Transliteration-based zero-shot domain adaptation for low-resource automatic speech recognition


Speech Samples


English to Cantonese

The original speech and transcription are in English, whereas the ASR model and transliteration are in Cantonese. In other words, the source language is English and the target langauge is Cantonese.


Original speech Synthesized speech (transcription) Synthesized speech (transliteration)
English: what have you done with them   English: what have you done with them   Cantonese: 或係會頓活站

Original speech Synthesized speech (transcription) Synthesized speech (transliteration)
English: good bye said the boy English: good bye said the boy Cantonese: 娟拜石背

Original speech Synthesized speech (transcription) Synthesized speech (transliteration)
English: you don't like him English: you don't like him Cantonese: 如準禮謙

Mandarin to Cantonese

The original speech and transcription are in Mandarin, whereas the ASR model and transliteration are in Cantonese. In other words, the source language is Mandarin and the target langauge is Cantonese.


Original speech Synthesized speech (transcription) Synthesized speech (transliteration)
Mandarin: 地势西高东低 Mandarin: 地势西高东低 Cantonese: 啲書詩交東啲

Original speech Synthesized speech (transcription) Synthesized speech (transliteration)
Mandarin: 其后共有五千人 Mandarin: 其后共有五千人 Cantonese: 市口供繞胡車人

Original speech Synthesized speech (transcription) Synthesized speech (transliteration)
Mandarin: 我也无法停止担心 Mandarin: 我也无法停止担心 Cantonese: 果夜壺化天住單線