End-to-end speech synthesis network based on embedded system

Abstract

The invention belongs to the technical field of embedded computers, and provides an end-to-end speech synthesis network based on an embedded system, which is characterized in that characters are coded and decoded to generate a Mel spectrogram, and then a picture is converted into a speech file through a vocoder. And meanwhile, under the condition of ensuring that the voice quality is not obviously attenuated, the reasoning speed is improved, real-time performance is realized, and finally, the network is deployed on an embedded platform. By adopting the end-to-end network and using the newest neural network method, the reasoning speed can be greatly improved under the condition of reducing parameters and model calculation amount, and the effect of rhythm adjustability is achieved by feeding forward a rhythm encoder. Characters pass through a front-end encoding and decoding part to generate a Mel spectrogram, and then the Mel spectrogram is converted into the speech file through the vocoder. The method is an end-to-end method. According to the method, the audio can be efficiently synthesized in real time, so that the method is deployed on a subway embedded platform.

Description

technical field [0001] The invention belongs to the technical field of embedded computers, and relates to an end-to-end model speech synthesis vocoder based on an embedded system. Background technique [0002] With more and more voice interactions with machines, speech synthesis technology is being used more and more in reality, such as AI synthesized anchors and subway broadcasts. However, the synthesis quality or small changes in speech can have a big impact on customer experience and customer preference. Therefore, high-quality real-time speech synthesis remains a challenging task. [0003] Currently, advanced speech synthesis models include statistical parameter neural network speech synthesis models and end-to-end speech synthesis models. Text-to-speech synthesis is usually divided into two parts. The first step is to convert the text into time-aligned features, such as mel-spectrograms. The second model is to convert these time-aligned features into audio samples. ...

AI Technical Summary

Problems solved by technology

However, the splicing method used in this invention is difficult to obtain high quality, compared with the neural network method widely used at present, and it is not real-time in terms of inference speed.

Images