Optimization method and system for single-channel speech recognition model

Abstract

The embodiment of the present invention provides an optimization method for a single-channel speech recognition model. The method includes: receiving each single-person voice with a real label vector, mixing the voices of multiple people, inputting the voice features extracted from each single-person voice to a target teacher model, and obtaining a target soft label vector corresponding to each single-person voice; The multi-person mixed speech is input to the end-to-end student model, and the output arrangement is determined; according to the output label vector of each person in the multi-person mixed speech with the determined output arrangement, the knowledge distillation loss and direct loss are determined; when determined according to the knowledge distillation loss and direct loss The end-to-end student model is optimized according to the joint error when the joint error of . The embodiment of the present invention also provides an optimization system for a single-channel speech recognition model. In the embodiment of the present invention, it is easier to learn good parameters, and at the same time, the model is relatively simplified, and the better parameters enable the trained student model to have better performance.

Description

technical field [0001] The invention relates to the field of speech recognition, in particular to an optimization method and system for a single-channel speech recognition model. Background technique [0002] With the development of intelligent voice, more and more devices have the function of speech recognition. However, due to the consideration of the usage scenarios of different devices, some devices are only equipped with a single microphone, and some devices are equipped with multiple microphones. Microphones are so-called single-channel and multi-channel. Since there is only a single microphone, this type of device has poor recognition performance when dealing with speech conversations such as banquets where multiple people speak at the same time and are mixed together. For this purpose, the knowledge distillation method of single-channel multi-speaker speech recognition based on bidirectional long-short-term memory network-cyclic neural network, or an end-to-end sing...

AI Technical Summary

Problems solved by technology

[0005] In order to at least solve the traditional model in the prior art is more complex, the training process is cumbersome, the training effect is not good, and the performance is poor

Images