lintoai/linto-stt-kaldi repository overview

⁠LinTO-STT-Kaldi

ASR API based on Kaldi⁠ and Vosk⁠. Supports offline and real-time transcription.

⁠Quick Start

⁠Prerequisites

• Docker⁠
• At least 7GB disk space for the image
• Up to 7GB RAM depending on model

⁠Models

LinTO-STT-Kaldi accepts two kinds of models:

LinTO models (acoustic + language model separately):

• Download from dl.linto.ai⁠
• Set MODEL_TYPE=lin
• Mount volumes: -v <AM_PATH>:/opt/AM -v <LM_PATH>:/opt/LM

Vosk models (all-in-one):

• Download from alphacephei.com/vosk/models⁠
• Set MODEL_TYPE=vosk
• Mount volume: -v <MODEL_PATH>:/opt/model

⁠Pull or Build

docker pull lintoai/linto-stt-kaldi

or

docker build -t linto-stt-kaldi:latest --build-arg STT_ENGINE=kaldi .

⁠Run

With LinTO models:

docker run --rm -p 8080:80 \
  -e SERVICE_MODE=http \
  -e MODEL_TYPE=lin \
  -v /path/to/AM:/opt/AM \
  -v /path/to/LM:/opt/LM \
  lintoai/linto-stt-kaldi

With a Vosk model:

docker run --rm -p 8080:80 \
  -e SERVICE_MODE=http \
  -e MODEL_TYPE=vosk \
  -v /path/to/vosk-model:/opt/model \
  lintoai/linto-stt-kaldi

For streaming, use -e SERVICE_MODE=websocket. For Celery task mode, use -e SERVICE_MODE=task with -v /shared/audio:/opt/audio.

For punctuation recovery, add a recasepunc model⁠:

-v /path/to/fr.24000:/opt/PUNCT -e PUNCTUATION_MODEL=/opt/PUNCT

⁠License

AGPLv3 (see LICENSE).

⁠Acknowledgments

• Vosk, speech recognition toolkit⁠
• Kaldi Speech Recognition Toolkit⁠