A Robust Speech Recognition Method Based on Acoustic Model Array

Abstract

The invention discloses a robust speech recognition method based on an acoustic model array, which includes a training phase and a testing phase. In the training phase, multiple upper frequency limits are set for the training voice according to the highest frequency of the voice, multiple sets of feature vectors are extracted, and model training is performed to obtain an array of acoustic models. In the test phase, firstly, based on a small amount of adaptive speech in the test environment, the upper limit frequency of the test speech is estimated; then the acoustic model matching the upper limit frequency of the test speech is selected from the acoustic model array, and its parameters are adjusted to obtain the test environment acoustics Finally, feature extraction is performed according to the upper limit frequency of the test speech to obtain the feature vector of the noisy test speech, and the acoustic decoding is performed with the test environment acoustic model to obtain the recognition result. The invention can improve the performance of the speech recognition system in the noise environment and improve the robustness of the system.

Description

technical field [0001] The invention belongs to the technical field of speech recognition, and specifically relates to extracting multiple sets of feature vectors in different frequency ranges according to a plurality of speech upper limit frequencies, constructing an array of acoustic models, and compensating the acoustic model matched with the upper limit frequency of the test speech to improve speech recognition. A model adaptation method for identifying system robustness. Background technique [0002] In the practical application of the speech recognition system, due to the influence of speech variability such as environmental noise, the pre-trained acoustic model often does not match the feature parameters extracted in the test environment, which will lead to a serious decline in the performance of the speech recognition system. Therefore, it is necessary to compensate the environment mismatch to improve the recognition performance of the speech recognition system. [...

AI Technical Summary

Problems solved by technology

However, the energy of the high-frequency part of the speech spectrum is small, and it is easily covered by noise in a noisy environment. Therefore, in a noisy test environment, the high-frequency part of the noisy speech spectrum is noise components, which not only cannot provide effective effects for speech recognition , and will adversely affect the model adaptation of the backend

Images