Speaker counting method and system

Abstract

Embodiments of the present invention provide a speaker counting method. The method includes: establishing an end-to-end speaker counting model based on a deep convolutional neural network; using an original audio waveform as an input of the end-to-end speaker counting model; determining according to an output result of the end-to-end speaker counting model number of speakers. Embodiments of the present invention provide a speaker counting system and also provide an optimization method and system for an aliased speech detection model. The embodiment of the present invention adopts the method of end-to-end speech aliasing detection and speaker counting of the original waveform input, and uses the neural network to directly extract the depth features from the original speech for subsequent tasks, so that it is easier to obtain the features matching the corresponding tasks, and the accurate determination It is more suitable for the scene of multiple people speaking at the same time in real life, and provides additional information for the back-end speech processing system, thereby promoting the recognition, separation and enhancement of aliased speech.

Description

technical field [0001] The invention relates to the field of voice detection, in particular to a speaker counting method and system. Background technique [0002] Although intelligent speech is constantly developing, the performance of the speech processing system will still be severely degraded under complex scene conditions, for example, a cocktail party scene that includes multiple talkers overlapping speech at the same time and involves other background noises. In this case, if the number of speakers is given in advance, the performance of overlapping speech processing can be significantly improved, and then accurate overlapping speech detection and speaker counting are very useful for later speech detection and recognition. For the above problems, the speaker counting method based on convolutional neural network is usually used. The input is the artificially designed acoustic features, including the speech signal envelope, histogram, Mel frequency cepstral coefficient, ...

AI Technical Summary

Problems solved by technology

[0005] In order to solve at least the traditional acoustic features that are artificially designed in the prior art, speech features that are highly correlated with speaker counts or aliasing speech detection cannot be extracted, and the artificially designed traditional acoustic features tend to fall into local optimum question

Images