WO2018113243A1 - 语音分割的方法、装置、设备及计算机存储介质 - Google Patents

语音分割的方法、装置、设备及计算机存储介质 Download PDF

Info

Publication number
WO2018113243A1
WO2018113243A1 PCT/CN2017/091310 CN2017091310W WO2018113243A1 WO 2018113243 A1 WO2018113243 A1 WO 2018113243A1 CN 2017091310 W CN2017091310 W CN 2017091310W WO 2018113243 A1 WO2018113243 A1 WO 2018113243A1
Authority
WO
WIPO (PCT)
Prior art keywords
segment
voice
speaker
mixed
segmentation
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2017/091310
Other languages
English (en)
French (fr)
Inventor
王健宗
郭卉
肖京
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Publication of WO2018113243A1 publication Critical patent/WO2018113243A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/04—Segmentation; Word boundary detection

Definitions

  • the present invention relates to the field of voice processing technologies, and in particular, to a voice segmentation method, apparatus, device, and computer storage medium.
  • Speech segmentation means that in the field of speech processing, when the speeches of multiple speakers are combined and recorded in one channel, the speech of each speaker in the signal is separately extracted.
  • the traditional speech segmentation technique is based on the global background model and the Gaussian mixture model. Due to the limitation of technology, the segmentation accuracy of this segmentation method is not high, especially for dialogues with frequent alternating and overlapping dialogue segmentation effects. difference.
  • the present invention provides a method for voice segmentation, characterized in that the method for voice segmentation comprises:
  • the present invention also provides a device for voice segmentation, the device for voice segmentation comprising:
  • a segmentation module configured to divide the mixed voice into a plurality of phrase segments and receive a corresponding speaker identifier for each phrase segment when receiving the mixed voice sent by the terminal;
  • An adjustment module configured to establish a voiceprint model for a phrase segment corresponding to each speaker identifier by using a time recurrent neural network, and adjust a corresponding segmentation boundary in the mixed voice based on the voiceprint model to segment each speaker identifier correspondingly Effective voice segment.
  • the present invention also provides a voice segmentation device including a memory, a processor, and a language stored on the memory and executable on the processor a tone segmentation program that, when executed by the processor, implements the following steps:
  • the present invention also provides a computer readable storage medium having a voice segmentation program stored thereon, the voice segmentation program being executed by a processor to implement the following steps:
  • the invention has the beneficial effects that the present invention first divides the mixed speech into a plurality of phrase segments, and each phrase segment corresponds to identify a speaker, and uses a time recurrent neural network to establish a voiceprint model for each phrase segment. Since the voiceprint model established by using the time recurrent neural network can correlate the sound information of the speaker across time points, the adjustment of the segmentation boundary of the phrase segment can be realized based on the voiceprint model, which can effectively improve the accuracy of voice segmentation, especially for The dialogue is frequent and there are overlapping voices, and the effect of voice segmentation is better.
  • FIG. 1 is a schematic flow chart of an embodiment of a method for voice segmentation according to the present invention
  • step S1 shown in FIG. 1;
  • step S2 shown in FIG. 1;
  • FIG. 4 is a schematic structural diagram of an apparatus for voice segmentation according to an embodiment of the present invention.
  • Figure 5 is a schematic structural view of the dividing module shown in Figure 4.
  • FIG. 6 is a schematic structural view of the adjustment module shown in FIG. 4;
  • FIG. 7 is a schematic structural diagram of a device in a hardware operating environment according to an embodiment of the present invention.
  • FIG. 1 is a schematic flowchart of a method for voice segmentation according to an embodiment of the present invention.
  • the voice segmentation method includes the following steps:
  • Step S1 when receiving the mixed voice sent by the terminal, dividing the mixed voice into a plurality of phrase segments, and marking a corresponding speaker identifier for each phrase segment;
  • the method proposed in this embodiment can be applied to an automatic answering system of a call center, such as an automatic answering system of an insurance call center, an automatic answering system of various customer service call centers, and the like.
  • Automatic answer The system receives the original mixed voice sent by the terminal, and the mixed voice is mixed with sounds generated by a plurality of different sound sources, such as a sound mixed by a plurality of people, a sound mixed by a plurality of people and a sound mixed with other noises, and the like.
  • the mixed voice can be segmented into a plurality of phrase segments by using a predetermined method.
  • a Gaussian Mixture Model (GMM) can be used to divide the mixed voice into a plurality of phrase segments.
  • GMM Gaussian Mixture Model
  • the method divides the mixed speech into multiple phrase segments.
  • each phrase segment should correspond to only one speaker, and there may be multiple phrase segments in the different phrase segments belonging to the same speaker, which will be different from the same speaker.
  • the voice segments are labeled the same.
  • Step S2 using a time recurrent neural network to establish a voiceprint model for a phrase segment corresponding to each speaker identifier, and adjusting a corresponding segmentation boundary in the mixed voice based on the voiceprint model to segment the effective corresponding to each speaker identifier Voice segment.
  • the Long-Short Term Memory (LSTM) model has a directional loop introduced by a recurrent neural network in a traditional forward feedback neural network to process the association before and after the input between layers and before and after the output in the layer. .
  • the speech signal characteristics across time points can be obtained, which can be used to process the speech sequence with associated information at any length and any position.
  • the time recurrent neural network model can design the multiple interaction layers in the neural network layer to memorize the information on the nodes at the farther time.
  • the “forget gate layer” is used to discard the information that is not related to the recognition task, and then Use the “input gate layer” to determine the state that needs to be updated, and finally determine the state that needs to be output and process the output.
  • a voice pattern is established by using a time recurrent neural network for a phrase segment corresponding to each speaker identifier, and the voice information of the speaker crossing the time point can be obtained by the voiceprint model, and the corresponding voice can be adjusted based on the voice information.
  • the segmentation boundary adjusts the segmentation boundary of all the phrase segments corresponding to each speaker, and finally divides the effective segment of speech corresponding to each speaker identifier, and the valid segment can be regarded as the complete speech of the corresponding speaker.
  • the embodiment first divides the mixed speech into a plurality of phrase segments, and each phrase segment corresponds to identify a speaker, and uses a time recurrent neural network to establish a voiceprint model for each phrase segment. Since the voiceprint model established by using the time recurrent neural network can correlate the sound information of the speaker across time points, the adjustment of the segmentation boundary of the phrase segment can be realized based on the voiceprint model, which can effectively improve the accuracy of voice segmentation, especially For speech with frequent alternating and overlapping voices, the effect of speech segmentation is better.
  • the foregoing step S1 includes:
  • Step S11 Acquire a silent segment in the mixed voice, and remove a silent segment in the mixed voice, to segment the mixed voice according to the silent segment, to obtain a long segment after the segmentation;
  • Step S12 framing the long speech segment to extract acoustic features of each long speech segment
  • Step S13 performing KL distance analysis on the acoustic features of each long speech segment, according to the KL distance
  • the segmentation segment is segmented by the analysis result to obtain a segmented phrase segment.
  • step S14 the speech segmentation of each phrase segment is performed by using a Gaussian mixture model, and the corresponding speaker identifier is marked with the phrase segment of the same speech class.
  • the initial segmentation is first performed according to the mute: determining the mute segment in the mixed speech, and removing the determined mute segment from the mixed speech, so as to realize the segmentation of the mixed speech according to the mute segment, and the mute segment is through the mixed speech.
  • the analysis of short-term speech energy and short-term zero-crossing rate is determined.
  • each person speaks for a fixed threshold Tu. If a certain voice is greater than the duration, more people may speak. If it is less than the duration, it is more likely that only one person will speak. Based on this assumption, the inter-frame KL distance analysis can be performed on the acoustic features of the speech segment whose silence length is longer than the fixed threshold Tu for each long speech segment. Of course, inter-frame KL distance analysis can also be performed on the acoustic characteristics of all long speech segments.
  • the obtained long speech segment is framed to obtain a speech frame of each long speech segment, the acoustic features of the speech frame are extracted, and KL distance (ie, relative entropy) is analyzed for the acoustic features of all the long speech segments.
  • KL distance ie, relative entropy
  • acoustic characteristics include, but are not limited to, linear prediction coefficients, cepstral coefficients MFCC, average zero-crossing rate, short-term spectrum, formant frequency, and bandwidth.
  • the KL distance is larger, the difference between the two PQs is larger, that is, the two sets of PQ are from the voices of two different people.
  • the long speech segment whose duration is greater than the preset time threshold is segmented at the maximum value of KL to improve the accuracy of speech segmentation.
  • the long speech segment is segmented to obtain a phrase segment, and the number of phrase segments is greater than the number of long segments.
