obtain the probability value that each speech signal component is the target speaker's speech signal (i.e., the probability that the target speaker's speech exists in each speech signal component). Further, each probability value can be multiplied by the corresponding speech signal component to obtain the extracted target speaker's speech signal.

[0088] S209, Perform ISTFT on the extracted speech signal of the target speaker.

[0089] After performing ISTFT on the extracted speech signal of the target speaker, the output can be used as the enhanced speech signal of the target speaker.

[0090] S210, determine whether the update conditions are met.

[0091] To enable adaptive updates to the target speaker's speaker ID, the enhanced speech signal of the target speaker can be input into a speech quality evaluation system. When the enhanced speech signal is input into this system, the feature vector d-vector of the enhanced speech signal is calculated and compared with the previously existing speaker ID to determine whether the update condition is met. For example, if the similarity is greater than or equal to a first threshold, the input speech signal is considered to have high speech quality and meets the update condition; in this case, steps S211 and S212 can be executed to update the target speaker's speaker ID. Conversely, if the similarity is less than the first threshold, the input speech signal is considered to have low speech quality and does not meet the update condition; in this case, the target speaker's speaker ID does not need to be updated.

[0092] S211, Generate a new speaker ID based on the extracted speech signal of the target speaker.

[0093] As one implementation method, the feature vector d-vector of the enhanced target speaker's speech signal can be used as the speaker ID of the newly generated target speaker.

[0094] S212, Weighted Update of Speaker ID

[0095] In this step, the previous speaker ID can be weighted based on the newly generated target speaker's speaker ID, thereby updating the target speaker's speaker ID.

[0096] It is understandable that the updated speaker ID of the target speaker can be used to extract speech from the mixed speech signal the next time it is acquired, thereby obtaining the speech signal of the target speaker. The mixed speech signal acquired next time can be the mixed speech signal from S201, or other mixed speech signals, without limitation.

[0097] According to the method of this embodiment, firstly, in order to obtain the speaker ID of the target speaker, a speech feature database of the target speaker can be established in advance. For example, the clean speech signal of the target speaker can be used as a reference signal input, and a feature vector d-vector is obtained through neural network training, serving as the speaker ID. Then, the mixed speech signal to be processed is subjected to STFT. After STFT transformation, the features of the mixed speech signal to be processed are extracted, and these features, along with the pre-established speaker ID of the target speaker, are input into the neural network model (speaker speech signal extraction model). After training the neural network, the probability of the target speaker's speech presence is obtained. Based on this probability, the speech signal of the target speaker can be extracted from the mixed speech signal. Finally, the extracted speech signal of the target speaker is subjected to ISTFT to obtain the enhanced speech signal of the target speaker.

[0098] To enable adaptive updates of the target speaker's speaker ID, after performing ISTFT on the extracted speech signal, a speech quality assessment system can be connected. When the extracted speech signal is input into the assessment system, the extracted feature vector (d-vector) of the target speaker can be calculated and compared with the previously existing speaker ID. If the update condition is met, the quality of the input speech signal is considered high. At this point, a new speaker ID can be generated and weighted with the original speaker ID, and stored in the network as the new speaker ID. This achieves adaptive updates of the target speaker's speaker ID, enabling more accurate recognition of the target speaker's speech and ultimately achieving speaker-adaptive speech enhancement.

[0099] Figure 3 This is a schematic diagram illustrating the principle of the speech processing method according to an embodiment of this application.

[0100] like Figure 3As shown, on the one hand, for the mixed speech signal, an STFT transform can be performed first, followed by a dimensionality transformation on the STFT-transformed mixed speech signal. On the other hand, for the clean speech signal of the target speaker used as reference input, a neural network can be used to extract features and generate a speaker ID. Then, based on the generated speaker ID, the probability of the target speaker's speech being present in the mixed speech signal can be determined, or in other words, the probability that each speech signal component in the mixed speech signal is the target speaker's speech signal. Subsequently, the probability of the target speaker's speech being present can be multiplied by the STFT-transformed mixed speech signal, and the result can be subjected to an ISTFT transform to obtain the enhanced target speaker's speech signal. Furthermore, if the ISTFT-transformed speech signal meets the update conditions, a new speaker ID can be generated. The newly generated speaker ID and the previous speaker ID can then be weighted to update the target speaker's speaker ID.

