Speech recognition acoustic model building method and device, and electronic equipment

Abstract

Embodiments of the invention provide a speech recognition acoustic model building method and device, and electronic equipment. The method includes building an acoustic model based on a deep neutral network, modeling units of the acoustic model being context free speech units, obtaining training data, converting speech signals in the training data into eigenvector sequences, and converting annotated texts corresponding to the speech signals into modeling unit based label sequences; training the acoustic model by using the eigenvector sequences and the label sequences; inputting the eigenvectorsequences into the acoustic model, and counting the recognition error rate of each modeling unit; splitting the modeling units whose recognition error rates are higher than threshold values into morethan one modeling units; updating the label sequences according to the modeling units obtained by splitting; and retraining the acoustic model according to the eigenvector sequences and updated labelsequences. According to the embodiments, the correct rates of speech recognition can be effectively enhanced.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to a method and device for establishing an acoustic model for speech recognition and electronic equipment. Background technique [0002] Speech is a typical time-sequence signal, which is very complex due to background noise, channel, scene, speaker and other factors. Speech recognition technology refers to the process of converting a speech signal into text. [0003] At present, speech recognition technology mainly adopts a hybrid modeling method based on HMM (Hidden Markov Model, Hidden Markov Model), for example, GMM (Gaussian Mixture Model, Gaussian Mixture Model) + HMM, DNN (DeepNeutral Network, deep neural network) +HMM, CNN (Convolution Neural Network, convolutional neural network) and LSTM (Long Short Time Memory, long short-term memory) + HMM, etc. [0004] In the prior art, hybrid modeling methods use state as a modeling unit to establish an acoustic model for...

AI Technical Summary

Problems solved by technology

Here, in order to ensure the correct rate of speech recognition, state modeling requires the modeling unit to be a pronunciation unit with a short duration, less internal changes, and context-dependent. The generalization performance of the model is poor, and the accuracy of speech recognition is low. Usually, if the data set, scene or speaker is changed, the accuracy of speech recognition will drop sharply.

Images