  • the phrase segment clustering is performed: clustering the segmented phrase segments to aggregate all the phrase segments into a plurality of phonetic classes, and labeling the corresponding speaker segments for each phrase segment, wherein, the same speaker identifier
  • the phrase segments of the speech class are labeled with the same speaker identifier, and the phrase segments that do not belong to the same phonetic class are labeled with different speaker identifiers.
  • the clustering method is: using the Gaussian mixture model of K components to fit each segment of the phrase, using the mean as the feature vector, and using k-means clustering method to cluster all the phrase segments into multiple categories.
  • the foregoing step S2 includes:
  • Step S21 using the time recurrent neural network to establish a voiceprint model for a phrase segment corresponding to each speaker identifier, and extracting a preset type vector representing the speaker identity feature based on the voiceprint model;
  • Step S22 Calculate, according to the preset type vector, a maximum posterior probability that each voice frame belongs to a corresponding speaker;
  • Step S23 adjusting the mixed Gaussian model of the speaker based on the maximum posterior probability and using a predetermined algorithm
  • Step S24 Acquire a speaker with the highest probability corresponding to each voice frame based on the adjusted mixed Gaussian model, and adjust a corresponding segmentation boundary in the mixed voice according to a probability relationship between the speaker with the highest probability and the voice frame;
  • Step S25 iteratively updating the voiceprint model n times, and iterating m times the mixed Gaussian model each time the voiceprint model is updated to obtain an effective voice segment corresponding to each speaker, where n and m are both greater than 1.
  • a temporal recursive neural network is used to establish a voiceprint model for a phrase segment corresponding to each speaker identifier, and a preset type vector representing the speaker identity feature is extracted based on the voiceprint model, preferably, the preset type
  • the vector is an i-vector vector, and the i-vector vector is an important feature that reflects the acoustic difference of the speaker.
  • the maximum posterior probability that each speech frame belongs to a certain speaker is calculated according to the preset type vector, and the mixed Gaussian model of the speaker is re-adjusted by the preset algorithm in the mixed speech by calculating the maximum posterior probability.
  • the speaker's mixed Gaussian model is re-adjusted by the Baum-Welch algorithm, which is a set of k (typically 3-5) Gaussian models.
  • the re-adjusted mixed Gaussian model is used to find the speaker with the highest probability of each speech frame.
  • the segmentation boundary of the mixed speech is adjusted according to the probability relationship between the speech frame and the found speaker, for example, the segmentation boundary is fine-tuned forward or fine-tuned backward.
  • the above-mentioned voiceprint model is iteratively updated n times, and the Gaussian model is iterated m times each time the voiceprint model is updated to obtain an effective speech segment corresponding to each speaker, and n and m are positive integers greater than 1.
  • a voiceprint model is established by means of a deep learning time recurrent neural network, and each voice frame corresponding to each speaker voice pattern is used to calculate a probability that the voice frame belongs to a certain speaker, and the model is modified based on the probability, and finally adjusted.
  • the boundary of speech segmentation can effectively improve the accuracy of speaker speech segmentation, reduce the error rate, and have good scalability.
  • the method further includes: acquiring corresponding response content based on the valid voice segment, and feeding back the response content to the terminal. .
  • the automatic response system is associated with a corresponding response library, and the response library stores response content corresponding to different questions.
  • the automatic response system After receiving the mixed voice sent by the terminal, the automatic response system divides the response into a speaker identifier. An effective voice segment, from which an effective voice segment related to the automatic answering system is obtained, and the valid voice segment is matched in the response library, and the matched response content is fed back to the terminal.
  • FIG. 4 is a schematic structural diagram of an apparatus for voice segmentation according to an embodiment of the present invention.
  • the device for voice segmentation includes:
  • the segmentation module 101 is configured to, when receiving the mixed voice sent by the terminal, divide the mixed voice into a plurality of phrase segments, and mark a corresponding speaker identifier for each phrase segment;
  • the apparatus for voice segmentation of the present embodiment can be applied to an automatic answering system such as an automatic answering system of an insurance call center, an automatic answering system of various customer service call centers, and the like.
  • Speech segmentation The dividing module 101 of the device receives the original mixed voice sent by the terminal, and the mixed voice is mixed with sounds generated by a plurality of different sound sources, for example, a voice mixed by a plurality of people, and a voice of a plurality of people is mixed with other noises. Sound and so on.
  • the mixed voice can be segmented into a plurality of phrase segments by using a predetermined method.
  • a Gaussian Mixture Model (GMM) can be used to divide the mixed voice into a plurality of phrase segments.
  • GMM Gaussian Mixture Model
  • the method divides the mixed speech into multiple phrase segments.
  • each phrase segment should correspond to only one speaker, and there may be multiple phrase segments in the different phrase segments belonging to the same speaker, which will be different from the same speaker.
  • the voice segments are labeled the same.
  • the adjusting module 102 is configured to establish a voiceprint model for a phrase segment corresponding to each speaker identifier by using a time recurrent neural network, and adjust a corresponding segmentation boundary in the mixed voice based on the voiceprint model to segment each speaker identifier Corresponding valid voice segment.
  • the Long-Short Term Memory (LSTM) model has a directional loop introduced by a recurrent neural network in a traditional forward feedback neural network to process the association before and after the input between layers and before and after the output in the layer. .
  • the speech signal characteristics across time points can be obtained, which can be used to process the speech sequence with associated information at any length and any position.
  • the time recurrent neural network model can design the multiple interaction layers in the neural network layer to memorize the information on the nodes at the farther time.
  • the “forget gate layer” is used to discard the information that is not related to the recognition task, and then Use the “input gate layer” to determine the state that needs to be updated, and finally determine the state that needs to be output and process the output.
  • a voice pattern is established by using a time recurrent neural network for a phrase segment corresponding to each speaker identifier, and the voice information of the speaker crossing the time point can be obtained by the voiceprint model, and the corresponding voice can be adjusted based on the voice information.
  • the segmentation boundary adjusts the segmentation boundary of all the phrase segments corresponding to each speaker, and finally divides the effective segment of speech corresponding to each speaker identifier, and the valid segment can be regarded as the complete speech of the corresponding speaker.
  • the segmentation module 101 includes:
  • the removing unit 1011 is configured to acquire a silent segment in the mixed voice, and remove a silent segment in the mixed voice to segment the mixed voice according to the silent segment to obtain a segmented long voice segment.
  • a framing unit 1012 configured to framing the long speech segment to extract an acoustic feature of each long speech segment
  • the segmentation unit 1013 is configured to perform KL distance analysis on the acoustic features of each long segment of speech, and segment the segment according to the KL distance analysis result to obtain a segmented phrase segment;
  • the clustering unit 1014 is configured to perform voice clustering on each phrase segment by using a Gaussian mixture model, and mark a corresponding speaker identifier for the phrase segment of the same phonetic class.
  • the initial segmentation is first performed according to the mute: determining the mute segment in the mixed speech, The determined silence segment is removed from the mixed speech to achieve splitting of the mixed speech according to the silent segment, which is determined by analyzing the short-term speech energy and the short-term zero-crossing rate of the mixed speech.
  • each person speaks for a fixed threshold Tu. If a certain voice is greater than the duration, more people may speak. If it is less than the duration, it is more likely that only one person will speak. Based on this assumption, the inter-frame KL distance analysis can be performed on the acoustic features of the speech segment whose silence length is longer than the fixed threshold Tu for each long speech segment. Of course, inter-frame KL distance analysis can also be performed on the acoustic characteristics of all long speech segments.
  • the obtained long speech segment is framed to obtain a speech frame of each long speech segment, the acoustic features of the speech frame are extracted, and KL distance (ie, relative entropy) is analyzed for the acoustic features of all the long speech segments.
  • KL distance ie, relative entropy
  • acoustic characteristics include, but are not limited to, linear prediction coefficients, cepstral coefficients MFCC, average zero-crossing rate, short-term spectrum, formant frequency, and bandwidth.
  • the KL distance is larger, the difference between the two PQs is larger, that is, the two sets of PQ are from the voices of two different people.
  • the long speech segment whose duration is greater than the preset time threshold is segmented at the maximum value of KL to improve the accuracy of speech segmentation.
  • the long speech segment is segmented to obtain a phrase segment, and the number of phrase segments is greater than the number of long segments.
