This is an adaptation of Modal's Whisper transcription example [blog, github]. Key changes:
- eliminated podcast search; hardcoded for TWIML
- replaced react front end with sveltekit/tailwind
-
Clone the repo
-
Create a Modal Account
Visit modal.com and sign up.
- Install Modal CLI
pip install modal # Note: Requires Python 3.11.x
python3 -m modal setup # this will open a web page allowing you to create a token to authorize your client
- Build the Front End
cd into the frontend directory, and run:
npm install
npx vite build --watch
The last command will start a watcher process that will rebuild your static frontend files whenever you make changes to the frontend code.
- Tokens and secrets
To make your tokens accessible on the modal platform you must navigate to https://modal.com/{username}/secrets & select Create new secret of type Custom. Then add your secrets for the apis specified below - which you can get from the respective api providers (OpenAI, Hugging Face, Pinecone etc.).
Please ensure the key (set in step 2 & used in the code for referring to the local env-var), and name (set in step 3 and used by modal for referring to the secret) you enter in the modal UI are the same
1. key & name: 'HUGGING_FACE_ACCESS_TOKEN'
- from `https://huggingface.co/docs/hub/en/security-tokens`
- Deploy to Modal
Once you have vite build running, in a separate shell run this to start an ephemeral/dev app on Modal:
modal serve backend.main
The modal app will print two URLs in its shell window. Open the one that looks like this to access the front end: Created fastapi_app => https://sbc--whisper-pod-transcriber-fastapi-app-dev.modal.run
The other URL is the modal admin/back end, which looks like this: View app at https://modal.com/sbc/apps/ap-hkENjVd3K2PDbPtuwoWKND
Pressing Ctrl+C will stop your app.