Noise-robustness acoustic modeling method based on posterior knowledge supervision

Abstract

The invention discloses a noise-robustness acoustic modeling method based on posterior knowledge supervision and belongs to the voice man-machine interaction technology field. The method comprises thefollowing steps of acquiring the posterior probability distribution of a clean voice through the training of a teacher model; and taking the posterior probability distribution of the clean voice as astandard so as to supervise the training of a student model so that the student model infinitely approaches the posterior probability distribution of the teacher model, wherein the teacher model is aclean voice training model and the student model is a voice training model with a noise. By using the modeling method, an established acoustic model has high environment robustness and shows excellent anti-noise performance.

Description

technical field [0001] The invention belongs to the technical field of voice human-computer interaction, in particular to a noise robust acoustic modeling method based on posterior knowledge supervision. Background technique [0002] In recent years, with the development of speech recognition, natural language processing, deep learning and other technologies and the continuous deepening of market demand, the research and development and application of voice interactive products have gradually become a new hot spot; on the other hand, due to the complexity of actual application scenarios , the operation of the voice interaction system is usually in a low signal-to-noise ratio environment. Due to the insufficient anti-interference ability to noise, the system interaction process often has low voice recognition accuracy or chaotic human-computer interaction, which leads to service objects The poor interactive experience has greatly restricted the market application and promotio...

AI Technical Summary

Problems solved by technology

[0007] Although the above four methods can effectively improve the environmental robustness of the acoustic model, there are two problems in theory and application: First, the above methods only supervise the noise reduction of noisy speech through clean speech or use noisy speech Fitting clean speech to reduce the difference between the two does not fully tap the implicit knowledge of clean speech, and the extraction of information is not sufficient; on the other hand, the acoustic feature extraction module in the above four types of methods and the subsequent The training recognition process is independent of each other, without considering the internal connection between the modeling and feature extraction units, which makes the objective function of model training deviate from the overall performance index of the system, and the extracted speech features contain some redundant information. These redundant information are usually not robust to noise, resulting in suboptimal performance of the entire acoustic network

Images