  • the phrase segment clustering is performed: clustering the segmented phrase segments to aggregate all the phrase segments into a plurality of phonetic classes, and labeling the corresponding speaker segments for each phrase segment, wherein, the same speaker identifier
  • the phrase segments of the speech class are labeled with the same speaker identifier, and the phrase segments that do not belong to the same phonetic class are labeled with different speaker identifiers.
  • the clustering method is: using the Gaussian mixture model of K components to fit each segment of the phrase, using the mean as the feature vector, and using k-means clustering method to cluster all the phrase segments into multiple categories.
  • the adjustment module 102 includes:
  • the modeling unit 1021 is configured to establish, by using the time recurrent neural network, a voiceprint model for a phrase segment corresponding to each speaker identifier, and extract a preset type vector that characterizes the speaker identity feature based on the voiceprint model;
  • the calculating unit 1022 is configured to calculate, according to the preset type vector, a maximum posterior probability that each voice frame belongs to a corresponding speaker;
  • a first adjusting unit 1023 configured to adjust the mixed Gaussian model of the speaker by using a predetermined algorithm based on the maximum posterior probability
  • the second adjusting unit 1024 is configured to acquire a speaker with the highest probability corresponding to each voice frame based on the adjusted mixed Gaussian model, and adjust a corresponding segment in the mixed voice according to a probability relationship between the speaker with the highest probability and the voice frame. boundary;
  • An iteration unit 1025 configured to iteratively update the voiceprint model n times, each time updating the voice pattern module
  • the type is iterative m times the mixed Gaussian model to obtain an effective speech segment corresponding to each speaker, and both n and m are positive integers greater than one.
  • a temporal recursive neural network is used to establish a voiceprint model for a phrase segment corresponding to each speaker identifier, and a preset type vector representing the speaker identity feature is extracted based on the voiceprint model, preferably, the preset type
  • the vector is an i-vector vector, and the i-vector vector is an important feature that reflects the acoustic difference of the speaker.
  • the maximum posterior probability that each speech frame belongs to a certain speaker is calculated according to the preset type vector, and the mixed Gaussian model of the speaker is re-adjusted by the preset algorithm in the mixed speech by calculating the maximum posterior probability.
  • the speaker's mixed Gaussian model is re-adjusted by the Baum-Welch algorithm, which is a set of k (typically 3-5) Gaussian models.
  • the re-adjusted mixed Gaussian model is used to find the speaker with the highest probability of each speech frame.
  • the segmentation boundary of the mixed speech is adjusted according to the probability relationship between the speech frame and the found speaker, for example, the segmentation boundary is fine-tuned forward or fine-tuned backward.
  • the above-mentioned voiceprint model is iteratively updated n times, and the Gaussian model is iterated m times each time the voiceprint model is updated to obtain an effective speech segment corresponding to each speaker, and n and m are positive integers greater than 1.
  • a voiceprint model is established by means of a deep learning time recurrent neural network, and each voice frame corresponding to each speaker voice pattern is used to calculate a probability that the voice frame belongs to a certain speaker, and the model is modified based on the probability, and finally adjusted.
  • the boundary of speech segmentation can effectively improve the accuracy of speaker speech segmentation, reduce the error rate, and have good scalability.
  • the apparatus for voice segmentation further includes: a feedback module, configured to acquire a corresponding response content based on the valid voice segment, and feed back the response content. To the terminal.
  • the automatic response system is associated with a corresponding response library, and the response library stores response content corresponding to different questions.
  • the segmentation module 101 divides the speaker into a speaker. Identifying a valid voice segment corresponding to the active voice segment, obtaining a valid voice segment related to the automatic answering system, matching the valid voice segment in the response library, and the feedback module feeding back the matched response content to the terminal .
  • FIG. 7 is a schematic structural diagram of a device in a hardware operating environment according to an embodiment of the present invention.
  • the voice segmentation device in the embodiment of the present invention may be a PC (Personal Computer), or may be a terminal device such as a smart phone, a tablet computer, an e-book reader, or a portable computer.
  • PC Personal Computer
  • terminal device such as a smart phone, a tablet computer, an e-book reader, or a portable computer.
  • the device includes a memory 11, a processor 12, a communication bus 13, and a network interface 14.
  • the memory 11 includes at least one type of readable storage medium including a flash memory, a hard disk, a multimedia card, a card type memory (for example, SD or DX memory, etc.), a random access memory (RAM), and a static memory.
  • the memory 11 may be an internal storage unit of the voice segmentation device, such as a hard disk or memory of the voice segmentation device, in some embodiments.
  • the memory 11 may also be an external storage device of the voice segmentation device in other embodiments, such as a plug-in hard disk equipped on the voice segmentation device, a smart memory card (SMC), and a secure digital (SD). Card, flash card, etc.
  • the memory 11 may also include both an internal storage unit of the voice segmentation device and an external storage device.
  • the memory 11 can be used not only for storing application software installed in the voice segmentation device and various types of data, such as codes of a voice segmentation program, but also for temporarily storing data that has been output or is to be output.
  • the processor 12 may be a Central Processing Unit (CPU), controller, microcontroller, microprocessor or other data processing chip for running program code or processing stored in the memory 11. Data, such as performing a voice segmentation program or the like.
  • CPU Central Processing Unit
  • controller microcontroller
  • microprocessor or other data processing chip for running program code or processing stored in the memory 11.
  • Data such as performing a voice segmentation program or the like.
  • Communication bus 13 is used to implement connection communication between these components.
  • the network interface 14 can optionally include a standard wired interface, a wireless interface (such as a WI-FI interface), and is typically used to establish a communication connection between the device and other electronic devices.
  • a standard wired interface such as a WI-FI interface
  • Figure 7 shows only the speech segmentation device with components 11-14 and the speech segmentation program, but it should be understood that not all illustrated components may be implemented, and more or fewer components may be implemented instead.
  • the device may further include a user interface
  • the user interface may include a display
  • an input unit such as a keyboard
  • the optional user interface may further include a standard wired interface and a wireless interface.
  • the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch sensor, or the like.
  • the display may also be appropriately referred to as a display screen or display unit for displaying information processed in the voice segmentation device and a user interface for displaying visualization.
  • the memory 11 as a computer storage medium may include an operating system and a voice segmentation program; wherein the network interface 14 is configured to perform a communication connection with the terminal to receive the mixture sent by the terminal.
  • Voice when the processor 12 executes the voice segmentation program stored in the memory 11, the following steps are implemented:
  • step S13 includes:
  • the KL distance analysis is performed on the acoustic features of each long speech segment, and the long speech segment whose duration is greater than the preset time threshold is segmented at the maximum value of the KL distance to obtain a segmented phrase segment.
  • step S1 includes:
  • S11 Obtain a silence segment in the mixed voice, and remove a silence segment in the mixed voice, to segment the mixed voice according to the silence segment, to obtain a segmented long voice segment;
  • S14 Perform a speech clustering on each phrase segment by using a Gaussian mixture model, and mark a corresponding speaker identifier for the phrase segment of the same phonetic class.
  • step S2 the method further includes:
  • the present invention also provides a computer readable storage medium having a speech segmentation program stored thereon, the speech segmentation program being executed by a processor to implement the following steps:
  • step S1 includes:
  • S11 Obtain a silence segment in the mixed voice, and remove a silence segment in the mixed voice, to segment the mixed voice according to the silence segment, to obtain a segmented long voice segment;
  • S14 Perform a speech clustering on each phrase segment by using a Gaussian mixture model, and mark a corresponding speaker identifier for the phrase segment of the same phonetic class.
  • step S13 includes:
  • the KL distance analysis is performed on the acoustic features of each long speech segment, and the long speech segment whose duration is greater than the preset time threshold is segmented at the maximum value of the KL distance to obtain a segmented phrase segment.
  • step S2 includes:
  • the foregoing embodiment method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be through hardware, but in many cases, the former is better.
  • Implementation Based on such understanding, the technical solution of the present invention, which is essential or contributes to the prior art, may be embodied in the form of a software product stored in a storage medium (such as ROM/RAM, disk,
  • the optical disc includes a number of instructions for causing a terminal device (which may be a cell phone, a computer, a server, or a network device, etc.) to perform the methods described in various embodiments of the present invention.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Telephonic Communication Services (AREA)
  • Machine Translation (AREA)