[0101] The main core aspects of the speech processing method in this application embodiment may include the following parts: 1) construction of the target speaker's voiceprint database; 2) adaptive update of the speaker ID; 3) feature extraction of mixed speech signals; 4) neural network model training; and 5) speech enhancement of the target speaker. The above five parts are described below.

[0102] 1) Construction of the target speaker's voiceprint database. First, a voiceprint database of the desired target speaker needs to be built beforehand, using the target speaker's clean speech signal as reference input. One method for building the target speaker's voiceprint database is to input the clean target speaker's speech signal into a neural network for feature extraction. During this process, the target speaker's speech signal can be labeled (i.e., generating a d-vector feature corresponding to the target speaker's speech signal). Subsequently, the d-vector output by the network can be used as the target speaker's speaker ID. The target speaker's speech features (such as the speaker ID) are placed in the built target speaker's voiceprint database for subsequent recognition and extraction of the target speaker's speech signal from mixed speech signals.

[0103] 2) Adaptive Speaker ID Update. Since the voice characteristics of the target speaker are not constant—for example, as the speaker ages and their voice changes, or when the speaker is ill—using the original speaker ID as the identification method would be ineffective. This solution continuously collects the target speaker's voice signal and updates the voiceprint features stored in the voiceprint database based on the collected voice characteristics. Specifically, the current target speaker's voice signal is collected, extracted by the network (i.e., after filtering out voice signals other than the target speaker), and then fed into the voice quality evaluation system. The current speaker's voiceprint features (d-vector) are extracted and compared with the previously existing speaker ID to determine if the update conditions are met. For example, if the similarity is greater than or equal to a first threshold, the input voice signal is considered to be of high quality and meets the update conditions, thus the target speaker's speaker ID can be updated; if the similarity is less than the first threshold, the input voice signal is considered to be of low quality and does not meet the update conditions, in which case the target speaker's speaker ID is not updated. This method achieves an adaptive speaker ID update, thereby improving the accuracy of speech recognition for the target speaker in multiple scenarios.

[0104] 3) Feature Extraction of Mixed Speech Signals. When the target speaker is in a complex environment, the acquired speech signal may be a mixture of the target speaker's speech signal, interference speech signals from other speakers, and environmental noise. For the acquired mixed speech signal, framing and windowing can be performed, and a short-time Fourier transform can be applied to each frame of the processed mixed speech signal to convert it to the frequency domain. Thus, the spectral features can be used as the features of the extracted mixed speech signal. Alternatively, the spectrum of each frame of the speech signal can be converted into a power spectrum, and a logarithmic transform can be performed to obtain the corresponding logarithmic power spectrum. This logarithmic power spectrum, or the result of a discrete cosine transform of the logarithmic power spectrum, can then be used as the feature of the mixed speech signal. The features of the mixed speech signal are then used as input to a neural network model (such as a speaker speech signal extraction model). After processing by the model, the target speaker's speech signal can be extracted.

[0105] 4) Speaker Speech Signal Extraction Model Training. To train the speaker speech signal extraction model, clean target speaker speech signal samples and mixed speech signal samples to be processed are required. The mixed speech signal may include, for example, the target speaker's speech signal and interference signals. Processing these samples allows for the extraction of features from both the clean target speaker's speech signal and the mixed speech signal. Specifically, training reference speech features and training identity features (i.e., the target speaker's speaker ID) can be extracted from the clean target speaker's speech signal. Therefore, the features of the mixed speech signal and the target speaker's speaker ID can be used as input, and the training reference speech features can be used as the target (or, in other words, the actual or standard value) for supervised learning to train the speaker speech signal extraction model.

[0106] 5) Target Speaker Speech Enhancement. After sufficient training of the aforementioned speaker speech signal extraction model, the trained model can be used to process the input target speaker's reference signal (such as the target speaker's speaker ID) and the mixed speech signal to be processed. During this process, the model can filter out speech signals other than the target speaker's speech signal from the mixed speech signal, thereby extracting (extracting) the target speaker's speech signal. After transforming (such as ISTFT transform) the extracted target speaker's speech signal, the enhanced target speaker's speech signal can be obtained.

