Speech enhancement model training and application method, device and equipment, equipment and storage medium

Abstract

The invention relates to the field of artificial intelligence speech enhancement, and particularly discloses a speech enhancement model training and application method, device and equipment, and a storage medium, and a speech enhancement model with small distortion and noise reduction capability is obtained through joint modeling of the speech enhancement model and a vocoder. The method comprises the following steps: performing analog noise addition on clean speech to obtain noisy speech, and determining a target time-frequency mask according to the clean voice and the noisy speech; extracting noisy Mel spectrum features from the noisy speech, inputting the noisy Mel spectrum features into the speech enhancement model, outputting a predicted time-frequency mask, and determining a first loss value according to the predicted time-frequency mask and a target time-frequency mask; obtaining de-noised Mel spectrum features according to the predicted time-frequency mask and the noisy Mel spectrum features; and inputting the de-noised Mel spectrum features into a vocoder to obtain synthetic speech, and determining a second loss value according to the synthetic speech and the clean speech. And optimizing parameters of the speech enhancement model and the vocoder according to the first loss value and the second loss value to obtain a trained speech enhancement model.

Description

Technical field [0001] The present application relates to artificial intelligent speech enhancements, and in particular, a training method, an application, apparatus, computer device, and storage medium, a speech enhancement model. Background technique [0002] Voice synthesis technology has been able to generate a relatively close voice, but to create a high quality speech synthesis system, high quality voice training data is required. High quality speech data typically requires recording in a muffle room equipped with high-end recording equipment and having a very low under noise. So the recording of high quality voice data will cost a lot of cost, and if it is recorded in ordinary indoor environment, the bottom noise and other environmental noise and reverberation will be collected by the recording equipment. If voice enhancement is performed using the current mainstream depth neural network, a larger distortion is often caused to make audio or acoustic characteristics before ...

AI Technical Summary

Problems solved by technology

Therefore, the recording of high-quality voice data will consume a lot of cost, and if it is recorded in an ordinary indoor environment, the background noise and other environmental noises and reverberation will be collected or even amplified by the recording equipment

If the current mainstream deep neural network method is used for speech enhancement, it will often cause large distortion, and the audio or acoustic features will suffer from large information loss before the speech synthesis model training.

Images