Voicemail-ChatGPT is an IVR telephony interface for communicating with ChatGPT. It is implemented in Node.js and Twilio Programmable Voice.
You will need a Twilio account so that you can provide your account SID and auth token in .env. You will also need a Twilio phone number. Details on getting a free phone number are here.
Instructions for deploying a Node.js app and configuring Twilio appropriately are here. Set your Twilio phone number's voice webhook to [YOUR BASE URL]/twilio-webhook, and set the HTTP method to GET. Once the web server is running and the webhook is configured, you can use the service by calling your Twilio phone number from any phone.
You will need to supply the environment variables with a .env file in the root folder. Environment variables are indicated throughout the code with process.env.[VARIABLE_NAME].
If you are deploying the app on Space, once you have generated your default Spacefile, add the .env file's environment variables to your Spacefile as follows (assuming Spacefile is in the root folder):
node env-to-spacefile
Note that if the value of ABSOLUTE_URL in Spacefile is not already set to the Space domain, it will need to be changed accordingly once you have run env-to-spacefile. Storing your Space domain in another environment variable, such as SPACE_URL, may be helpful for making this change.
Spacefile will need to include the line public: true to allow Twilio's server to make a POST request to your app.
WARNING: running env-to-spacefile more than once will result in duplicate environmental variables being appended to Spacefile.
When the server is running, users can talk to ChatGPT as follows:
- User calls the Twilio phone number.
- ChatGPT responds with a greeting.
- If the user waits, ChatGPT will repeat the last question it asked in its greeting.
- User speaks a prompt (equivalent to typing a prompt into ChatGPT's text interface).
- When the user stops speaking for TWILIO_SPEECH_TIMEOUT_SECONDS, their prompt will be sent to ChatGPT.
- There will be a brief sound as ChatGPT processes the prompt (based on whatever sound is at WAIT_URL).
- ChatGPT will speak its response.
- User can press any number key to skip to the end of the message (in case they don't want to wait for the entire response before asking a follow-up prompt).
- If the user waits, ChatGPT will repeat the last question it asked in its response, or ask a generic question if the response did not include a question.
- Repeat steps 4-9, until user hangs up.
During each call, the conversation history is re-submitted to ChatGPT each time the user speaks their next prompt, so a single phone call is equivalent to a single "New Chat".