Abstract

一种语音分割的方法及装置，语音分割的方法包括：在接收到终端发送的混合语音时，将混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识（S1）；利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于声纹模型调整混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段（S2）。该方法能够有效提高语音分割的精度，特别是对于对话交替频繁、以及有交叠的语音，语音分割的效果较好。

Description

语音分割的方法、装置、设备及计算机存储介质
优先权申明
本申请基于巴黎公约申明享有2016年12月19日递交的申请号为CN201611176791.9、名称为“语音分割的方法及装置”的中国专利申请的优先权，该中国专利申请的整体内容以参考的方式结合在本申请中。
技术领域
本发明涉及语音处理技术领域，尤其涉及一种语音分割的方法、装置、设备及计算机存储介质。
背景技术
目前，呼叫中心接收到的语音很多都混杂有多人的语音，这时需要先对语音进行语音分割(speaker diarization)，才能进一步对目标语音进行语音分析。语音分割是指：在语音处理领域，当多个说话人的语音被合并录在一个声道中时，把信号中每个说话人的语音分别进行提取。传统的语音分割技术是基于全局背景模型和高斯混合模型进行分割，由于技术的限制，这种语音分割的方法分割的精度并不高，特别是对于对话交替频繁、以及有交叠的对话分割效果差。
发明内容
本发明的目的在于提供一种语音分割的方法、装置、设备及计算机存储介质，旨在有效提高语音分割的精度。
为实现上述目的，本发明提供一种语音分割的方法，其特征在于，所述语音分割的方法包括：
S1，在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
S2，利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
为实现上述目的，本发明还提供一种语音分割的装置，所述语音分割的装置包括：
分割模块，用于在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
调整模块，用于利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
为实现上述目的，本发明还提供一种语音分割设备，所述语音分割设备包括存储器、处理器以及存储在所述存储器上并可在所述处理器上运行的语 音分割程序，所述语音分割程序被所述所述处理器执行时实现如下步骤：
S1，在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
S2，利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
为实现上述目的，本发明还提供计算机可读存储介质，所述计算机可读存储介质上存储有语音分割程序，所述语音分割程序被处理器执行时实现如下步骤：
S1，在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
S2，利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
本发明的有益效果是：本发明首先将混合语音进行分割，分割成多个短语音段，每一短语音段对应标识一个说话人，利用时间递归神经网络对各短语音段建立声纹模型，由于利用时间递归神经网络建立的声纹模型能够关联说话人跨时间点的声音信息，因此基于该声纹模型实现对短语音段的分割边界的调整，能够有效提高语音分割的精度，特别是对于对话交替频繁、以及有交叠的语音，语音分割的效果较好。
附图说明
图1为本发明语音分割的方法一实施例的流程示意图；
图2为图1所示步骤S1的细化流程示意图；
图3为图1所示步骤S2的细化流程示意图；
图4为本发明语音分割的装置一实施例的结构示意图；
图5为图4所示分割模块的结构示意图；
图6为图4所示调整模块的结构示意图；
图7为本发明实施例方案涉及的硬件运行环境的设备结构示意图。
具体实施方式
以下结合附图对本发明的原理和特征进行描述，所举实例只用于解释本发明，并非用于限定本发明的范围。
如图1所示，图1为本发明语音分割的方法一实施例的流程示意图，该语音分割的方法包括以下步骤：
步骤S1，在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
本实施例提出的方法可应用于呼叫中心的自动应答系统中，例如保险呼叫中心的自动应答系统、各种客服呼叫中心的自动应答系统等等。自动应答 系统接收到终端发送的原始的混合语音，该混合语音中混合有多种不同的声源产生的声音，例如有多人说话混合的声音，多人说话的声音与其他噪声混合的声音等等。
本实施例可以利用预定的方法将混合语音分割成多个短语音段，例如可以利用高斯混合模型(Gaussian Mixture Model，GMM)将混合语音分割成多个短语音段，当然，也可以利用其他传统的方法将混合语音分割成多个短语音段。
其中，经本实施例的语音分割后，每一短语音段应只对应一说话人，不同的短语音段中可能有多个短语音段属于同一个说话人，将同一个说话人的不同短语音段进行相同的标识。
步骤S2，利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
本实施例中，时间递归神经网络模型(Long-Short Term Memory，LSTM)拥有递归神经网络在传统前向反馈神经网络中引入的定向循环，用以处理层间输入前后、层内输出前后的关联。用时间递归神经网络在语音序列上建模，可以得到跨越时间点的语音信号特征，可以用于对关联信息处于任何长度、任何位置的语音序列进行处理。时间递归神经网络模型通过神经网络层内设计多个交互层，可以记忆到更远时间节点上的信息，在时间递归神经网络模型中用“忘记门层”丢弃与识别任务不相关的信息，接着用“输入门层”决定需要更新的状态，最后确定需要输出的状态并处理输出。
本实施例对于各说话人标识对应的短语音段，利用时间递归神经网络建立声纹模型，通过该声纹模型可以得到说话人跨越时间点的声音信息，基于这些声音信息可以调整混合语音中对应的分割边界，以对每一说话人对应的所有短语音段调整其分割边界，最终分割出各说话人标识对应的有效语音段，该有效语音段可以看作对应的说话人的完整语音。
与现有技术相比，本实施例首先将混合语音进行分割，分割成多个短语音段，每一短语音段对应标识一个说话人，利用时间递归神经网络对各短语音段建立声纹模型，由于利用时间递归神经网络建立的声纹模型能够关联说话人跨时间点的声音信息，因此基于该声纹模型实现对短语音段的分割边界的调整，能够有效提高语音分割的精度，特别是对于对话交替频繁、以及有交叠的语音，语音分割的效果较好。
在一优选的实施例中，如图2所示，在上述图1的实施例的基础上，上述步骤S1包括：
步骤S11，获取所述混合语音中的静音段，去除所述混合语音中的静音段，以根据所述静音段对所述混合语音进行分割，得到分割后的长语音段；
步骤S12，对所述长语音段进行分帧，以提取每一长语音段的声学特征；
步骤S13，对每一长语音段的声学特征进行KL距离分析，根据KL距 离分析结果对所述语音段进行切分，得到切分后的短语音段，
步骤S14，利用高斯混合模型对各短语音段进行语音聚类，并对同一语音类的短语音段标注对应的说话人标识。
本实施例中，首先根据静音进行初步分割：确定混合语音中的静音段，将确定的静音段从混合语音中去除，以实现将混合语音根据静音段进行分割，静音段是通过对混合语音的短时语音能量和短时过零率的分析来确定的。
去除静音段后，首先假设在整个混合语音中，每人每次讲话时长为固定阈值Tu，若某段语音大于该时长，则可能多人说话，若小于该时长，则更可能只有一个人说话，基于这种假设，可以对静音分割后的每个长语音段的时长大于固定阈值Tu的语音段的声学特征进行帧间KL距离分析。当然，也可以对所有的长语音段的声学特征进行帧间KL距离分析。具体地，对得到的长语音段进行分帧，以得到每一长语音段的语音帧，提取语音帧的声学特征，对所有长语音段的声学特征进行KL距离(也即相对熵)分析，其中，声学特征包括但不限定于线性预测系数、倒频谱系数MFCC、平均过零率、短时频谱、共振峰频率及带宽。
其中，KL距离分析的含义是对于两个离散型的声学特征概率分布集合P＝{p1，p2，…，pn}和Q＝{q1，q2，…，qn}，P和Q间的KL距离：
