Nanjing Silicon Intelligence Technology Co., Ltd.

Zhiqiang Mao

Huapeng Sima

Xuefei Gong

The present disclosure proposes a speech conversion scheme for non-parallel corpus training, to get rid of dependence on parallel text and resolve a technical problem that it is difficult to achieve speech conversion under conditions that resources and equipment are limited. A voice conversion system and a training method therefor are included. Compared with the prior art, according to the embodiments of the present disclosure: a trained speaker-independent automatic speech recognition model can be used for any source speaker, that is, the speaker is independent; and bottleneck features of audio are more abstract as compared with phonetic posteriorGram features, can reflect decoupling of spoken content and timbre of the speaker, and meanwhile are not closely bound with a phoneme class, and are not in a clear one-to-one correspondence relationship. In this way, a problem of inaccurate pronunciation caused by a recognition error in ASR is relieved to some extent. Pronunciation accuracy of audio obtained by performing voice conversion by the bottleneck feature is obviously higher than that of a phonetic posteriorGram based method, and timbre is not significantly different. By means of a transfer learning mode, dependence on training corpus can be greatly reduced.

Voice conversion system and training method therefor

The present disclosure provides a GAN-based speech synthesis model, a training method, and a speech synthesis method. According to the speech synthesis method, to-be-converted text is obtained and is converted into a text phoneme, the text phoneme is further digitized to obtain text data, and the text data is converted into a text vector to be input into a speech synthesis model. In this way, target audio corresponding to the to-be-converted text is obtained. When a target Mel-frequency spectrum is generated by using a trained generator, accuracy of the generated target Mel-frequency spectrum can reach that of a standard Mel-frequency spectrum. Through constant adversary between the generator and a discriminator and the trainings thereof, acoustic losses of the target Mel-frequency spectrum are reduced, and acoustic losses of the target audio generated based on the target Mel-frequency spectrum are also reduced, thereby improving accuracy of audio synthesized from speech.

GAN-based speech synthesis model and training method

Embodiments of the present application provide a method and system for outputting target audio, a readable storage medium, and an electronic device. The method includes: inputting source audio into a phonetic posteriorgram PPG classification network model to obtain a PPG feature vector, where the PPG feature vector is used for indicating a phoneme label corresponding to each frame of the source audio, and the PPG feature vector contains text information and prosodic information of the source audio; inputting the PPG feature vector into a voice conversion network model, and outputting an acoustic feature vector of target audio based on the phoneme label corresponding to the PPG feature vector, where the target audio contains a plurality pieces of audio with different timbres; and inputting the acoustic feature vector of the target audio into a voice coder, and outputting the target audio through the voice coder.

Growing community of inventors

Jiangsu, China

Zhiqiang Mao