Language recognition method and device, and device for language recognition

Abstract

The embodiment of the invention provides a language recognition method and device and a device for language recognition. The language recognition method specifically comprises the following steps: performing acoustic feature extraction on a voice segment to obtain acoustic features based on a frame sequence; inputting the acoustic features into a deep neural network, the deep neural network comprising a bottleneck layer and a time recursion layer, and the output of the bottleneck layer being connected to the time recursion layer; extracting a bottleneck feature sequence from the bottleneck layer, and inputting the bottleneck feature sequence into the time recursion layer to output a high-level feature sequence through the time recursion layer; and determining a language type correspondingto the voice segment according to the high-level feature sequence. According to the embodiment of the invention, the accuracy of language recognition can be improved.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a language recognition method, device and a device for language recognition. Background technique [0002] Language recognition refers to the technology of automatically processing a piece of speech and judging its language type through a computer. Language recognition technology is mainly used in the front end of the multilingual speech processing system. Automatic classification of speech through language recognition technology can save resources, avoid tedious manual classification, and greatly improve work efficiency. [0003] At present, language recognition systems usually use i-Vector features, which use a fixed-length low-dimensional vector to represent a segment of speech. In the process of extracting i-Vector features, sufficient statistics of acoustic features need to be calculated, therefore, the effective data of each speech segment must be long enough to ensure t...

AI Technical Summary

Problems solved by technology

However, such conditions are difficult to guarantee in practical applications. For example, in the language recognition system of mobile phones, longer speech segments (such as more than 30 seconds) can be used for training in the training phase, while in the testing phase, the length of the speech segment is generally Only 3 seconds to 5 seconds

Since the extraction of i-Vector features requires the estimation of sufficient statistics of the acoustic features, insufficient representation of short test speech segments will lead to a decrease in system performance, making it difficult to improve the accuracy of the language recognition system

Images