[0107] Figure 4 This illustration shows a process for processing mixed speech signals using a speaker speech signal extraction model, as provided in an embodiment of this application. The speaker speech signal extraction model may include, for example, CNN / DNN, Long Short-Term Memory (LSTM) neural networks, and fully connected layers (FC). Figure 4 As shown, firstly, sample data of the mixed speech (mixed speech signal) and acoustic feature data of the target speaker's speech (target speaker's speaker ID) are obtained as inputs to the speaker speech signal extraction model. The mixed speech feature data is then processed in the time-frequency domain (e.g., STFT transform) and used as input to a CNN. The output of the CNN and the target speaker's speaker ID are used as inputs to an LSTM. Finally, after processing by a fully connected layer, the probability of the target speaker's speech presence in the mixed speech signal is output. Further, the probability of the target speaker's speech presence is multiplied by the time-frequency information of the mixed speech signal, and then subjected to an inverse time-frequency transform (e.g., ISTFT transform) to obtain the enhanced speech signal of the target speaker.

[0108] The technical solution of this application can produce the following beneficial effects:

[0109] 1) When users use this system, they can accurately obtain the voice of the target speaker, resulting in a good communication experience and usage effect.

[0110] 2) This scheme can automatically adjust the speaker ID of the target speaker based on the speaker's speech in different states, significantly improving the speech recognition rate. For example, as the speech characteristics of the target speaker change, the speaker ID will also be updated, thus enabling more accurate recognition of the target speaker's speech, ultimately achieving a speaker-adaptive speech enhancement effect.

[0111] 3) This model can also suppress ambient noise, making the user's auditory experience better.

[0112] The solution in this application has the following innovative features:

[0113] 1) Accurate extraction of the target speaker's voiceprint features. This solution incorporates an adaptive update module for the target speaker's speaker ID, enabling the extraction and recognition of speaker voiceprint features across a wider range of scenarios. It addresses the issue of low voiceprint recognition rates and negative impacts on speech enhancement caused by changes in the speaker's acoustic features due to variations in age and state.

[0114] 2) Suppression of ambient noise around the speaker. This scheme can not only separate the speech signals of multiple speakers and extract the speech signal of the target speaker, but also effectively suppress ambient noise, thereby achieving better adaptive speech enhancement.

[0115] It is understood that in the embodiments of this application... Figures 1 to 4 The examples provided are merely to facilitate understanding of the embodiments of this application by those skilled in the art, and are not intended to limit the embodiments of this application to the specific scenarios illustrated. Figures 1 to 4 The examples are obviously subject to various equivalent modifications or changes, and such modifications or changes also fall within the scope of the embodiments of this application.

[0116] It is also understood that the solutions in the various embodiments of this application can be used in reasonable combinations, and the explanations or descriptions of the various terms appearing in the embodiments can be referenced or explained to each other in the various embodiments, without limitation.

[0117] It should also be understood that the various numerical sequences in the embodiments of this application do not imply the order of execution, but are merely a distinction for the convenience of description, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0118] Based on the foregoing embodiments, this application provides a voice processing device, which includes the included modules and can be implemented by a processor in a terminal; of course, it can also be implemented by specific logic circuits; in the implementation process, the processor can be a central processing unit (CPU), a microprocessor unit (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0119] Figure 5 The diagram illustrates the structural composition of a speech processing apparatus 500 provided in an embodiment of this application. For example... Figure 5 As shown, the device 500 may include:

[0120] The first acquisition module 501 is used to acquire the first speech signal to be processed; the filtering module 502 is used to filter the speech signals other than the speech signal of the target speaker in the first speech signal based on the first voiceprint features of the target speaker, and obtain the second speech signal; the updating module 503 is used to update the first voiceprint features of the target speaker based on the second speech signal, and obtain the second voiceprint features of the target speaker.

[0121] In some embodiments, the apparatus 500 further includes: a first determining module, configured to determine the similarity between the second speech signal and the first voiceprint feature of the target speaker; and an updating module 503, specifically configured to: update the first voiceprint feature of the target speaker based on the second speech signal if the similarity between the second speech signal and the first voiceprint feature of the target speaker is greater than or equal to a first threshold.

[0122] In some embodiments, the updating module 503 includes: a weighting submodule, used to perform weighting processing on the second speech signal and the first voiceprint feature of the target speaker to obtain a weighted speech signal; and a determining submodule, used to determine the weighted speech signal as the second voiceprint feature of the target speaker.

[0123] In some embodiments, the apparatus 500 further includes: a second determining module, configured to determine the weight of the second speech signal and the weight of the first voiceprint feature of the target speaker; and a weighting submodule, specifically configured to: perform weighted processing on the second speech signal and the first voiceprint feature of the target speaker based on the weight of the second speech signal and the weight of the first voiceprint feature of the target speaker.

