Wakeup words support for Chinese

respeaker / respeaker_python_library

To build voice enabled objects/applications with Python and ReSpeaker

Apache License 2.0

149 stars 76 forks source link

Wakeup words support for Chinese #1

Open juncaofish opened 7 years ago

juncaofish commented 7 years ago

Hello, Just need a local support feature for voice wakeup~ I notice this file respeaker/pocketsphinx-data/dictionary.txt . So my question is if I want to set a customised wakeup word in Chinese, like "魔镜", can I add to the dictionary something like 'mojing' M UE J IN ? Thanks.

xiongyihui commented 7 years ago

To recognize Chinese words, you need another acoustic model, for example https://sourceforge.net/projects/cmusphinx/files/Acoustic%20and%20Language%20Models/Mandarin/.

But the mandarin acoustic model is not good enough. You may need some training for the model.

juncaofish commented 7 years ago

Thanks for replying. So is MT7688 capable enough to run the mandarin acoustic model?

xiongyihui commented 7 years ago

The mandarin acoustic model is too big for MT7688 to do continuous speech recognizer. You need a small one.

juncaofish commented 7 years ago

I find a tool https://github.com/Kitt-AI/snowboy which seems to be a perfect solution for Chinese hotword local detection. However, it seems it cannot be targeted to MT7688 either for the hardware architecture limit. Do you have any suggestions on how to take less effort for Chinese hotword detection task? Thanks very much.

xiongyihui commented 7 years ago

It seems the snowboy is the most accurate. Ask for kitt-ai to compile a MIPS version of snowboy maybe work. There are old reversions of mandrarin acoustic model in cmusphinx repo which is smaller and should be able to run on MT7688, or train a smaller model by hand.

juncaofish commented 7 years ago

I'm afraid the MT7688 is not a supported target for snowboy. Take a reference at discussion here: https://groups.google.com/a/kitt.ai/forum/#!topic/snowboy-discussion/iL4rRZRUk4U

juncaofish commented 7 years ago

is it possible to use Xunfei wakeup on Respeaker, I mean compile the sdk for 7688's mips platform? It seems like a good solution if cost is not considered as a cons.

xiongyihui commented 7 years ago

I applied a sdk of xunfei's stt and stt running on 7688, it worded. The wakeup sdk should work too.

juncaofish commented 7 years ago

Since I have tried the online sdk on Raspberry Pi, I guess you mean the online tts/stt service of xunfei. Do you need to submit the compile chain to their engineer? I remember the sdk only provides .so for x86/x64 .

xiongyihui commented 7 years ago

Yes, for c sdk for 7688, we need to submit the toolchain.

calvinwongsc commented 7 years ago

I am intersted in using xunfei's wakeup sdk.... i have download the pack from there, and how can i use it in respeaker?

jiaoxu commented 7 years ago

yihui 你好我在mac上执行bing_translator.py的时候，提示DEBUG:urllib3.connectionpool:https://oxford-speech.cloudapp.net:443 "POST /token/issueToken HTTP/1.1" 401 143 00000000000recognition connection failed，我已经提供了申请的密钥 1: ，这个错误是什么问题，国内网络限制吗？

solidssss commented 3 years ago

Is there any way to actually run snowboy detect on 7688, if so how? Thanks