nicklonlee / speech-to-text-to-speech

A fully fledged conversation bot built on top of OpenAI's GPT-3. Insert your API key, and have a nice conversation with an AI! (Careful, can be costly if you let it run wildly)

IMPORTANT NOTE!

This script will not run without an API key from openai.com

About the Script

I use three modules, openai, speech_recognition and pyttsx3. Together they work to take in text and input it into whatever GPT-3 set up you’ve got going on. I primarily use it to chat.I’m sharing this because I really don’t see all that many ‘basic’ scripts using the openai library and it was kind of annoyingly obtuse to get this running. I spent around a day just to get this working, which added up to a lot of questions to a poor and confused API that was sick of me by the time I was done.(Seriously, that was really cool. GPT-3 got bored of me asking the same question on completely different prompts.) Learned quite a decent amount, not a big python guy, but I think the script runs well and takes tweaks fluently.

Notes

Running the script

Installation

Install the required libraries through pip, by running 'pip install -r requirements.txt' the file is in the repo

Running the script

Run a terminal and navigate to the repo file and run 'python Speech-to-text-to-Speech.py' It will ask you to speak. Do so and wait a bit for it to process. Boom. RESULTS!

Variables

The only Variable you're probably interested in should be 'myPrompt' and the '{SpeechText}' string inside of it. You can change the prompt, but definitely look at how I laid it out. The API reacts a bit differently here and I think that's due to a lack of punctuation and on top of that, some issues with new lines. So the best way I found for this to work was the way I got it laid out, aka 'Human:{SpeechText} \nAI:' That way GPT-3 just picks it up from the point of interest.

Some things I’d like to develop further:

1.) The speech to text system is not great at taking in commas and periods. It really misses a lot of the nuance that this AI seems to thrive on. So any recommendations on libraries are welcome.

2.) Hopefully I find a more calming voice than Microsoft Sam (Despite his nostalgic value). If anyone recommends some solution for that, it would be awesome.

3.) Some visuals would be nice, maybe see if I can make API calls with the AI results and get a sentiment grading on it and represent that in the graphic (frowny face/smile etc. etc.)

Feel free to share any tips and to fork/use this freely.