minutely-notes

AI-powered meeting transcription project for MinutelyAI.

This repository includes two browser-based experiences:

• meeting.html: live, multi-user transcription with shared real-time output
• minutely.html: upload a recorded audio/video file and get diarized transcription (speaker-labeled)

What It Uses

• Deepgram WebSocket API for live speech-to-text
• Firebase Realtime Database for syncing live transcripts across participants
• Modal GPU endpoint for offline transcription
• Whisper medium + pyannote speaker diarization on the backend

Project Files

• meeting.html: live meeting page
• minutely.html: recorded file transcription page
• modal_app.py: Modal backend (FastAPI + Whisper + pyannote)
• config.example.js: frontend config template
• requirements.txt: backend Python dependencies

Prerequisites

You need:

• A Deepgram API key
• A Firebase project (Realtime Database enabled)
• A Hugging Face token with access to pyannote models
• A Modal account and CLI
• Python 3.11+ for local backend setup/deploy

Quick Start

1. Configure Frontend Keys

Create config.js from the template:

cp config.example.js config.js

Then update:

• DEEPGRAM_KEY
• MODAL_URL (after deploying backend)
• FIREBASE object values

Important: config.js is gitignored and should never be committed.

2. Run Live Meeting Page

Open meeting.html in Chrome.

Flow:

1. Enter your name and create/join a meeting
2. Share the invite link
3. Each participant starts microphone capture
4. Transcript updates in real time for everyone

3. Run Recorded Transcription Page

Open minutely.html in a browser, upload one file, and click transcribe.

Supported formats include .mp3, .wav, .mp4, .m4a, .webm.

Deploy Backend to Modal

Install dependencies and deploy:

pip install -r requirements.txt
modal secret create minutely-secrets HF_TOKEN=your_hf_token
modal deploy modal_app.py

Take the generated endpoint and set it as MODAL_URL in config.js.

Backend API Contract

POST / expects JSON payload:

{
	"audio": "<base64-audio-bytes>",
	"language": "en",
	"min_speakers": 1,
	"max_speakers": 8
}

Response includes:

• full transcript text
• detected language
• speaker segments with timestamps

Troubleshooting

• If mic capture fails, verify browser mic permissions and use HTTPS when required.
• If live transcript is empty, confirm DEEPGRAM_KEY and Firebase config are valid.
• If recorded transcription fails, verify MODAL_URL and that Modal app is deployed.
• If diarization fails, verify HF_TOKEN is set in minutely-secrets.

Security Notes

• Never commit config.js.
• Rotate any API key that was accidentally exposed.
• Restrict Firebase rules appropriately for production.