Figure PCTCN2017091310-appb-000001
当KL距离越大时，PQ两者差异越大，即PQ这两个集合来自两个不同人的语音。优选地，对时长大于预设时间阈值的长语音段在KL的最大值处进行切分，以提高语音分割的精度。
长语音段经过切分后得到短语音段，短语音段的数量大于长语音段的数量。然后进行短语音段聚类：对切分后的短语音段进行聚类，以将所有短语音段聚为多个语音类，并为各个短语音段标注对应的说话人标识，其中，属于同一语音类的短语音段标注相同的说话人标识，不属于同一语音类的短语音段标注不同的说话人标识。聚类方法是：采用K个成分的高斯混合模型拟合每段短语音段，以均值作为特征向量，使用k-means聚类方法把所有短语音段聚为多类。
在一优选的实施例中，如图3所示，在上述的实施例的基础上，上述步骤S2包括：
步骤S21，利用所述时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型提取表征说话人身份特征的预设类型向量；
步骤S22，基于所述预设类型向量计算每一语音帧属于对应的说话人的最大后验概率；
步骤S23，基于所述最大后验概率并利用预定算法调整该说话人的混合高斯模型；
步骤S24，基于调整后的混合高斯模型获取每一语音帧对应的概率最大的说话人，并根据概率最大的说话人与语音帧的概率关系调整所述混合语音中对应的分割边界；
步骤S25，迭代更新所述声纹模型n次，每次更新所述声纹模型时迭代m次所述混合高斯模型，以得到各说话人对应的有效语音段，n及m均为大于1的正整数。
本实施例中，利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型提取表征说话人身份特征的预设类型向量，优选地，该预设类型向量为i-vector向量，i-vector向量是反映说话人声学差异的一个重要特征。
在整个混合语音中，根据预设类型向量计算每一语音帧属于某一说话人的最大后验概率，利用计算最大后验概率，在混合语音中通过预设算法重新调整说话人的混合高斯模型，例如，通过Baum-Welch算法重新调整说话人的混合高斯模型，该混合高斯模型为k(一般为3-5个)个高斯模型的集合。利用重新调整后的混合高斯模型寻找每一语音帧概率最大的说话人。根据语音帧与寻找到的该说话人的概率关系调整混合语音的分割边界，例如将分割边界向前微调或者向后微调。最后，迭代更新上述声纹模型n次，每次更新声纹模型时迭代m次混合高斯模型，以得到各个说话人对应的有效语音段，n及m均为大于1的正整数。
本实施例借助深度学习的时间递归神经网络建立声纹模型，用各说话人声纹对应的身份特征对应各语音帧以计算语音帧属于某一说话人的概率，基于该概率修正模型，最终调整语音分割的边界，可以有效提高说话人语音分割的精度，降低错误率，且可扩展性好。
在一优选的实施例中，在上述的实施例的基础上，该方法在上述步骤S2之后还包括：基于所述有效语音段获取对应的应答内容，并将所述应答内容反馈给所述终端。
本实施例中，自动应答系统关联对应的应答库，该应答库中存储有不同的问题对应的应答内容，自动应答系统在接收到终端发送的混合语音后，将其分割为说话人标识对应的有效语音段，从这些有效语音段中获取与该自动应答系统有关问题的一个有效语音段，针对该有效语音段在应答库中进行匹配，并将匹配得到的应答内容反馈给终端。
如图4所示，图4为本发明语音分割的装置一实施例的结构示意图，该语音分割的装置包括：
分割模块101，用于在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
本实施例的语音分割的装置可以应用于自动应答系统，例如保险呼叫中心的自动应答系统、各种客服呼叫中心的自动应答系统等等。该语音分割的 装置的分割模块101接收到终端发送的原始的混合语音，该混合语音中混合有多种不同的声源产生的声音，例如有多人说话混合的声音，多人说话的声音与其他噪声混合的声音等等。
本实施例可以利用预定的方法将混合语音分割成多个短语音段，例如可以利用高斯混合模型(Gaussian Mixture Model，GMM)将混合语音分割成多个短语音段，当然，也可以利用其他传统的方法将混合语音分割成多个短语音段。
其中，经本实施例的语音分割后，每一短语音段应只对应一说话人，不同的短语音段中可能有多个短语音段属于同一个说话人，将同一个说话人的不同短语音段进行相同的标识。
调整模块102，用于利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
本实施例中，时间递归神经网络模型(Long-Short Term Memory，LSTM)拥有递归神经网络在传统前向反馈神经网络中引入的定向循环，用以处理层间输入前后、层内输出前后的关联。用时间递归神经网络在语音序列上建模，可以得到跨越时间点的语音信号特征，可以用于对关联信息处于任何长度、任何位置的语音序列进行处理。时间递归神经网络模型通过神经网络层内设计多个交互层，可以记忆到更远时间节点上的信息，在时间递归神经网络模型中用“忘记门层”丢弃与识别任务不相关的信息，接着用“输入门层”决定需要更新的状态，最后确定需要输出的状态并处理输出。
本实施例对于各说话人标识对应的短语音段，利用时间递归神经网络建立声纹模型，通过该声纹模型可以得到说话人跨越时间点的声音信息，基于这些声音信息可以调整混合语音中对应的分割边界，以对每一说话人对应的所有短语音段调整其分割边界，最终分割出各说话人标识对应的有效语音段，该有效语音段可以看作对应的说话人的完整语音。
在一优选的实施例中，如图5所示，在上述图4的实施例的基础上，上述分割模块101包括：
去除单元1011，用于获取所述混合语音中的静音段，去除所述混合语音中的静音段，以根据所述静音段对所述混合语音进行分割，得到分割后的长语音段；
分帧单元1012，用于对所述长语音段进行分帧，以提取每一长语音段的声学特征；
切分单元1013，用于对每一长语音段的声学特征进行KL距离分析，根据KL距离分析结果对所述语音段进行切分，得到切分后的短语音段；
聚类单元1014，用于利用高斯混合模型对各短语音段进行语音聚类，并对同一语音类的短语音段标注对应的说话人标识。
本实施例中，首先根据静音进行初步分割：确定混合语音中的静音段， 将确定的静音段从混合语音中去除，以实现将混合语音根据静音段进行分割，静音段是通过对混合语音的短时语音能量和短时过零率的分析来确定的。
去除静音段后，首先假设在整个混合语音中，每人每次讲话时长为固定阈值Tu，若某段语音大于该时长，则可能多人说话，若小于该时长，则更可能只有一个人说话，基于这种假设，可以对静音分割后的每个长语音段的时长大于固定阈值Tu的语音段的声学特征进行帧间KL距离分析。当然，也可以对所有的长语音段的声学特征进行帧间KL距离分析。具体地，对得到的长语音段进行分帧，以得到每一长语音段的语音帧，提取语音帧的声学特征，对所有长语音段的声学特征进行KL距离(也即相对熵)分析，其中，声学特征包括但不限定于线性预测系数、倒频谱系数MFCC、平均过零率、短时频谱、共振峰频率及带宽。
其中，KL距离分析的含义是对于两个离散型的声学特征概率分布集合P＝{p1，p2，…，pn}和Q＝{q1，q2，…，qn}，P和Q间的KL距离：
Figure PCTCN2017091310-appb-000002
当KL距离越大时，PQ两者差异越大，即PQ这两个集合来自两个不同人的语音。优选地，对时长大于预设时间阈值的长语音段在KL的最大值处进行切分，以提高语音分割的精度。
长语音段经过切分后得到短语音段，短语音段的数量大于长语音段的数量。然后进行短语音段聚类：对切分后的短语音段进行聚类，以将所有短语音段聚为多个语音类，并为各个短语音段标注对应的说话人标识，其中，属于同一语音类的短语音段标注相同的说话人标识，不属于同一语音类的短语音段标注不同的说话人标识。聚类方法是：采用K个成分的高斯混合模型拟合每段短语音段，以均值作为特征向量，使用k-means聚类方法把所有短语音段聚为多类。
在一优选的实施例中，如图6所示，在上述实施例的基础上，上述调整模块102包括：
建模单元1021，用于利用所述时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型提取表征说话人身份特征的预设类型向量；
计算单元1022，用于基于所述预设类型向量计算每一语音帧属于对应的说话人的最大后验概率；
第一调整单元1023，用于基于所述最大后验概率并利用预定算法调整该说话人的混合高斯模型；
