Speech recognition method based on CLDNN+CTC acoustic model

Abstract

The invention discloses a voice recognition method based on a CLDNN+CTC acoustic model, and the method comprises the steps: 1, obtaining a real-time voice signal, carrying out the feature extraction of the voice signal, and obtaining a frame-by-frame acoustic feature sequence; 2, taking the acoustic characteristic sequence as the input of a CLDNN+CTC acoustic model, and outputting a phoneme sequence; and 3, establishing a decoding model for converting the phoneme sequence into the character sequence, taking the phoneme sequence as the input of the decoding model, and outputting the character sequence through the decoding model. The method is a voice recognition method based on two-stage end-to-end (seq2seq), and comprises an end-to-end model of a 'voice-phoneme sequence' and an end-to-endmodel of a 'phoneme sequence-character sequence', the two models are different from the end-to-end model of an existing 'voice-character sequence', the two models do not need super-large-scale corpustraining, the advantages of the two parts can be complementary, and a language model can make up for the deficiency of an acoustic model in a noise environment.

Description

technical field [0001] The invention relates to the field of speech recognition, in particular to a speech recognition method based on a CLDNN+CTC acoustic model. Background technique [0002] Speech is the most common and effective way of human interaction, and it has always been an important part of the research field of human-computer communication and human-computer interaction. Human-computer voice interaction technology, which is composed of speech synthesis, speech recognition and natural language understanding, is recognized as a difficult and challenging technical field in the world. At the same time, speech recognition technology can enter various industries such as industrial production, electronic communication, automotive electronics, medical care, service education, etc., and will lead the information technology revolution to a new level. [0003] Speech recognition, also known as automatic speech recognition (Automatic Speech Recognition, ASR). Automatic spe...

AI Technical Summary

Problems solved by technology

This strategy of adding noise is not universal, the noise is different in different scenarios, adding noise to augment data is not a universal solution;

[0019] The speech recognition method disclosed in Invention Patent 3 uses a multi-channel convolutional neural network as the acoustic model. The same voice data enters the same three-channel convolutional network respectively, which cannot extract more discriminative features, and at the same time makes the network structure more Complex, requires a large amount of training data, and is prone to overfitting;

[0020] The speech recognition technology disclosed in Invention Patent 4 is based on a simple DCNN network model and outputs speech sequences end-to-end. Since it uses a CNN-based structure, it has limited processing capacity for data with strong temporal characteristics such as speech; at the same time , the entire model has only 9 layers. For speech recognition with a large Chinese vocabulary, the model fitting ability is limited;

This modeling method involves three models, and the three models are interdependent. The shortcomings of any model will restrain other models, resulting in a sharp drop in overall performance.

The model combines syllables and acoustic features to determine whether the speech is the text, which cannot substantially improve the recognition accuracy

Images