vdspeak's speech synthesis system uses the WaveNet deep learning architecture to break through the mechanical feel of traditional splicing synthesis technology. Its acoustic model has undergone 2,000 hours/language of targeted training, and is able to capture the pronunciation rules and intonation characteristics of specific languages. Actual tests show that the naturalness MOS score of English to Chinese dubbing reaches 4.2 out of 5, especially when dealing with specialized terminology with an accuracy rate superior to common market solutions.
At the level of technical realization, the system will first perform voiceprint separation of the original video to eliminate background noise interference; then maintain the similarity between the dubbed timbre and the original speaker through cross-language speech cloning technology. For complex morphological languages such as Spanish, an attention mechanism is also integrated to deal with verb conjugation. User cases show that after an educational institution used its German dubbing function, the viewing time in Germany increased by 3 times.
This answer comes from the articlevdspeakThe
English 
简体中文 
日本語 
Deutsch 
Português do Brasil