Vaani · lab

Vaaniवाणी

Welcome to the lab.

Try, fine-tune, benchmark and publish small text and speech models for Indian languages, on your own machine.

Step 1

Your profile

Saved on this machine only, in runs/profile.json.

API · https://vaani-lab.centralindia.cloudapp.azure.com · checking…

Display nameHugging Face token (optional)Needed only for the Publish tab. If left empty, HF_TOKEN from the repo's .env is used.

Step 2

A project keeps its own chats, training runs, champions and benchmark results.

New project nameLetters, numbers, dot, dash or underscore. Start with a letter or number.

Every project has the same five tabs. Work left to right: try the models, train a better one, prove it is better, then ship it.

01
Chat
Talk to the text model in Hindi or English. It can call tools. Every turn is saved to the project.
02
Voice
Turn text into speech with any installed engine, or transcribe a clip to text.
03
Train
Fine-tune a text or speech model here or on Colab. Promote the best run to champion.
04
Eval
Benchmark speech engines on frozen test sets, run blind listening tests, score the text model.
05
Publish
Push a champion to a private Hugging Face repo.