Child speech recognition model training corpus screening method

Abstract

The invention discloses a child speech recognition model training corpus screening method. The method comprises the following steps: inputting a phoneme sequence and a child reference audio into a TTS synthesizer to obtain a plurality of synthesized audios; acquiring reference feature information of the child reference audio and a plurality of pieces of synthetic feature information of the plurality of synthetic audios; and screening the plurality of synthetic audios according to the reference feature information and the plurality of pieces of synthetic feature information. According to the child speech recognition model training corpus screening method provided by the invention, the TTS synthesizer is adopted to generate the synthetic audio, and the synthetic audio is screened according to the child reference audio used for generating the synthetic audio to obtain the high-quality synthetic audio, so that the corpus for training the child speech recognition model is expanded. Therefore, the invention solves the problem that a child corpus is difficult to collect, the quality of corpora in the child corpus is ensured, and a child speech recognition model with good performance can be trained.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to a method for screening children's speech recognition model training corpus, electronic equipment and a storage medium. Background technique [0002] The performance of automatic speech recognition (ASR) systems has improved significantly since the introduction of deep neural networks. With a large amount of training data and advanced model structures, ASR models are now able to achieve human-equal performance. However, to the best of our knowledge, speech recognition in children remains a challenging task despite many efforts. [0003] One challenge of speech recognition in children is the lack of data, as child corpora are difficult to collect. In addition, children have inherently high variability in physical and vocal features and expressions. To overcome these difficulties, channel length normalization is proposed to reduce the acoustic variability between speak...

AI Technical Summary

Problems solved by technology

However, using synthetic speech generated by TTS systems trained on child speech data is problematic because child speech involves substandard or unclear articulation

As a result, the quality of synthesized speech is inconsistent in this case

Images