第二调整单元1024，用于基于调整后的混合高斯模型获取每一语音帧对应的概率最大的说话人，并根据概率最大的说话人与语音帧的概率关系调整所述混合语音中对应的分割边界；
迭代单元1025，用于迭代更新所述声纹模型n次，每次更新所述声纹模 型时迭代m次所述混合高斯模型，以得到各说话人对应的有效语音段，n及m均为大于1的正整数。
本实施例中，利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型提取表征说话人身份特征的预设类型向量，优选地，该预设类型向量为i-vector向量，i-vector向量是反映说话人声学差异的一个重要特征。
在整个混合语音中，根据预设类型向量计算每一语音帧属于某一说话人的最大后验概率，利用计算最大后验概率，在混合语音中通过预设算法重新调整说话人的混合高斯模型，例如，通过Baum-Welch算法重新调整说话人的混合高斯模型，该混合高斯模型为k(一般为3-5个)个高斯模型的集合。利用重新调整后的混合高斯模型寻找每一语音帧概率最大的说话人。根据语音帧与寻找到的该说话人的概率关系调整混合语音的分割边界，例如将分割边界向前微调或者向后微调。最后，迭代更新上述声纹模型n次，每次更新声纹模型时迭代m次混合高斯模型，以得到各个说话人对应的有效语音段，n及m均为大于1的正整数。
本实施例借助深度学习的时间递归神经网络建立声纹模型，用各说话人声纹对应的身份特征对应各语音帧以计算语音帧属于某一说话人的概率，基于该概率修正模型，最终调整语音分割的边界，可以有效提高说话人语音分割的精度，降低错误率，且可扩展性好。
在一优选的实施例中，在上述的实施例的基础上，所述语音分割的装置还包括：反馈模块，用于基于所述有效语音段获取对应的应答内容，并将所述应答内容反馈给所述终端。
本实施例中，自动应答系统关联对应的应答库，该应答库中存储有不同的问题对应的应答内容，自动应答系统在接收到终端发送的混合语音后，分割模块101将其分割为说话人标识对应的有效语音段，从这些有效语音段中获取与该自动应答系统有关问题的一个有效语音段，针对该有效语音段在应答库中进行匹配，反馈模块将匹配得到的应答内容反馈给终端。
如图7所示，图7为本发明实施例方案涉及的硬件运行环境的设备结构示意图。
本发明实施例语音分割设备可以是PC(Personal Computer，个人电脑)，也可以是智能手机、平板电脑、电子书阅读器、便携计算机等终端设备。
参照图7所示，该设备包括：存储器11、处理器12，通信总线13，以及网络接口14。其中，存储器11至少包括一种类型的可读存储介质，所述可读存储介质包括闪存、硬盘、多媒体卡、卡型存储器(例如，SD或DX存储器等)、随机访问存储器(RAM)、静态随机访问存储器(SRAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、可编程只读存储器(PROM)、磁性存储器、磁盘、光盘等。
存储器11在一些实施例中可以是语音分割设备的内部存储单元，例如该语音分割设备的硬盘或内存。存储器11在另一些实施例中也可以是语音分割设备的外部存储设备，例如语音分割设备上配备的插接式硬盘，智能存储卡(Smart Media Card,SMC)，安全数字(Secure Digital,SD)卡，闪存卡(Flash Card)等。进一步地，存储器11还可以既包括语音分割设备的内部存储单元也包括外部存储设备。存储器11不仅可以用于存储安装于语音分割设备的应用软件及各类数据，例如语音分割程序的代码等，还可以用于暂时地存储已经输出或者将要输出的数据。
处理器12在一些实施例中可以是一中央处理器(Central Processing Unit,CPU)、控制器、微控制器、微处理器或其他数据处理芯片，用于运行存储器11中存储的程序代码或处理数据，例如执行语音分割程序等。
通信总线13用于实现这些组件之间的连接通信。
网络接口14可选的可以包括标准的有线接口、无线接口(如WI-FI接口)，通常用于在该设备与其他电子设备之间建立通信连接。
图7仅示出了具有组件11-14以及语音分割程序的语音分割设备，但是应理解的是，并不要求实施所有示出的组件，可以替代的实施更多或者更少的组件。
可选地，该设备还可以包括用户接口，用户接口可以包括显示器(Display)、输入单元比如键盘(Keyboard)，可选的用户接口还可以包括标准的有线接口、无线接口。可选地，在一些实施例中，显示器可以是LED显示器、液晶显示器、触控式液晶显示器以及OLED(Organic Light-Emitting Diode，有机发光二极管)触摸器等。其中，显示器也可以适当的称为显示屏或显示单元，用于显示在语音分割设备中处理的信息以及用于显示可视化的用户界面。
在图7所示的装置实施例中，作为一种计算机存储介质的存储器11中可以包括操作系统、以及语音分割程序；其中，网络接口14用于与终端进行通信连接，以接收终端发送的混合语音，处理器12执行存储器11中存储的语音分割程序时实现如下步骤：
S1，在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
S2，利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
进一步地，步骤S13包括：
对每一长语音段的声学特征进行KL距离分析，对时长大于预设时间阈值的长语音段在KL距离的最大值处进行切分，得到切分后的短语音段。
进一步地，步骤S1包括：
S11，获取所述混合语音中的静音段，去除所述混合语音中的静音段，以根据所述静音段对所述混合语音进行分割，得到分割后的长语音段；
S12，对所述长语音段进行分帧，以提取每一长语音段的声学特征；
S13，对每一长语音段的声学特征进行KL距离分析，根据KL距离分析结果对所述语音段进行切分，得到切分后的短语音段；
S14，利用高斯混合模型对各短语音段进行语音聚类，并对同一语音类的短语音段标注对应的说话人标识。
进一步地，步骤S2之后还包括：
基于所述有效语音段获取对应的应答内容，并将所述应答内容反馈给所述终端。
本发明还提供一种计算机可读存储介质，该计算机可读存储介质上存储有语音分割程序，所述语音分割程序被处理器执行时实现如下步骤：
S1，在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
S2，利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
进一步地，步骤S1包括：
S11，获取所述混合语音中的静音段，去除所述混合语音中的静音段，以根据所述静音段对所述混合语音进行分割，得到分割后的长语音段；
S12，对所述长语音段进行分帧，以提取每一长语音段的声学特征；
S13，对每一长语音段的声学特征进行KL距离分析，根据KL距离分析结果对所述语音段进行切分，得到切分后的短语音段；
S14，利用高斯混合模型对各短语音段进行语音聚类，并对同一语音类的短语音段标注对应的说话人标识。
进一步地，步骤S13包括：
对每一长语音段的声学特征进行KL距离分析，对时长大于预设时间阈值的长语音段在KL距离的最大值处进行切分，得到切分后的短语音段。
进一步地，步骤S2包括：
S21，利用所述时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型提取表征说话人身份特征的预设类型向量；
S22，基于所述预设类型向量计算每一语音帧属于对应的说话人的最大后验概率；
S23，基于所述最大后验概率并利用预定算法调整该说话人的混合高斯模型；
S24，基于调整后的混合高斯模型获取每一语音帧对应的概率最大的说话人，并根据概率最大的说话人与语音帧的概率关系调整所述混合语音中对应的分割边界；
S25，迭代更新所述声纹模型n次，每次更新所述声纹模型时迭代m次所述混合高斯模型，以得到各说话人对应的有效语音段，n及m均为大于1 的正整数。
本发明计算机可读存储介质的具体实施例与上述语音分割的方法和装置各实施例基本相同，在此不作赘述。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本发明的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端设备(可以是手机，计算机，服务器，或者网络设备等)执行本发明各个实施例所述的方法。
以上所述仅为本发明的较佳实施例，并不用以限制本发明，凡在本发明的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。

