58 results about "Speech segmentation" patented technology

The invention comprises a bimodal man-man conversation sentiment analysis system and a bimodal man-man conversation sentiment analysis method based on machine learning. The bimodal man-man conversation sentiment analysis system is characterized by comprising a speech recognition module, a text deep-layer feature extraction module, a speech segmentation module, an acoustic feature extraction module, a feature fusion module and an sentiment analysis module, wherein the speech recognition module is used for recognizing speech content and a time label; the text deep-layer feature extraction module is used for completing the extraction of text deep-layer word level features and text deep-layer sentence level features; the speech segmentation module is used for segmenting single sentence speech from entire speech; the acoustic feature extraction module is used for completing the extraction of acoustic features of the speech; the feature fusion module is used for fusing the obtained text deep-layer features with the acoustic features; and the sentiment analysis module is used for acquiring sentiment polarities of the speech to be subjected to sentiment analysis. The bimodal man-man conversation sentiment analysis method can integrate the text and audio modals for recognizing conversation sentiment, and fully utilizes features of word vectors and sentence vectors, thereby improving the precision of recognition.

A speech segment determination device includes a frame division portion, a power spectrum calculation portion, a power spectrum operation portion, a spectral entropy calculation portion and a determination portion. The frame division portion divides an input signal in units of frames. The power spectrum calculation portion calculates, using an analysis length, a power spectrum of the input signal for each of the frames that have been divided. The power spectrum operation portion adds a value of the calculated power spectrum to a value of power spectrum in each of frequency bins. The spectral entropy calculation portion calculates spectral entropy using the power spectrum whose value has been increased. The determination portion determines, based on a value of the spectral entropy, whether the input signal is a signal in a speech segment.

The invention relates to a speech segmentation method and a speech segmentation device. The speech segmentation method comprises the following steps: segmenting mixed speech into a plurality of short speech segments when the mixed speech, which is transmitted from a terminal, is received by virtue of an automatic response system, and labeling the various short speech segments with corresponding speaker identifiers; and establishing a vocal print model for the short speech segments corresponding to the various speaker identifiers by virtue of a time recurrent neural network, and regulating corresponding segmentation boundaries in the mixed speech on the basis of the vocal print model, so as to segment out effective speech segments corresponding to the various speaker identifiers. With the application of the method and the device provided by the invention, precision of speech segmentation can be effectively enhanced; and especially for speeches that conversations are frequently alternated and are overlapped, a relatively good effect of speech segmentation is achieved.

The invention discloses an emotional Chinese text human voice synthesis method, which mainly comprises the steps of: (1) constructing an emotional corpus; (2) and performing emotional speech synthesisbased on waveform splicing. The emotional corpus establishment is mainly implemented by the steps of: (11) segmenting terms and acquiring parts of speech of the terms; (12) performing speech segmentation, and acquiring audio data corresponding to segmented terms based on speech data features and text corpora; (13) and performing emotion analysis, and acquiring emotional feature values of terms, clauses and whole sentences based on text term segmentation and audio features. The emotional speech synthesis based on waveform splicing is implemented by the steps of: (21) segmenting terms and performing emotion analysis on a text to be synthesized, and acquiring parts of speech of words, sentence patterns and emotional features in the text to be synthesized; (22) selecting the optimal corpus, and carrying out matching to obtain the optimal corpus set based on text eigenvalues; (23) and perfomring speech synthesis and waveform splicing, extracting a word audio sequence set from the corpus set, and synthesizing the audio to output a final speech. The emotional Chinese text human voice synthesis method is used for synthesizing and outputting a true human voice speech with emotional features.

A method and system for obtaining a pool of speech syllable models. The model pool is generated by first detecting a training segment using unsupervised speech segmentation or speech unit spotting. If the model pool is empty, a first speech syllable model is trained and added to the model pool. If the model pool is not empty, an existing model is determined from the model pool that best matches the training segment. Then the existing module is scored for the training segment. If the score is less than a predefined threshold, a new model for the training segment is created and added to the pool. If the score equals the threshold or is larger than the threshold, the training segment is used to improve or to re-estimate the model.

A method segments an audio signal including frames into non-speech and speech segments. First, high-dimensional spectral features are extracted from the audio signal. The high-dimensional features are then projected non-linearly to low-dimensional features that are subsequently averaged using a sliding window and weighted averages. A linear discriminant is applied to the averaged low-dimensional features to determine a threshold separating the low-dimensional features. The linear discriminant can be determined from a Gaussian mixture or a polynomial applied to a bi-model histogram distribution of the low-dimensional features. Then, the threshold can be used to classify the frames into either non-speech or speech segments. Speech segments having a very short duration can be discarded, and the longer speech segments can be further extended. In batch-mode or real-time the threshold can be updated continuously.

