A dockerized environment to run all kinds of ASR/transcription AI models.
You can run it via CLI, or as OpenAI Compatible Transcription API Server.
Supported local models:
- openai-whisper
- faster-whisper
- glm-asr-nano-2512
Supported remote providers:
- openai
- assemblyai
- lemonfoxai
more comming.
- Docker
- GPU for local models
- API Key for remote providers
touch .env
# override env vars in .env, e.g.:
# PORT=8000
# TRANSCRIBER_BACKEND=glm
make build
# run as api server
make server
./test_server.sh
# or run via cli
make shell
# you should be in container now
./transcribers.py --help
./transcribers.py data/audio.mp3 --backend lemonfoxai --format srt
For a complete speech-to-text experience, you can use these frontend applications that are compatible with this GLM-ASR server:
- NeuralWhisper - A modern web-based frontend for speech transcription with real-time capabilities
- WhisperSqueak - A lightweight desktop application for audio transcription
Both frontends are designed to work seamlessly with this GLM-ASR server's OpenAI-compatible API endpoints.
To use it in Spokenly:
URL: http://192.168.20.9:8000
API Key: (empty)
Model: (empty)
Uses the GLM-ASR-Nano-2512 model from the ZAI organization, which provides efficient speech recognition with minimal computational overhead.
The GLM-ASR project is developed by the ZAI team and represents state-of-the-art multimodal speech recognition capabilities.
- Input audio is resampled to 16kHz (optimal for the model)
- Supports up to 30-second chunks, automatically batched for longer audio
- Inference runs in bfloat16 precision for efficiency
This project builds upon the excellent work of:
- GLM-ASR - The underlying speech recognition model by the ZAI organization (zai-org/GLM-ASR-Nano-2512)
- faster-whisper-server - Inspired by Fedir Zadniprovskyi's architecture for OpenAI-compatible speech API servers
- FastAPI - For the excellent Python web framework
- HuggingFace - For the Transformers library and model hub
MIT License - See LICENSE file for details
Contributions are welcome. Please feel free to submit a pull request.
We especially welcome enhancements to the Dockerfile to make it smaller and more modern. If you have ideas for optimizing the Docker image (multi-stage builds, better layer caching, Alpine Linux compatibility, etc.), we'd love to see your contributions.
If you use GLM-ASR in your research, please cite the original GLM-ASR model from ZAI organization:
@misc{glm-asr,
title={GLM-ASR: Global Large-scale Multimodal Model for Automatic Speech Recognition},
author={ZAI Organization},
year={2024},
url={https://huggingface.co/zai-org}
}