Non-autoregressive speech recognition training decoding method based on parameter sharing and system thereof

Abstract

The invention discloses a non-autoregressive speech recognition training decoding method based on parameter sharing and a system thereof, and the training method comprises the steps: extracting the features of speech training data, and forming an acoustic feature sequence; performing acoustic coding on the acoustic feature sequence, and outputting an acoustic coding state sequence; performing non-autoregressive decoding on the acoustic coding state sequence and the blank filling sequence, and calculating non-autoregressive cross entropy loss in combination with text labeling training data; performing autoregression decoding on the acoustic coding state sequence and the text labeling training data, and calculating autoregression cross entropy loss in combination with the text labeling training data; carrying out weighting according to the non-autoregression cross entropy loss and the autoregression cross entropy loss to obtain joint loss, calculating a gradient, and carrying out back propagation; and circularly executing until the training is completed; and the decoding method comprises the step of performing speech recognition through the trained model; wherein the system comprises an acoustic feature sequence extraction module, an acoustic encoder, a non-autoregression decoder, an autoregression decoder and a joint loss calculation module.

Description

technical field [0001] The invention relates to the technical field of electronic signal processing, in particular to a non-autoregressive speech recognition training and decoding method and system based on parameter sharing. Background technique [0002] As the entrance of human-computer interaction, speech recognition is an important research direction in the field of artificial intelligence. End-to-end speech recognition discards the pronunciation dictionary, language model and decoding network that the hybrid speech recognition model relies on, and realizes the direct conversion of audio feature sequences to text sequences. The classic encoding and decoding model uses word autoregressive decoding, and its encoder encodes the input speech into a high-level feature representation; the decoder starts from the initial symbol, and gradually predicts the corresponding text based on the output of the editor sequence until the end marker is predicted. The timing dependence of ...

AI Technical Summary

Problems solved by technology

However, the non-autoregressive model faces problems such as training difficulties, and

Images