51 results about "Speech Acoustics" patented technology

A common narrow-band speech signal is expanded into a wide-band speech signal. The expanded signal gives the impression of a wide-band speech signal regardless of what type of vocoder is used in a receiver. The robust techniques suggested herein are based on speech acoustics and fundamentals of human hearing. That is the techniques extend the harmonic structure of the speech signal during voiced speech segments and introduce a linearly estimated amount of speech energy in the wide frequency-band. During unvoiced speech segments, a fricated noise may be introduced in the upper frequency-band.

The invention discloses a method for scoring interpretation quality by using a computer, which comprehensively uses computer voice identification, voice pronunciation estimation, text translation quality confirmation technology to obtain the interpretation quality of a testee. The method comprises the following steps: establishing a database aiming at the characteristics of the tested pronunciation crowd; training the database by using a large vocabulary continuous speech acoustic model training platform to obtain an acoustic model; collecting corresponding expert knowledge and translated text linguistic data for each translation question to form a language model, a scoring model and a standard adjustment model required for identification; and finally, integrating the output result of a speech identifier and a linguistic processing mechanism to output the score of the interpretation quality of the testee, and providing feedback suggestion. The effect of machine evaluation already basically reaches the level close to expert evaluation; and meanwhile, the method also can provide some suggestions for testee pronunciation, vocabulary use and sentence pattern use during evaluating to guide the testee to correct.

The invention belongs to the technical field of signal processing in the electronic industry, and aims at solving a problem that the discrimination performance of an acoustic model of a target language with just a little mark data is low. In order to solve the above problem, the invention provides a small data speech acoustic modeling method in speech recognition, and the method comprises the steps: carrying out the adversarial training of the acoustic features of many languages through a language adversarial discriminator, so as to build a multi-language adversarial bottleneck network model;taking the acoustic features of a target language as the input of the multi-language adversarial bottleneck network model, so as to extract the bottleneck features which is irrelevant to the language;carrying out the fusion of the bottleneck features which is irrelevant to the language with the acoustic features of the target language, so as to obtain fusion features; carrying out the training through the fusion features, so as to build an acoustic model of the target language. The method effectively irons out the defects, caused by a condition that the bottleneck information comprises the information correlated with the language, of the unremarkable improvement of the recognition performance of the target language and even the negative migration phenomenon in the prior art, thereby improving the voice recognition precision of the target language.

To achieve an improvement in recognition performance, a non-speech acoustic model correction unit adapts a non-speech acoustic model representing a non-speech state using input data observed during an interval immediately before a speech recognition interval during which speech recognition is performed, by means of one of the most likelihood method, the complex statistic method, and the minimum distance-maximum separation theorem.

The present disclosure provides a far field speech acoustic model training method and system. The method comprises: blending near field speech training data with far field speech training data to generate blended speech training data, wherein the far field speech training data is obtained by performing data augmentation processing for the near field speech training data; using the blended speech training data to train a deep neural network to generate a far field recognition acoustic model. The present disclosure can avoid the problem of spending a lot of time costs and economic costs in recording the far field speech data in the prior art; and reduce time and economic costs of obtaining the far field speech data, and improve the far field speech recognition effect.

The invention discloses a voice conversion method with emotion and rhythm, which comprises a training stage and a conversion stage, and is characterized in that a style coding layer with an attentionmechanism is used for calculating a style coding vector of a speaker; and the style coding vector and the speaker voice acoustic features are input into a self-coding network with a box link togetherfor training and conversion, and finally the acoustic features are converted into audio through a vocoder. On the basis of a traditional voice conversion method, rhythm and emotion information of a speaker is introduced, so that the converted voice has emotion and rhythm of the voice of a target speaker, and the method has high similarity and high voice quality in many-to-many speaker voice conversion tasks such as intra-set pairing, intra-set pairing, extra-set pairing, extra-set pairing and the like.

The invention discloses a multi-speaker and multi-language speech synthesis method. The method comprises the following steps: extracting speech acoustic features; processing the texts in different languages into a unified representation mode, and aligning the audio with the texts to obtain duration information; constructing a speaker space and a language space, generating a speaker id and a language id, extracting a speaker vector and a language vector, adding the speaker vector and the language vector into the initial speech synthesis model, and training the initial speech synthesis model byusing the aligned text, duration information and speech acoustic features to obtain a speech synthesis model; processing the to-be-synthesized text to generate the speaker id and the language id; andinputting the speaker id, the text and the language id into a speech synthesis model, outputting speech acoustic features and converting the speech acoustic features into audio. A system is also disclosed. According to the method, unentanglement of the characteristics of the speaker and the language characteristics is realized, and conversion of the speaker or the language can be realized only bychanging the id.

