Training method for hybrid frequency acoustic recognition model and speech recognition method

Abstract

The invention discloses a training method for a hybrid frequency acoustic recognition model and a speech recognition method, and belongs to the technical field of speech recognition. The training method comprises the steps of obtaining a first type of speech features of a first speech signal, and processing the first type of speech features so as to obtain corresponding first speech training data;obtaining a first type of speech features of a second speech signal, and processing the first type of speech features so as to obtain corresponding second speech training data; obtaining a second type of speech features of the first speech signal and a second type of speech features of the second speech signal according to a power spectrum; forming a preliminary recognition model of the hybrid frequency acoustic recognition model according to pre-training of the first speech signal and the second speech signal; and performing supervised parameter training on the preliminary recognition modelaccording to the first speech training data, the second speech training data and the second type of speech features so as to form the hybrid frequency acoustic recognition model. The technical schemehas the beneficial effect that the recognition model has good robustness and generalization.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to a training method for a mixed-frequency acoustic recognition model and a speech recognition method. Background technique [0002] In the prior art, due to differences in usage environments, data transmission requirements, and technical means used for transmission, there are large differences between different recording devices and storage methods, and the main difference lies in the difference in sampling frequency. For example, voice data with a sampling frequency of 8kHz usually comes from telephone recordings, so in traditional voice recognition methods, the telephone data is specially used for training to form an acoustic recognition model of 8kHz data. Correspondingly, the speech data with a sampling frequency of 16kHz usually comes from desktop recordings, and the desktop data is also used to train an acoustic recognition model of 16kHz data. Therefore, the trad...

AI Technical Summary

Problems solved by technology

[0003] Although the use of a dedicated acoustic model can make the test environment and the training environment more compatible, it will also bring many disadvantages: First, the update and maintenance of the recognition model will be very cumbersome, and it is necessary to conduct special training for each dedicated acoustic model. Update and maintenance; Second, training each dedicated acoustic model separately will make the training data of each model insufficient, and the robustness and generalization of the model will also be limited

Images