Method for building acoustic model and speech decoding method based on acoustic model

Abstract

The invention provides a method for building an acoustic model and a speech decoding method based on the acoustic model. The method comprises the following steps: (101) based on training data, calculating three-factor Gaussian mixture model statistics which are required by the acoustic model; (102) adopting a self-clustering method to calculate a decision tree problem set according to the statistics, and adopting a decision tree algorithm to conduct partitioning clustering on the statistics based on the decision tree problem set to obtain isotopic phonemes; (103) combining a basic phoneme set with the isotopic phonemes to obtain a phoneme set containing the isotopic phonemes, and adopting the decision tree algorithm to treat an original speech annotation to obtain a speech annotation containing the isotopic phonemes; (104) based on the phoneme set containing the isotopic phonemes and the speech annotation containing the isotopic phonemes, carrying out acoustic model training according to an acoustic model training method to build the acoustic model containing the isotopic phonemes. The method for building the acoustic model aims to solve the problem that the degree of acoustic model confusion is high in an agglutinative language speech recognition system.

Description

technical field [0001] The invention relates to the field of speech recognition, and is mainly aimed at an agglutinative speech recognition system. Background technique [0002] In language morphology, according to whether the language needs to rely on the change of word endings to express its grammatical relationship, it can be divided into analytic language and synthetic language. Classification. Adhesive language is a kind of synthetic language, which belongs to the comprehensive language with high inflection. Its word-level units are usually composed of a large number of morpheme connections, which are called agglutinative characteristics. Since the speech recognition system was originally designed for analytic languages ​​and quasi-analytic languages, such as Chinese and English, the emergence of sticky features has brought many new problems to the traditional speech recognition system, which requires further improvement and improvement. improved. [0003] The proble...

AI Technical Summary

Problems solved by technology

However, the number of word-level units in agglutinative language has increased dramatically due to the existence of agglutinative properties, and the number of common words has far exceeded the size that the dictionary can accommodate, so it is not suitable as the basic modeling unit of the language model; at the same time, the secondary natural language unit phoneme (or word , depending on the language and the sub-units are different) is not suitable as the basic modeling unit of the language model, because the sticky characteristics will aggravate the short-sighted phenomenon of this level unit

Second, in terms of acoustic models, the cohesion of phonemes in agglutinative languages ​​will lead to a large number of co-articulation phenomena, that is, the same phoneme will have many different pronunciations depending on its location.

But the second problem has not yet found an effective solution, which is one of the difficulties that plague the speech recognition of agglutinative language

Images