The present invented method comprises speech sampling and preprocessing, speech acoustic feature extraction, syllable fuzzy division, speech template training, speech pattern matching and recognitiondecision. Said method adopts uniquie quick method for processing signal to support Chinese phrase recognition, and processes the advantages of less resource consumption, quick operation speed and high recognition rate. Said method is applicable to the products of toy, personal digital processing device and communication terminal, etc.

The invention discloses a personalized speech translation method based on speaker features, wherein the personalized speech translation method comprises the following steps: collecting speaker speech,extracting speech acoustic features of the speaker speech, and converting the speech acoustic features into speaker feature vectors; carrying out speaker text recognition by combining the speaker feature vectors with speaker voice acoustic features; translating a text of a speaker into a text of a target language; combining a text code of the target language generated in the previous step with the speaker feature vectors generated in the first step to obtain a target text vector with speaker features; and generating a target voice from the target text vector generated in the previous step through a text-to-voice model. By adding the speaker feature extraction network, different speaker mood tones can be added into the speech recognition and text-to-speech conversion process, and the meaning of the speaker can be more accurately translated. The invention further discloses a personalized speech translation device based on the speaker features.

The invention relates to the technical field of voice synthesis, voice recognition and voice cloning, and provides a voice cloning implementation scheme based on Bottleneck features (language featuresof audio) by combining a voice synthesis technology, a voice recognition technology and a transfer learning technology. A training system and a training method are included. The TTS service with highnaturalness and similarity is provided by using a small number of samples, so that the TTS service with target user characteristics is provided, and problems of large service sample size, long manufacturing period and high labor cost of a voice synthesis technology are solved. The training system comprises a data acquisition module, an acoustic feature extraction module, a voice recognition module, a rhythm module, a multi-person voice acoustic module and a voice synthesis module. The invention further provides a training method based on the system. The training method comprises the steps oftraining corpus preparation, acoustic feature extraction, training and fine adjustment of all modules and speech synthesis.

The invention relates to the technical field of speech recognition, in particular to a speech recognition method and device, electronic equipment and a storage medium, which can be applied to various scenes such as cloud technology, artificial intelligence, intelligent traffic and auxiliary driving and are used for efficiently and accurately realizing speech recognition of multiple dialect target languages. The method comprises the following steps: acquiring to-be-recognized voice data of a target language; extracting voice acoustic features corresponding to each frame of voice data in the to-be-recognized voice data; performing deep feature extraction on the voice acoustic features to obtain corresponding dialect embedding features; encoding the voice acoustic features to obtain corresponding acoustic encoding features; and based on the dialect embedding feature and the acoustic coding feature, performing dialect speech recognition on the to-be-recognized speech data to obtain target text information and a target dialect category corresponding to the to-be-recognized speech data. According to the method, the dialect embedding feature and the acoustic coding feature are combined for comprehensive learning, so that speech recognition for recognizing various dialects can be efficiently and accurately realized.

A method is developed which includes 1) defining a switching state space model for a continuous valued hidden production-related parameter and the observed speech acoustics, and 2) approximating a posterior probability that provides the likelihood of a sequence of the hidden production-related parameters and a sequence of speech units based on a sequence of observed input values. In approximating the posterior probability, the boundaries of the speech units are not fixed but are optimally determined. Under one embodiment, a mixture of Gaussian approximation is used. In another embodiment, an HMM posterior approximation is used.

The embodiment of the invention discloses a multi-dialect speech recognition method, device and apparatus, and a storage medium. The method comprises the steps: obtaining an input speech acoustic feature through a multi-dialect encoder, and outputting a first vector code with a fixed length; identifying the first vector code through a dialect identifier to obtain a corresponding dialect vector; and decoding the first vector code through a multi-dialect decoder according to the dialect vector to obtain a recognition text corresponding to the speech acoustic feature. According to the scheme, dialect recognition efficiency is improved, a large amount of sample data is not needed, and recognition accuracy is better than that of an existing scheme.