The embodiment of the invention discloses a single channel-based non-supervision target speaker speech extraction method comprising a teacher language detection step and a teacher language model training step; the teacher language detection step comprises the following parts: obtaining speech data from a classroom recording; processing speech signals; speech segmentation and modeling, the speech segmentation comprises steps of segmenting the classroom speech at equal length, aiming at each segment of speech and extracting corresponding MFCC features, and building each segment speech GMM modelaccording to the MFCC features; teacher speed detection, calculating the similarity between the GMM model of each segment speech except for teacher speech types and a GGMM, tagging the GMM models smaller than a set threshold as teacher speech types, thus obtaining the final teacher speech types; the teacher language GGMM model training step comprises the following parts: clustering the speech dataobtained in S3; obtaining an initial teacher speech type, and extracting the GGMM model according to the initial teacher speech type. The method can effectively improve the system adaptability and intelligence in real applications, thus laying foundation for following applications and researches.

The invention discloses a phonetic segmentation-based isolate word recognition method. In the method, a continuous HMM model obtained by voice training takes syllable or semi-syllable as a unit set, trained unit set models are spliced into whole word models according to the syllables or semi-syllables of isolate words in a word list, and a Viterbi algorithm is adopted in recognition. The method has the advantages of improving recognition performance because each segment of HMM model in recognition results (accurately recognized isolate words) can be better matched with each segment of characteristic parameters of voices to be recognized, and shortening recognition response time because the recognition of the isolate words with relatively smaller probability values can be directly refused every time recognition operation is performed on the tail state of one syllable or semi-syllable.

The invention provides a speech segmentation method, a device, a device and a computer storage medium, wherein the method comprises the following steps: determining the cross correlation degree of a first speech and a second speech, wherein the second speech is the speech obtained after recording the first speech, and the first speech is formed by splicing two or more first speech segments; Calibrating a time tag based on the cross-correlation, the time tag comprising a start time and an end time of each first speech segment in a first speech; Using the calibrated time stamp, the second speechis segmented to obtain more than two second speech segments. The invention can better align the calibrated time label with the second speech, thereby improving the segmentation accuracy of the secondspeech.

In a method for performing a segmentation operation upon a synthesizing speech signal and an input speech signal, a synthesized speech signal and a speech element duration signal are generated from the synthesizing speech signal A first feature parameter is extracted from the synthesized speech signal, and a second feature parameter is extracted from the input speech signal. A dynamic programming matching operation is performed upon the second feature parameter with reference to the first feature parameter and the speech element duration signal to obtain segmentation points of the input speech signal.

The invention provides a speech recognition method and device of a joint model, and computer equipment. Firstly, to-be-recognized speech is input into a first speech recognition model and a second speech recognition model; a first recognition sequence corresponding to the first speech recognition model and a second recognition sequence corresponding to the second speech recognition model are obtained, wherein the first speech recognition model is a speech recognition model based on an HMM, and the second speech recognition model is an end-to-end speech recognition model. A system corrects thefirst recognition sequence according to the second recognition sequence to obtain a final recognition result. According to the invention, the recognition sequences of two different types of speech recognition models are combined with each other, and the first recognition sequence is corrected through the second recognition sequence, so that the speech segmentation accuracy is effectively improved.

The embodiment of the disclosure discloses a voice segmentation model training method and device, electronic equipment and a computer readable storage medium. The training method of the voice segmentation model comprises the following steps: acquiring a voice feature map of a sample voice file; acquiring annotation information of a target voice in the voice feature map; initializing model parameters of a voice segmentation model; inputting the voice feature map into the voice segmentation model to obtain prediction information of a target voice output by the voice segmentation model; calculating an error between the prediction information and the annotation information according to a target function; updating parameters of the voice segmentation model according to the error; and inputtingthe voice feature map into the voice segmentation model after parameter updating to iterate the parameter updating process until the error is less than a first threshold. According to the method, thespeech segmentation model is trained through the speech feature image, and a technical problem of inaccurate speech segmentation caused by complex speech signals in the prior art is solved.

The invention discloses a speech segmentation, recombination and output method and device.The speech segmentation, recombination and output method and device are used for reducing occupied IO port resources and reducing production and manufacturing cost while providing a speech content alteration function.The speech segmentation, recombination and output device comprises a flash memory module and a single-chip microcomputer module.The flash memory module is used for storing several segments of speech broadcasting information and an address directory corresponding to the segments of speech broadcasting information.The single-chip microcomputer module comprises a speech segmentation unit, a segment processing unit and a reading unit, wherein the speech segmentation unit is used for decomposing target speech information into several ordered segments of speech information, the segment processing unit is used for converting the segments of speech information decomposed by the speech segmentation unit into address directory numbers in order, and the reading unit is used for reading and outputting the corresponding speech broadcasting information in the flash memory module in order according to the converted address directory numbers, so that recombination and output of segmented speech are realized.

