GPT-3 Interactive
This work is based on AI Dungeon CLI.
Speech-enabled Interactive GPT-3 CLI is integrated with Google Speech Recognition, Google Translation and Apple Text-To-Speech.
Examples
The system records spoken Korean language and translates into English using Google Translation. GPT-3 generates responses and the system is translating the conversation back into Korean language and speaking in verbal Korean language using OS/X "Say". You can find out how smooth the conversation flows.
[English]
Q: I had presbyopia in my early 40s. What should I do?
A: In the old days, you would simply have got reading glasses for slightly closer work such as reading the newspaper.
Q: Since I'm in my early 40s, I think people will think it's strange to use a magnifying glass. Is there any other alternative?
A: If you can affort to, then you should see an eye doctor to see if anything can be done for your presbyopia.
Q: What methods do ophthalmologists have to treat presbyopia?
A: You may be able to benefit from new technologies such as monovision lens implants or small incision lenticule extraction (LASE) and no surgery is required.
Q: Wouldn't such a procedure be expensive?
A: That's all about how much money you have, if you can't affort that, you might want to consider wearing reading glasses or bifocals.
[Korean]
Q : 40 대 초반에 노안이있었습니다. 어떻게해야합니까?
A : 예전에는 신문을 읽는 것과 같이 좀 더 가까운 작업을 위해 돋보기를 가지고 있었을 것입니다.
Q : 40 대 초반이라 사람들이 돋보기를 사용하는 것이 이상하다고 생각할 것 같아요. 다른 대안이 있습니까?
A : 노력할 수 있다면 안과 의사를 만나 노안을 치료할 수 있는지 확인해야합니다.
Q : 안과 의사는 노안을 치료하기 위해 어떤 방법을 사용해야합니까?
A : Monovision 렌즈 임플란트 또는 LASE (small incision lenticule extract)와 같은 새로운 기술의 혜택을받을 수 있으며 수술이 필요하지 않습니다.
Q : 그러한 절차는 비싸지 않습니까?
A : 그것이 얼마나 많은 돈이 있는지에 관한 것입니다. 그렇게 노력할 수 없다면 독서 용 안경이나 이중 초점을 고려하는 것이 좋습니다.
Installation
$ python3 -m pip install -e .
Running
In any case, you first need to create a configuration file.
$ ai-dungeon-cli --auth-token [your auth token] --scene scenes/qa --locale ko-KR --voice Yuna
Specified voice font should be available with OS/X Say command.
Commands
[text]
Default command, nothing special. It will append [text] to the end of the context and request GPT-3 to generate further more.
/qa [question]
To use in interactive way, you can use /qa command. If the question is not specified, it will be caught using record and speech API.
/r[rrr...]
If you don't like the progression of GPT-3, you can revert using /r command. /r command tries to revert as many blocks as you ask.
/s[sss...]
If you want a synthesized voice output for the progress, /s command will read loudly the last n blocks using OS/X "Say" command.
Configuration (optional)
Several things can be tuned by resorting to a config file.
Create a file config.yml either:
- in the same folder in your home folder:
$HOME/.config/ai-dungeon-cli/config.yml - in the same folder as the sources:
./ai-dungeon-cli/ai_dungeon_cli/config.yml
Authentication
Sniff a Authentication Token and use it directly:
auth_token: '<MY-AUTH-TOKEN>'To get this token, you need to first login in a web browser to play.aidungeon.io.
Then you can find the token either in your browser localStorage or in the content of the connection_init message of the websocket communication (first sent message).
Either way, developer tools (F12) is your friend.
Prompt
The default user prompt is '> '.
You can customize it with e.g. :
prompt: 'me: 'Dependencies
Please have a look at requirements.txt.
Limitations and future improvements
Right now, the code is over-optimistic: we don't catch cleanly when the backend is down.
Support
As you might have heard, hosting AI Dungeon costs a lot of money.
This cli client relies on the same infrastructure as the online version (play.aidungeon.io).
So don't hesitate to help support the hosting fees to keep the game up and running.
Author
Nako Sung @nakosung.
Contributors & acknowledgements
- scenes/qa from google search results
- scenes/play from https://monologueblogger.com/policy/
- TBU