The invention relates to the technical field of voice recognition, and particularly relates to an intelligent recognition method for incomplete voice of elderly people. The intelligent recognition method for incomplete voice of elderly people comprises the following steps: S1, preprocessing voice data, acquiring an original voice signal, and preprocessing the original voice signal, specifically including windowing and framing of the voice signal; detecting voice signal endpoints; performing voice enhancement processing by adopting a signal subspace enhancement algorithm; S2, extracting voice features, and fusing the extracted feature parameters; and S3, establishing a voice acoustic model. According to the intelligent recognition technology for incomplete voice of elderly people, the problems of slight sound amplitude and great influence of environmental noise caused by aging of vocal organs of the elderly people can be reduced, and the voice features fused by adopting the sound parameters can be closer to the voice features of the elderly people, so that data comprehensively representing the voice features of the elderly people can be acquired, and the recognition degree of incomplete voice and fuzzy voice of the elderly people is improved.

The invention discloses a multi-speaker voice synthesis method based on a variational auto-encoder. The method comprises the following steps: extracting a phoneme-level duration parameter and a frame-level acoustic parameter of a to-be-synthesized speaker clean voice, inputting the normalized phoneme-level duration parameter into a first variational auto-encoder, and outputting a duration speakerlabel; inputting the normalized frame-level acoustic parameter into a second variational auto-encoder, and outputting an acoustic speaker label; extracting frame-level linguistic features and phoneme-level linguistic features from speech signals to be synthesized, wherein the speech signals include a plurality of speakers; inputting the duration speaker label and the normalized phoneme-level linguistic features into a duration prediction network, and outputting a current phoneme prediction duration; obtaining the frame-level linguistic characteristics of the phoneme through the current phonemeprediction duration, inputting the frame-level linguistic characteristics and the acoustic speaker label into an acoustic parameter prediction network, and outputting the normalized acoustic parameters of the prediction voice; and inputting the normalized predicted speech acoustic parameters into a vocoder, and outputting a synthesized speech signal.

The invention provides a voice control scene method and a voice control scene system which are applied to a system of mobile equipment. The method comprises the following steps: obtaining voice data in real time; acquiring acoustic features from the voice data according to a pre-trained voice-acoustic neural network model; detecting the acoustic features, and if the acoustic features contain a preset system wake-up acoustic feature, waking up the system of the mobile device; acquiring voiceprint features from the voice data according to a pre-trained voice-voiceprint neural network model; detecting the voiceprint features, and if the voiceprint features are the same as those of the pre-trained user voiceprint, unlocking the system of the awakened mobile device; converting the voice data into character data according to a pre-trained voice-character neural network model; and performing scene event message matching according to the character data and a pre-trained event database to obtain a corresponding scene event message, and executing the scene event message. The efficiency of controlling the scene of the mobile device through voice can be improved, and the user experience of using the mobile device is improved.

The invention discloses a controllable emotional speech synthesis system and method based on an emotional category label. The method comprises steps of text feature extraction, extracting speech text features from an input phoneme sequence; a voice style feature extraction step for receiving acoustic features of a target voice corresponding to the phoneme sequence and extracting voice style features from the acoustic features; a voice style characteristic memorizing step used for obtaining emotional style characteristics of the target voice according to the voice style characteristics; and an acoustic feature prediction step used for predicting synthetic emotional speech acoustic features according to the speech text features and the emotional style features. According to the method, the decoupling degree of the voice style characteristics and the voice text characteristics can be improved, so the style regulation and control result of the synthesized voice is not limited by the text content, controllability and flexibility of the synthesized voice are improved, and emotional labels and emotional data distribution information of the voice in the corpus can be effectively utilized; therefore, the voice style characteristics of each emotion can be extracted more efficiently.

The embodiment of the invention discloses a streaming end-to-end speech recognition method and device, and electronic equipment. The streaming end-to-end speech recognition method comprises the following steps: performing speech acoustic feature extraction on a received speech stream by taking a frame as a unit, and encoding; performing block processing on the coded frame, and predicting the number of activation points which are included in the same block and need to be coded and output; and determining the position of the activation point needing to be decoded and output according to the prediction result so that a decoder decodes at the position of the activation point and outputs an identification result. According to the embodiment of the invention, the robustness of a streaming end-to-end speech recognition system to a noise can be improved, and the system performance and accuracy are increased.