DaCiDian is an open-sourced lexicon for Chinese Automatic Speech Recognition(ASR)
In mainstream ASR system, lexicon is a core component, that maps word into acoustic modeling units(such as phone). In DaCiDian, we break the mapping into 2 independent layers:
word --> PinYin syllable --> phoneme
The purpose of this design is as follows:
-
Anyone who is familiar with PinYin (basically every mandarin speaker), can enrich DaCiDian's vocabulary, by adding new entry(word) into the layer-1 mapping.
-
ASR system developers can easily adapt DaCiDian to their own phone set by defining their own layer-2 mapping.
- word and its pronunciation(s) are seperated by tab
- multiple pronunciations are seperated by ;
- each pronunciation is a sequence of Chinese PinYin syllables, seperated by space
- tone infomation are encoded into 0,1,2,3,4
examples:
...
裤子 KU_4 ZI_0
好事 HAO_4 SHI_4;HAO_3 SHI_4
教授 JIAO_1 SHOU_4;JIAO_4 SHOU_4
...
语音识别 YU_3 YIN_1 SHI_2 BIE_2
傅里叶变换 FU_4 LI_3 YE_4 BIAN_4 HUAN_4
pinyin_to_phone is a user-defined mapping from PinYin syllables to target phone set
Take traditional PinYin's Initial-Final structure for example, a mapping should be defined as follows:
A $0 a
AI $0 ai
AN $0 an
ANG $0 ang
AO $0 ao
BA b a
BAI b ai
BAN b an
BANG b ang
BAO b ao
...
...
...
ZONG z ong
ZOU z ou
ZU z u
ZUAN z uan
ZUI z ui
ZUN z un
ZUO z uo
- syllable and its phoneme representation are seperated by tab
- $0 is the NULL-Initial phone.
- English word in Chinese context is a problem. An upgraded version covering both English and Chinese can be found here BigCiDian