ML
hindcribe
Offline Hindi speech-to-text and Hindi→English translation — no API keys, no cloud, built for a government grievances hackathon.
Public-grievance audio comes in Hindi, and the systems that need to read it want English text — but you can’t ship citizens’ recordings off to a third-party cloud API. hindcribe does the whole pipeline on-device: upload a Hindi audio clip and get back both a Hindi transcript and an English translation, with nothing ever leaving the machine.
What it does
- Upload → transcript → translation — drop an MP3, get Hindi text (Vosk) and an English rendering (Argos Translate) back.
- Fully offline, open-source stack — no OpenAI, no Google Cloud, no keys. Everything runs locally, which is the point for anything touching citizen data.
- Normalizes messy input — FFmpeg converts whatever audio format shows up into the WAV shape the recognizer expects before it ever hits the model.
How it fits together
It’s a small, honest pipeline where each stage does one thing: FFmpeg for audio prep, Vosk for offline ASR, Argos (vendored, so it works with no network) for translation, Flask to glue it into a web UI. Each utility has its own unit tests — for a hackathon build, having the transcription and translation steps individually verifiable saved a lot of “why is the output garbage” debugging.
Context
Built for DARPG 2024, the government’s public-grievances hackathon — where “don’t send the data to a cloud you don’t control” isn’t a nice-to-have, it’s the whole constraint.
The interesting part wasn’t the models — it was proving you can do useful Hindi ASR and translation with an entirely offline, auditable stack.
Want something like this built?
I take on freelance backend and API work — auth, payments, media, and the security review to go with it.
Start a conversation →