Ghost and iLPCnet-based Mongolian speech synthesis method

• Main Authors: DAI QIN, ZHANG WENJING, SIRLING GER RILE, REN-QING DAOERJI, SA HEYA
• Format: Patent
• Language: Chinese; English
• Published: 29.07.2022
• Subjects: ACOUSTICS; CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; HANDLING RECORD CARRIERS; MUSICAL INSTRUMENTS; PHYSICS; PRESENTATION OF DATA; RECOGNITION OF DATA; RECORD CARRIERS; SPEECH ANALYSIS OR SYNTHESIS; SPEECH OR AUDIO CODING OR DECODING; SPEECH OR VOICE PROCESSING; SPEECH RECOGNITION

The invention discloses a ghost and iLPCnet-based Mongolian speech synthesis method, and the method comprises the steps: carrying out the alignment of a Mongolian phoneme information sequence based on a Bang pre-training model; on the basis of a ghost acoustic model, acoustic features are generated according to the phoneme sequence; the iLPCnet model is used as a vocoder, and conversion from acoustic features to voice waveforms is carried out. The Mongolian text is converted into phonemes by using the Encoder-Decoder model, then the phonemes are directly generated into the mel frequency spectrum by using the ghost-based acoustic model, and the mel frequency spectrum is directly converted into the voice waveform by the iLPCnet vocoder, so that the method can be seamlessly integrated to an end-to-end TTS system, the requirement on parameters is reduced, the speed of voice synthesis is improved, and the method is suitable for voice synthesis of small languages. 本发明公开一种基于ghost和iLPCnet的蒙古语语音合成方法，基于Bang预训练模型，对齐蒙古语音