Noisy speech gender identification method and system based on lightweight neural network

Abstract

The invention discloses a noisy speech gender identification method and system based on a lightweight neural network. The noisy speech gender identification method comprises the following steps of: synthesizing noisy speech based on pure male and female speech audios and pure noise audios; extracting audio features of the noisy speech, wherein the audio features only comprise a plurality of BFCC features, and first-order time derivatives and second-order time derivatives, fundamental tone gain values, fundamental frequency period values and voice short-time zero-crossing rates of part of the BFCC features; constructing a lightweight neural network model based on the audio features and performing training, wherein the lightweight neural network model comprises a voice activity branch, a noise reduction branch and a gender classification branch; and carrying out noisy speech gender prediction based on the lightweight neural network model. According to the noisy speech gender identification method, the lightweight neural network model containing noise reduction branches and male and female sound classification branches is established within a time range of 30 ms, the accuracy rate ishigh, and the method is suitable for actual application scenes containing unknown noise.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to a method and system for gender recognition of noisy speech based on a lightweight neural network. Background technique [0002] A good voice will give the listener an inexplicable appeal in daily communication. Gladstone, who served as the British Prime Minister for four times, said: "Voice is the most powerful instrument in communication." Certain special occupations and occasions require voice Higher, such as hosting, broadcasting, live broadcast and game voice chat, etc., but not everyone has a nice voice, so fine adjustment of male and female voices is required. [0003] The first step in the fine-tuning of male and female voices is to realize real-time speech gender recognition, which is a classic binary classification problem. There are two methods for speech gender recognition in the prior art, one is the traditional machine learning method, and the Perform fea...

AI Technical Summary

Problems solved by technology

[0003] The first step in the fine-tuning of male and female voices is to realize real-time speech gender recognition, which is a classic binary classification problem. There are two methods for speech gender recognition in the prior art, one is the traditional machine learning method, and the Perform feature extraction to achieve data dimensionality reduction, and input it into the Gaussian mixture model or SVM model to train the model parameters. The traditional machine learning method has a small model, but it relies heavily on the accuracy of audio feature extraction. In the environment of unknown noise, the existing It is difficult for feature extraction methods to extract desired audio features, such as pitch pitch extraction. Therefore, in a noisy environment, the accuracy of traditional machine learning for male and female voice recognition is not high. Noise reduction algorithm is usually a trade-off between audio quality and noise reduction range, and some noise will still remain after noise reduction. In addition, the noise reduction algorithm will also introduce delay. When the noise reduction algorithm and the gender classification algorithm When connected in series, the time delay is not lower than the superposition of the two; the other is based on the deep learning neural network method, extracting speech acoustic features, building a neural network model, and then using softmax for classification. Compared with traditional machine learning The method, the way of neural network can avoid the dependence on the accuracy of the audio features of the input model. The acoustic features can be high-dimensional primary features, which can get higher recognition accuracy in the case of certain noise, but based on deep learning neural network In this method, since the audio segment input into the network is usually 1s~4s in length, although high accuracy is obtained, real-time performance is sacrificed, and the input of the neural network is usually a high-dimensional feature such as time-spectrum STFT or MFCC, and there is network training. There are many parameters and large models, and it is difficult to apply them to embedded devices

Images