End-to-end speech recognition model based on multi-level identification and modeling method

Abstract

The invention provides an end-to-end speech recognition modeling method based on multi-level identification, and the method comprises the steps: decoding inference which employs a post-inference algorithm; the post-inference algorithm comprises the steps that a model corresponding to a fine-grained text sequence generates a posterior probability output sequence, the output sequence can uniquely correspond to a coarse-grained sub-sequence calculation model to generate a log-likelihood value of the coarse-grained sub-sequence, and the log-likelihood value serves as cross validation of an existing prediction output sequence; and according to the likelihood probability scores obtained through calculation in the two steps, an existing decoding path is cut, and it is guaranteed that the search path is controlled within the beam width range.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to an end-to-end speech recognition model and modeling method based on multi-level identification. Background technique [0002] End-to-End (E2E) Automatic Speech Recognition (ASR) based on the encoding-decoding framework directly models the sequence mapping relationship between the input audio sequence and the output text. The advantages of a concise framework and no need for linguistic background knowledge make this structure gradually sought after by academia and industry. [0003] In end-to-end ASR, input speech sequences can be mapped to text sequences at different levels. The mapping relationship between speech sequences and text sequences is one-to-many. In Chinese ASR, the text sequence can be composed of pinyin and Chinese characters; the English and Chinese text sequence can be composed of words (word) and characters (character). [0004] In general, modeling...

AI Technical Summary

Problems solved by technology

Its shortcomings lie in the large amount of training data required and the uneven distribution of samples.

However, the construction of character-level text units does not take into account the influence between adjacent units in the output text sequence, and cannot take into account issues such as co-pronunciation and non-pronunciation of speech.

Images