Claims (20)

  1. 一种语音分割的方法，其特征在于，所述语音分割的方法包括：
    S1，在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
    S2，利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
  2. 根据权利要求1所述的语音分割的方法，其特征在于，所述步骤S1包括：
    S11，获取所述混合语音中的静音段，去除所述混合语音中的静音段，以根据所述静音段对所述混合语音进行分割，得到分割后的长语音段；
    S12，对所述长语音段进行分帧，以提取每一长语音段的声学特征；
    S13，对每一长语音段的声学特征进行KL距离分析，根据KL距离分析结果对所述语音段进行切分，得到切分后的短语音段；
    S14，利用高斯混合模型对各短语音段进行语音聚类，并对同一语音类的短语音段标注对应的说话人标识。
  3. 根据权利要求2所述的语音分割的方法，其特征在于，所述步骤S13包括：
    对每一长语音段的声学特征进行KL距离分析，对时长大于预设时间阈值的长语音段在KL距离的最大值处进行切分，得到切分后的短语音段。
  4. 根据权利要求2所述的语音分割的方法，其特征在于，所述步骤S2包括：
    S21，利用所述时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型提取表征说话人身份特征的预设类型向量；
    S22，基于所述预设类型向量计算每一语音帧属于对应的说话人的最大后验概率；
    S23，基于所述最大后验概率并利用预定算法调整该说话人的混合高斯模型；
    S24，基于调整后的混合高斯模型获取每一语音帧对应的概率最大的说话人，并根据概率最大的说话人与语音帧的概率关系调整所述混合语音中对应的分割边界；
    S25，迭代更新所述声纹模型n次，每次更新所述声纹模型时迭代m次所述混合高斯模型，以得到各说话人对应的有效语音段，n及m均为大于1的正整数。
  5. 根据权利要求3所述的语音分割的方法，其特征在于，所述步骤S2包括：
    S21，利用所述时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型提取表征说话人身份特征的预设类型向量；
    S22，基于所述预设类型向量计算每一语音帧属于对应的说话人的最大后验概率；
    S23，基于所述最大后验概率并利用预定算法调整该说话人的混合高斯模型；
    S24，基于调整后的混合高斯模型获取每一语音帧对应的概率最大的说话人，并根据概率最大的说话人与语音帧的概率关系调整所述混合语音中对应的分割边界；
    S25，迭代更新所述声纹模型n次，每次更新所述声纹模型时迭代m次所述混合高斯模型，以得到各说话人对应的有效语音段，n及m均为大于1的正整数。
  6. 根据权利要求1所述的语音分割的方法，其特征在于，所述步骤S2之后还包括：
    基于所述有效语音段获取对应的应答内容，并将所述应答内容反馈给所述终端。
  7. 一种语音分割的装置，其特征在于，所述语音分割的装置包括：
    分割模块，用于在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
    调整模块，用于利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
  8. 根据权利要求7所述的语音分割的装置，其特征在于，所述分割模块包括：
    去除单元，用于获取所述混合语音中的静音段，去除所述混合语音中的静音段，以根据所述静音段对所述混合语音进行分割，得到分割后的长语音段；
    分帧单元，用于对所述长语音段进行分帧，以提取每一长语音段的声学特征；
    切分单元，用于对每一长语音段的声学特征进行KL距离分析，根据KL距离分析结果对所述语音段进行切分，得到切分后的短语音段；
    聚类单元，用于利用高斯混合模型对各短语音段进行语音聚类，并对同一语音类的短语音段标注对应的说话人标识。
  9. 根据权利要求8所述的语音分割的装置，其特征在于，所述切分单元具体用于对每一长语音段的声学特征进行KL距离分析，对时长大于预设时间阈值的长语音段在KL距离的最大值处进行切分，得到切分后的短语音段。
  10. 根据权利要求8所述的语音分割的装置，其特征在于，所述调整模块包括：
    建模单元，用于利用所述时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型提取表征说话人身份特征的预设类型向量；
    计算单元，用于基于所述预设类型向量计算每一语音帧属于对应的说话人的最大后验概率；
    第一调整单元，用于基于所述最大后验概率并利用预定算法调整该说话人的混合高斯模型；
    第二调整单元，用于基于调整后的混合高斯模型获取每一语音帧对应的概率最大的说话人，并根据概率最大的说话人与语音帧的概率关系调整所述混合语音中对应的分割边界；
    迭代单元，用于迭代更新所述声纹模型n次，每次更新所述声纹模型时迭代m次所述混合高斯模型，以得到各说话人对应的有效语音段，n及m均为大于1的正整数。
  11. 根据权利要求9所述的语音分割的装置，其特征在于，所述调整模块包括：
    建模单元，用于利用所述时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型提取表征说话人身份特征的预设类型向量；
    计算单元，用于基于所述预设类型向量计算每一语音帧属于对应的说话人的最大后验概率；
    第一调整单元，用于基于所述最大后验概率并利用预定算法调整该说话人的混合高斯模型；
    第二调整单元，用于基于调整后的混合高斯模型获取每一语音帧对应的概率最大的说话人，并根据概率最大的说话人与语音帧的概率关系调整所述混合语音中对应的分割边界；
    迭代单元，用于迭代更新所述声纹模型n次，每次更新所述声纹模型时迭代m次所述混合高斯模型，以得到各说话人对应的有效语音段，n及m均为大于1的正整数。
  12. 根据权利要求7所述的语音分割的装置，其特征在于，所述语音分割的装置还包括：反馈模块，用于基于所述有效语音段获取对应的应答内容， 并将所述应答内容反馈给所述终端。
  13. 一种语音分割设备，其特征在于，所述语音分割设备包括存储器、处理器以及存储在所述存储器上并可在所述处理器上运行的语音分割程序，所述语音分割程序被所述所述处理器执行时实现如下步骤：
    S1，在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
    S2，利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
  14. 根据权利要求13所述的语音分割设备，其特征在于，所述步骤S13包括：
    对每一长语音段的声学特征进行KL距离分析，对时长大于预设时间阈值的长语音段在KL距离的最大值处进行切分，得到切分后的短语音段。
  15. 根据权利要求14所述的语音分割设备，其特征在于，所述步骤S1包括：
    S11，获取所述混合语音中的静音段，去除所述混合语音中的静音段，以根据所述静音段对所述混合语音进行分割，得到分割后的长语音段；
    S12，对所述长语音段进行分帧，以提取每一长语音段的声学特征；
    S13，对每一长语音段的声学特征进行KL距离分析，根据KL距离分析结果对所述语音段进行切分，得到切分后的短语音段；
    S14，利用高斯混合模型对各短语音段进行语音聚类，并对同一语音类的短语音段标注对应的说话人标识。
  16. 根据权利要求14所述的语音分割设备，其特征在于，所述步骤S2之后还包括：
    基于所述有效语音段获取对应的应答内容，并将所述应答内容反馈给所述终端。
  17. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质上存储有语音分割程序，所述语音分割程序被处理器执行时实现如下步骤：
    S1，在接收到终端发送的混合语音时，将所述混合语音分割成多个短语音段，并对各短语音段标注对应的说话人标识；
    S2，利用时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型调整所述混合语音中对应的分割边界，以分割出各说话人标识对应的有效语音段。
  18. 根据权利要求17所述的计算机可读存储介质，其特征在于，所述步骤S1包括：
    S11，获取所述混合语音中的静音段，去除所述混合语音中的静音段，以根据所述静音段对所述混合语音进行分割，得到分割后的长语音段；
    S12，对所述长语音段进行分帧，以提取每一长语音段的声学特征；
    S13，对每一长语音段的声学特征进行KL距离分析，根据KL距离分析结果对所述语音段进行切分，得到切分后的短语音段；
    S14，利用高斯混合模型对各短语音段进行语音聚类，并对同一语音类的短语音段标注对应的说话人标识。
  19. 根据权利要求18所述的计算机可读存储介质，其特征在于，所述步骤S13包括：
    对每一长语音段的声学特征进行KL距离分析，对时长大于预设时间阈值的长语音段在KL距离的最大值处进行切分，得到切分后的短语音段。
  20. 根据权利要求18所述的计算机可读存储介质，其特征在于，所述步骤S2包括：
    S21，利用所述时间递归神经网络对各说话人标识对应的短语音段建立声纹模型，基于所述声纹模型提取表征说话人身份特征的预设类型向量；
    S22，基于所述预设类型向量计算每一语音帧属于对应的说话人的最大后验概率；
    S23，基于所述最大后验概率并利用预定算法调整该说话人的混合高斯模型；
    S24，基于调整后的混合高斯模型获取每一语音帧对应的概率最大的说话人，并根据概率最大的说话人与语音帧的概率关系调整所述混合语音中对应的分割边界；
    S25，迭代更新所述声纹模型n次，每次更新所述声纹模型时迭代m次所述混合高斯模型，以得到各说话人对应的有效语音段，n及m均为大于1的正整数。
PCT/CN2017/091310 2016-12-19 2017-06-30 语音分割的方法、装置、设备及计算机存储介质 Ceased WO2018113243A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN2016111767919 2016-12-19
CN201611176791.9A CN106782507B (zh) 2016-12-19 2016-12-19 语音分割的方法及装置

Publications (1)

Publication Number Publication Date
WO2018113243A1 true WO2018113243A1 (zh) 2018-06-28

Family

ID=58889790

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/091310 Ceased WO2018113243A1 (zh) 2016-12-19 2017-06-30 语音分割的方法、装置、设备及计算机存储介质

Country Status (3)

Country Link
CN (1) CN106782507B (zh)
TW (1) TWI643184B (zh)
WO (1) WO2018113243A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111213205A (zh) * 2019-12-30 2020-05-29 深圳市优必选科技股份有限公司 一种流式语音转换方法、装置、计算机设备及存储介质
CN112201256A (zh) * 2020-10-09 2021-01-08 深圳前海微众银行股份有限公司 声纹分割方法、装置、设备及可读存储介质
US20220301554A1 (en) * 2019-01-28 2022-09-22 Pindrop Security, Inc. Unsupervised keyword spotting and word discovery for fraud analytics
CN115938386A (zh) * 2022-12-28 2023-04-07 思必驰科技股份有限公司 基于多说话人语音检测的语音分离方法、系统和电子设备
CN120766051A (zh) * 2025-09-10 2025-10-10 吉林建筑大学 冰雪装备状态检测方法及系统