[0124] In some embodiments, the second determining module is specifically configured to: determine the weight of the second speech signal and the weight of the first voiceprint feature of the target speaker based on the similarity between the second speech signal and the first voiceprint feature of the target speaker; wherein the similarity between the second speech signal and the first voiceprint feature of the target speaker is negatively correlated with the weight of the second speech signal.

[0125] In some embodiments, the second determining module is specifically configured to: determine the time difference between the time when the first speech signal to be processed is generated and the time when the first voiceprint feature of the target speaker is generated; and determine the weight of the second speech signal and the weight of the first voiceprint feature of the target speaker based on the time difference; wherein the time difference is positively correlated with the weight of the second speech signal.

[0126] In some embodiments, the weight of the second speech signal is less than the weight of the first voiceprint feature of the target speaker.

[0127] In some embodiments, the filtering module 502 is specifically used to: perform similarity matching between the first voiceprint feature of the target speaker and each speech signal component in the first speech signal to obtain a probability value of each speech signal component being the speech signal of the target speaker; and multiply each probability value by the corresponding speech signal component to obtain a second speech signal.

[0128] In some embodiments, the device 500 further includes: a second acquisition module, configured to acquire a first voiceprint feature of the target speaker, wherein the first voiceprint feature is obtained by feature extraction from a clean speech signal of the target speaker; wherein in the clean speech signal, the ratio of the target speaker's speech signal to the noise signal is greater than or equal to a second threshold.

[0129] The descriptions of the apparatus embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. In some embodiments, the functions or modules included in the apparatus provided in this application can be used to perform the methods described in the method embodiments above. For technical details not disclosed in the apparatus embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0130] It should be noted that, in the embodiments of this application, if the above-described methods are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware, software, or firmware, or any combination of hardware, software, and firmware.

[0131] This application provides a voice processing device, including a memory and a processor. The memory stores at least one program or at least one instruction that can be executed by the processor. When the processor executes the program or instruction, it implements some or all of the steps in the above-described method. In some embodiments, the voice processing device is a chip.

[0132] This application provides a chip including a processor and a communication interface. The processor reads instructions stored in the memory through the communication interface to implement some or all of the steps in the above method. In some...

[0133] In one embodiment, as an implementation, the chip further includes a memory that stores at least one program or at least one instruction. The processor is used to execute the program or instruction stored in the memory. When the program or instruction is executed, the processor is used to perform some or all of the steps in the above method.

[0134] This application provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements some or all of the steps in the above-described method. The computer-readable storage medium can be transient or non-transient.

[0135] This application provides a computer program including computer-readable code, wherein when the computer-readable code is executed in a computer device, a processor in the computer device performs some or all of the steps in the above-described method.

[0136] This application provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program. When the computer program is read and executed by a computer, it implements some or all of the steps in the above-described method. This computer program product can be implemented specifically through hardware, software, or a combination thereof. In some embodiments, the computer program product specifically...

[0137] In some embodiments, the computer program product is embodied as a computer storage medium, and in others, it is specifically embodied as a software product, such as a software development kit (SDK).

[0138] It should be noted that the descriptions of the various embodiments above tend to emphasize the differences between them, while their similarities or commonalities can be referenced interchangeably. The descriptions of the above embodiments of the device, storage medium, computer program, and computer program product are similar to the descriptions of the above method embodiments, and are therefore consistent with each other.

[0139] Similar beneficial effects can be achieved with the method embodiments. For technical details not disclosed in the device, storage medium, computer program, and computer program product embodiments of this application, please refer to the description of the method embodiments of this application.

[0140] And understanding.

[0141] It should be noted that, Figure 6 This is a schematic diagram of a hardware entity of the voice processing device in an embodiment of this application, such as... Figure 6 As shown, the hardware entity of the voice processing device 600 includes: a processor 601, a communication interface 602, and a memory 603, wherein:

[0142] The processor 601 typically controls the overall operation of the voice processing device 600.

[0143] The communication interface 602 enables the voice processing device 600 to communicate with other terminals or servers via a network.

[0144] The memory 603 is configured to store instructions and applications executable by the processor 601, and can also cache data to be processed or already processed (e.g., image data, audio data, voice communication data, and video communication data) in the processor 601 and various modules in the voice processing device 600. It can be implemented using flash memory or random access memory (RAM). Data transfer between the processor 601, the communication interface 602, and the memory 603 can be performed via bus 604.

[0145] The voice processing device 600 can implement the corresponding processes in the various methods of the embodiments of this application, which will not be described in detail here for the sake of brevity.