The invention provides a sound acquisition method and device based on a microphone array. The method comprises the following steps of acquiring voices in an environment and sounder information in theenvironment by utilizing the microphone array to obtain multi-channel voices, wherein the sounder information comprises sounding directions of sounders and the number of the sounders; converting the multi-channel voices into single-channel voices; carrying out sentence segmentation on the single-channel voices to obtain voice segment data streams containing a preset type of sound; matching the voice segment data streams with each sounder to obtain a single-separated voice of each sounder; and respectively synthesizing the single-separated voice matched with each sounder into a respective voicesegment. The technical scheme provided by the invention is high in applicability, can be suitable for near-field and far-field voice environments, has the characteristics of high sound detection accuracy, and can separate the voices of a plurality of sounders, so that each sounder corresponds to one voice separated voice segment.

The embodiment of the invention provides a voice information processing method and electronic equipment. The method comprises the following steps: dividing voice information into a plurality of voicesegments according to preset segment parameters; generating segmented icons; under the condition that an editing instruction for the voice information is received, displaying the sorted segmented icons; receiving a first input of a user to a target segmented icon in the segmented icons; and in response to the first input, processing the voice information by taking the voice segment corresponding to the target segment icon as a reference. According to the voice information processing method provided by the embodiment of the invention, a user can flexibly trigger the electronic equipment to insert the newly added voice content into the voice information, delete the voice segments in the voice information or continuously play the voice information, the operation is convenient, and the use experience of the user can be improved.

The invention discloses a rail transit standard entity relationship automatic completion method based on artificial intelligence. The method comprises the steps of constructing an entity relationshipcompletion model, inputting the rail transit specification and part-of-speech segmentation of nouns into an entity relationship completion model; judging whether the input specification is a simple sentence or not; if yes, searching entity related attributes in the rail transit specification; generating an entity relationship triple; if not, extracting later sentence attribute words and entities of the rail transit specification, matching the former sentence entities and later sentence attribute words in a n:n manner, or judging whether the the former sentence grammar is subject-verb-object ornot and the latter sentence grammar is object complement; if yes, directly matching the former sentence entities with the objects and directly matching the later sentence keywords with the object entities to generate entity relationship triples, and if not, outputting the entities of which the vocabulary relevancy exceeds a threshold value and the entity relationships to generate the entity relationship triples to obtain complete semantic structure entity specifications, thereby finishing automatic completion of the rail transit specification entity relationships.

The invention discloses a synthetic speech detection method based on speech segmentation, belongs to the field of speech detection, and aims to solve the problem of low detection precision in the prior art. The method comprises the following steps of: extracting two kinds of features in an audio, namely a CQCC feature of a voiced segment of the audio and an average zero-crossing rate feature of a silent (mute) segment of the audio; and adopting two GMM models to fit the two kinds of features respectively, giving different weights to the two GMM models, carrying out testing, and finding the most appropriate weight. The detection precision of synthetic speech is obviously improved.

The embodiment of the invention provides an audio data processing method, equipment and system, which are used for classifying conference audio data according to identities of spokesmen. The conference record processing method specifically comprises the steps that the conference record processing device obtains audio data of a first conference place, sound source orientation information corresponding to the audio data and an identity recognition result, and additional domain information comprises the sound source orientation information corresponding to the audio data; the identity recognition result is used for indicating a corresponding relation between spokesman identity information obtained through a portrait recognition method and speaking time information of a spokesman; then the conference record processing device performs voice segmentation on the audio data to obtain first segmented audio data of the audio data; and finally, the conference record processing device determines a spokesman corresponding to the first segment of audio data according to the voiceprint feature of the first segment of audio data and the identity recognition result.

A speech processing device includes: first segment means for dividing first speech into a plurality of first speech segments; second segment means for dividing second speech into a plurality of second speech segments; primary speaker recognition means for calculating scores indicating similarities between the plurality of first and second speech segments; threshold value calculation means for calculating a threshold value based on scores indicating similarities between the plurality of first speech segments; speaker clustering means for classifying each of the plurality of second speech segments into one or more clusters having a similarity higher than the similarity indicated by the threshold value; and secondary speaker recognition means for calculating a similarity between each of the one or more clusters and the first speech and determining based on a result of the calculation whether speech corresponding to the first speech is contained in any of the one or more clusters.

The invention belongs to the technical field of voice signal processing, and in particular relates to a keyword search method based on unified representation. The method comprises the following steps:training a neural network voice auto-encoder with a bottleneck layer by adopting abundant voice data, thus obtaining an acoustics representation vector extractor; training a neural network text auto-encoder with a bottleneck layer by adopting abundant text data, thus obtaining a language representation vector extractor; extracting corresponding acoustics representation vectors and language representation vectors respectively by adopting abundant voice data segments and corresponding text data segments for training a unified vector extractor; acquiring inquiring vectors of a text keyword through the language representation vector extractor and the unified vector extractor; acquiring inquiring vectors of a voice keyword through the acoustics representation vector extractor and the unified vector extractor; and for to-be-inquired voice, acquiring a plurality of index vectors in segments in the sequence through the acoustics representation vector extractor and the unified vector extractor, and calculating the distances among the inquiring vectors, wherein if the value is smaller than a preset threshold, considering that the keyword is hit.