Families Citing this family (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106782507B (zh) * 2016-12-19 2018-03-06 平安科技（深圳）有限公司 语音分割的方法及装置
CN107358945A (zh) * 2017-07-26 2017-11-17 谢兵 一种基于机器学习的多人对话音频识别方法及系统
CN108257592A (zh) * 2018-01-11 2018-07-06 广州势必可赢网络科技有限公司 一种基于长短期记忆模型的人声分割方法及系统
CN108335226A (zh) * 2018-02-08 2018-07-27 江苏省农业科学院 农业种质资源信息实时智能采集系统
CN108597521A (zh) * 2018-05-04 2018-09-28 徐涌 音频角色分割与识别文字的交互系统、方法、终端及介质
CN109300470B (zh) * 2018-09-17 2023-05-02 平安科技（深圳）有限公司 混音分离方法和混音分离装置
CN109461447B (zh) * 2018-09-30 2023-08-18 厦门快商通信息技术有限公司 一种基于深度学习的端到端说话人分割方法及系统
CN109346083A (zh) * 2018-11-28 2019-02-15 北京猎户星空科技有限公司 一种智能语音交互方法及装置、相关设备及存储介质
CN109743624B (zh) * 2018-12-14 2021-08-17 深圳壹账通智能科技有限公司 视频切割方法、装置、计算机设备和存储介质
CN109616097B (zh) * 2019-01-04 2024-05-10 平安科技（深圳）有限公司 语音数据处理方法、装置、设备及存储介质
US11031017B2 (en) 2019-01-08 2021-06-08 Google Llc Fully supervised speaker diarization
CN110211595B (zh) * 2019-06-28 2021-08-06 四川长虹电器股份有限公司 一种基于深度学习的说话人聚类系统
CN110675858A (zh) * 2019-08-29 2020-01-10 平安科技（深圳）有限公司 基于情绪识别的终端控制方法和装置
CN110910891B (zh) * 2019-11-15 2022-02-22 复旦大学 基于长短时记忆深度神经网络的说话人分段标注方法
CN110930984A (zh) * 2019-12-04 2020-03-27 北京搜狗科技发展有限公司 一种语音处理方法、装置和电子设备
CN111524527B (zh) * 2020-04-30 2023-08-22 合肥讯飞数码科技有限公司 话者分离方法、装置、电子设备和存储介质
CN111681644B (zh) * 2020-06-30 2023-09-12 浙江同花顺智能科技有限公司 一种说话人分割方法、装置、设备和存储介质
CN112397057B (zh) * 2020-12-01 2024-07-02 平安科技（深圳）有限公司 基于生成对抗网络的语音处理方法、装置、设备及介质
CN112562682A (zh) * 2020-12-02 2021-03-26 携程计算机技术（上海）有限公司 基于多人通话的身份识别方法、系统、设备及存储介质
CN113707130B (zh) * 2021-08-16 2024-06-14 北京搜狗科技发展有限公司 一种语音识别方法、装置和用于语音识别的装置
CN113793592B (zh) * 2021-10-29 2024-07-16 浙江核新同花顺网络信息股份有限公司 一种区分说话人的方法和系统
US12223945B2 (en) 2021-10-13 2025-02-11 Hithink Royalflush Information Network Co., Ltd. Systems and methods for multiple speaker speech recognition
CN114299962A (zh) * 2021-12-22 2022-04-08 携程旅游信息技术（上海）有限公司 基于音频流的对话角色分离方法、系统、设备及存储介质
CN114999453B (zh) * 2022-05-25 2023-05-30 中南大学湘雅二医院 一种基于语音识别的术前访视系统及相应语音识别方法
CN116548717B (zh) * 2023-05-16 2026-01-02 江西恒必达实业有限公司 一种充电式吹风机及其控制方法

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1716380A (zh) * 2005-07-26 2006-01-04 浙江大学 基于决策树和说话人改变检测的音频分割方法
US7930179B1 (en) * 2002-08-29 2011-04-19 At&T Intellectual Property Ii, L.P. Unsupervised speaker segmentation of multi-speaker speech data
CN102543063A (zh) * 2011-12-07 2012-07-04 华南理工大学 基于说话人分割与聚类的多说话人语速估计方法
CN102760434A (zh) * 2012-07-09 2012-10-31 华为终端有限公司 一种声纹特征模型更新方法及终端
CN105161093A (zh) * 2015-10-14 2015-12-16 科大讯飞股份有限公司 一种判断说话人数目的方法及系统
CN105913849A (zh) * 2015-11-27 2016-08-31 中国人民解放军总参谋部陆航研究所 一种基于事件检测的说话人分割方法
CN106782507A (zh) * 2016-12-19 2017-05-31 平安科技（深圳）有限公司 语音分割的方法及装置

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6304842B1 (en) * 1999-06-30 2001-10-16 Glenayre Electronics, Inc. Location and coding of unvoiced plosives in linear predictive coding of speech
US8595007B2 (en) * 2006-06-15 2013-11-26 NITV Federal Services, LLC Voice print recognition software system for voice identification and matching
TW201513095A (zh) * 2013-09-23 2015-04-01 Hon Hai Prec Ind Co Ltd 語音處理系統、裝置及方法
CN106228045A (zh) * 2016-07-06 2016-12-14 吴本刚 一种身份识别系统

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7930179B1 (en) * 2002-08-29 2011-04-19 At&T Intellectual Property Ii, L.P. Unsupervised speaker segmentation of multi-speaker speech data
CN1716380A (zh) * 2005-07-26 2006-01-04 浙江大学 基于决策树和说话人改变检测的音频分割方法
CN102543063A (zh) * 2011-12-07 2012-07-04 华南理工大学 基于说话人分割与聚类的多说话人语速估计方法
CN102760434A (zh) * 2012-07-09 2012-10-31 华为终端有限公司 一种声纹特征模型更新方法及终端
CN105161093A (zh) * 2015-10-14 2015-12-16 科大讯飞股份有限公司 一种判断说话人数目的方法及系统
CN105913849A (zh) * 2015-11-27 2016-08-31 中国人民解放军总参谋部陆航研究所 一种基于事件检测的说话人分割方法
CN106782507A (zh) * 2016-12-19 2017-05-31 平安科技（深圳）有限公司 语音分割的方法及装置

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220301554A1 (en) * 2019-01-28 2022-09-22 Pindrop Security, Inc. Unsupervised keyword spotting and word discovery for fraud analytics
US11810559B2 (en) * 2019-01-28 2023-11-07 Pindrop Security, Inc. Unsupervised keyword spotting and word discovery for fraud analytics
CN111213205A (zh) * 2019-12-30 2020-05-29 深圳市优必选科技股份有限公司 一种流式语音转换方法、装置、计算机设备及存储介质
CN111213205B (zh) * 2019-12-30 2023-09-08 深圳市优必选科技股份有限公司 一种流式语音转换方法、装置、计算机设备及存储介质
CN112201256A (zh) * 2020-10-09 2021-01-08 深圳前海微众银行股份有限公司 声纹分割方法、装置、设备及可读存储介质
CN112201256B (zh) * 2020-10-09 2023-09-19 深圳前海微众银行股份有限公司 声纹分割方法、装置、设备及可读存储介质
CN115938386A (zh) * 2022-12-28 2023-04-07 思必驰科技股份有限公司 基于多说话人语音检测的语音分离方法、系统和电子设备
CN120766051A (zh) * 2025-09-10 2025-10-10 吉林建筑大学 冰雪装备状态检测方法及系统

Also Published As

Publication number Publication date
CN106782507B (zh) 2018-03-06
CN106782507A (zh) 2017-05-31
TWI643184B (zh) 2018-12-01
TW201824250A (zh) 2018-07-01

Similar Documents

Publication Publication Date Title
WO2018113243A1 (zh) 语音分割的方法、装置、设备及计算机存储介质
US11620104B2 (en) User interface customization based on speaker characteristics
US20230206909A1 (en) Speech recognition using neural networks
US9589560B1 (en) Estimating false rejection rate in a detection system
US9311915B2 (en) Context-based speech recognition
US10332507B2 (en) Method and device for waking up via speech based on artificial intelligence
US8731936B2 (en) Energy-efficient unobtrusive identification of a speaker
US8831947B2 (en) Method and apparatus for large vocabulary continuous speech recognition using a hybrid phoneme-word lattice
US9299347B1 (en) Speech recognition using associative mapping
US20170194006A1 (en) Individualized hotword detection models
CN109360572B (zh) 通话分离方法、装置、计算机设备及存储介质
US10089978B2 (en) Detecting customers with low speech recognition accuracy by investigating consistency of conversation in call-center
CN107209842A (zh) 隐私保护训练语料库选择
CN112201275B (zh) 声纹分割方法、装置、设备及可读存储介质
CN107680584B (zh) 用于切分音频的方法和装置
US10522135B2 (en) System and method for segmenting audio files for transcription
US9378735B1 (en) Estimating speaker-specific affine transforms for neural network based speech recognition systems
CN114067793A (zh) 音频处理方法和装置、电子设备及可读存储介质
CN109688271A (zh) 联系人信息输入的方法、装置及终端设备
JP2017045054A (ja) 言語モデル改良装置及び方法、音声認識装置及び方法
CN114596877B (zh) 一种话者分离方法、装置、电子设备及存储介质
US10872615B1 (en) ASR-enhanced speech compression/archiving
CN110415708A (zh) 基于神经网络的说话人确认方法、装置、设备及存储介质
US10783882B2 (en) Acoustic change detection for robust automatic speech recognition based on a variance between distance dependent GMM models
US11398239B1 (en) ASR-enhanced speech compression

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17883133

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 05/09/2019)

122 Ep: pct application non-entry in european phase

Ref document number: 17883133

Country of ref document: EP

Kind code of ref document: A1