[0146] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above steps / processes do not imply a sequential order of execution; the execution order of each step / process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above embodiments of this application are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0147] 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.

[0148] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0149] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0150] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0151] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0152] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence or the part that contributes to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, magnetic disks, or optical disks.

[0153] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method of speech processing, characterized by, The method comprises: acquiring a first voice signal to be processed; filtering, based on a first voiceprint feature of a target speaker, voice signals other than voice signals of the target speaker in the first voice signal, to obtain a second voice signal; updating the first voiceprint feature of the target speaker based on the second voice signal, to obtain a second voiceprint feature of the target speaker; wherein the updating of the first voiceprint feature of the target speaker based on the second voice signal, to obtain the second voiceprint feature of the target speaker, comprises: determining a time difference between a time when the first voice signal to be processed is generated and a time when the first voiceprint feature of the target speaker is generated; determining a weight of the second voice signal and a weight of the first voiceprint feature of the target speaker based on the time difference; wherein the time difference is positively correlated with the weight of the second voice signal; performing weighted processing on the second voice signal and the first voiceprint feature of the target speaker based on the weight of the second voice signal and the weight of the first voiceprint feature of the target speaker, to obtain a weighted voice signal; determining the weighted voice signal as the second voiceprint feature of the target speaker.

2. The method of claim 1, wherein, The method further comprises: determining a similarity between the second voice signal and the first voiceprint feature of the target speaker; the updating of the first voiceprint feature of the target speaker based on the second voice signal, to obtain the second voiceprint feature of the target speaker, comprises: in a case where the similarity between the second voice signal and the first voiceprint feature of the target speaker is greater than or equal to a first threshold, updating the first voiceprint feature of the target speaker based on the second voice signal.

3. The method of claim 1, wherein, The determining of the weight of the second voice signal and the weight of the first voiceprint feature of the target speaker comprises: determining the weight of the second voice signal and the weight of the first voiceprint feature of the target speaker based on the similarity between the second voice signal and the first voiceprint feature of the target speaker; wherein the similarity between the second voice signal and the first voiceprint feature of the target speaker is negatively correlated with the weight of the second voice signal.

4. The method of any one of claims 1 to 3, characterized in that: the weight of the second voice signal is less than the weight of the first voiceprint feature of the target speaker.

5. The method according to claim 1 or 2, characterized in that, The filtering, based on the first voiceprint feature of the target speaker, of voice signals other than voice signals of the target speaker in the first voice signal, to obtain a second voice signal, comprises: performing similarity matching between the first voiceprint feature of the target speaker and each voice signal component in the first voice signal, to obtain a probability value that each voice signal component is a voice signal of the target speaker; multiplying each probability value with a corresponding voice signal component, to obtain the second voice signal.

6. The method of claim 1 or 2, wherein, The method further comprises: The first voiceprint feature of the target speaker is obtained, and the first voiceprint feature is obtained by performing feature extraction on a clean speech signal of the target speaker; in the clean speech signal, a ratio of a speech signal of the target speaker to a noise signal is greater than or equal to a second threshold.

7. An apparatus for speech processing, characterized by The device comprises: A first obtaining module is configured to obtain a first speech signal to be processed. A filtering module is configured to filter, based on a first voiceprint feature of a target speaker, speech signals other than a speech signal of the target speaker in the first speech signal, to obtain a second speech signal. An updating module is configured to update the first voiceprint feature of the target speaker based on the second speech signal, to obtain a second voiceprint feature of the target speaker. The device further comprises a second determining module, and the updating module comprises a weighting submodule and a determining submodule. The second determining module is configured to: determine a time difference between a time when the first speech signal to be processed is generated and a time when the first voiceprint feature of the target speaker is generated; and based on the time difference, determine a weight of the second speech signal and a weight of the first voiceprint feature of the target speaker; wherein the time difference is positively correlated with the weight of the second speech signal. The weighting submodule is configured to perform weighting processing on the second speech signal and the first voiceprint feature of the target speaker based on the weight of the second speech signal and the weight of the first voiceprint feature of the target speaker, to obtain a weighted speech signal. The determining submodule is configured to determine the weighted speech signal as the second voiceprint feature of the target speaker.

8. A speech processing device, characterized by The speech processing device comprises: A memory is configured to store at least one program or at least one instruction. A processor is connected to the memory and is configured to execute the program or the instruction, to implement the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by at least one processor to implement the method of any one of claims 1 to 6.

Images