Speech recognition method and device supporting multi-language mixing, equipment and storage medium

Abstract

The invention relates to artificial intelligence, and provides a speech recognition method supporting multi-language mixing. The method comprises the following steps: acquiring speech features of training data; acquiring high-dimensional features respectively corresponding to the voice features through at least two parallel networks; performing feature splicing on the high-dimensional features output by the parallel network, and obtaining splicing features corresponding to the training data; training a neural network model based on the spliced features until the neural network model converges to a preset range, and forming a speech recognition model; and performing voice recognition on a to-be-recognized multi-language mixed signal through the voice recognition model. According to the invention, the recognition precision of the multi-language mixed voice can be improved.

Description

technical field [0001] The present invention relates to the field of artificial intelligence, in particular to a method, device, equipment, electronic equipment and computer-readable storage medium supporting multilingual mixed speech recognition. Background technique [0002] With the development of artificial intelligence technology, more and more intelligent hardware has entered people's lives, and voice input, as the most natural and convenient way in human-computer interaction, has gradually become the mainstream interaction method. Therefore, the performance of speech recognition directly determines the level of interaction quality. At the same time, due to the continuous deepening of globalization, cultural and language exchanges between different regions are becoming more and more frequent. People's voices are often mixed with different languages, such as Mandarin-English, Cantonese-English, and Mandarin-Cantonese. The current speech recognition system has a good pe...

AI Technical Summary

Problems solved by technology

[0004] Among the above-mentioned schemes, scheme 1 and scheme 2 need to rely on high-quality language classifiers, and consume a lot of computing resources, while scheme 3 has a simple system design and low computational complexity, but it does not have enough language discrimination, especially For similar pronunciation units and languages ​​with a small proportion of time in the training corpus, the recognition performance is generally poor, resulting in low recognition accuracy and poor effect

Images