Speech synthesis method and system

Abstract

The invention provides a speech synthesis method. The speech synthesis method comprises the steps that interactive instructions are acquired and analyzed, when the interactive instructions comprise aspeech broadcasting instruction, the speech broadcasting instruction is responded to; an acoustic model generated by a completely parallel attention mechanism is called, and to-be-synthesized text data or to-be-synthesized phonetic symbol data is input into the acoustic model in real time to obtain to-be-synthesized acoustic features; the to-be-synthesized acoustic features are input into an acoustic synthesizer, synthesized speech data is obtained through the acoustic synthesizer, and the speech data is output. The invention provides the end-to-end acoustic model and a training mode, which can completely utilize the advantages of parallel computing. The completely parallel attention mechanism and a convolution structure are utilized, so that the time required for generating the acoustic features is greatly shortened compared with the prior art, and the quality of the generated acoustic features can also be ensured, and therefore, under the situation of ensuring the quality of synthetic speech, the synthetic time is shortened, and the efficiency of the speech synthesis is improved.

Description

technical field [0001] The invention relates to the field of artificial intelligence, in particular to a speech synthesis method and system. Background technique [0002] Speech synthesis, also known as text-to-speech (Text to Speech) technology, can convert any text information into a standard and smooth voice in real time. It involves many disciplines and technologies such as acoustics, linguistics, digital signal processing, and computer science. The main problem to be solved is how to convert text information into audible sound information. [0003] Currently, the data used for speech synthesis training generally consists of plain text (or phonetic symbols) and corresponding audio. There is no boundary information of phonetic symbols in text (or phonetic symbols), which makes it very difficult to determine the sequence correspondence between acoustic features and text (phonetic symbols). Moreover, the current acoustic training model can only generate acoustic features ...

AI Technical Summary

Problems solved by technology

There is no boundary information of phonetic symbols in the text (or phonetic symbols), which makes it very difficult to determine the sequence correspondence between the acoustic features and the text (phonetic symbols)

Moreover, the current acoustic training model can only generate acoustic features one time period at a time, and cannot make full use of the advantages of parallel computing, making its time delay very long

Images