Method for training filter model and speech recognition method

Abstract

The application provides a method and device for training a filter model. The method comprises the following steps: determining N original syllables, wherein the N original syllables are syllables included in the actual pronunciation of a first corpus; determining N recognition syllables, wherein the N recognition syllables are syllables of a recognition result obtained by first speech recognitionprocessing of a sound signal of the first corpus, and the first speech recognition processing includes filter processing based on the filter model and recognition processing based on a speech recognition engine; determining N syllable distances according to the N original syllables and the N recognition syllables, wherein the N syllable distances correspond to the N syllables one to one, and theN original syllables and the N recognition syllables constitute N syllable pairs, each syllable pair includes one original syllable and one recognition syllable which correspond to each other; each syllable distance is used for indicating the similarity between the original syllable and the recognition syllable included in the corresponding syllable pair. Thus, the recognition accuracy of the speech recognition engine can be improved.

Description

technical field [0001] The present application relates to the technical field of speech recognition, and more specifically, to a method for training a filtering model, a method for speech recognition, a training device, a speech recognition device, and a speech recognition system. Background technique [0002] Automatic speech recognition (Automatic Speech Recognition, ASR) is a key technology of the voice interaction system, and the ASR engine (also called a speech recognition engine) is responsible for converting speech signals into text. figure 1 It is a schematic diagram of an example of speech recognition performed by an ASR engine. Such as figure 1 Said, the sound is collected through the sound pickup device, and the obtained speech signal is converted into a syllable sequence (for example, the initial and final sequence in Chinese Pinyin) by the acoustic model after the feature module extracts (frequency domain) features. Then, a character sequence (for example, a C...

AI Technical Summary

Problems solved by technology

[0003] However, in the prior art, the corresponding relationship between the syllable sequence and the character sequence in the language model is obtained based on the preset database training. In actual use, it may be affected by the environment and the user's pronunciation habits, etc. The sound signal of the voice in the database is inconsistent with the actual collected sound signal, resulting in the final failure to recognize the voice

Images