IndicSTT

IndicSTT is a FastAPI web service and CLI for Punjabi/Indic speech-to-text using a local NeMo .nemo model.

For a reproducible installation on another Linux computer, see docs/install.md.

Repository: https://github.com/sanbroz/IndicSTT

Quick Start

# From the project root
conda env create -f conda_indicstt_nobuilds.yml
conda activate IndicSTT

# Put the model here before starting the app:
# models/indicconformer_stt_pa_hybrid_rnnt_large.nemo

bash indicstt.sh

The web app listens on http://0.0.0.0:8010.

Important Runtime Notes

• ffmpeg must be installed on the operating system.
• The model file is expected at models/indicconformer_stt_pa_hybrid_rnnt_large.nemo.
• Real data/users.toml and data/api_keys.toml files are local runtime state; use the .example.toml files as templates.
• Set INDICSTT_SECRET_KEY to a long random value before deploying outside local development.
• With the pinned NeMo runtime, this archived multilingual model falls back from RNNT to CTC automatically.
• Transcription output is normalized to remove SentencePiece ▁ markers.