Skip to content

README.md

Latest commit

 

History

History
72 lines (59 loc) · 1.99 KB

README.md

File metadata and controls

72 lines (59 loc) · 1.99 KB

FireRedASR Microservice

A FastAPI-based microservice for speech-to-text transcription using FireRedASR.

Setup

  1. Download the FireRedASR model:

    cd server
mkdir -p models
# Download FireRedASR-AED-L & PUNC-BERT model and place it in models/
git clone https://huggingface.co/FireRedTeam/FireRedChat-punc models/PUNC-BERT
git clone https://huggingface.co/hfl/chinese-lert-base models/PUNC-BERT/chinese-lert-base
pushd models/PUNC-BERT && git lfs pull && popd
git clone https://huggingface.co/FireRedTeam/FireRedASR-AED-L models/FireRedASR-AED-L
pushd models/FireRedASR-AED-L && git lfs pull && popd

  2. Build the Docker image:

    # 1. Modify Dockerfile based on your CUDA version
#    for Ampere and newer GPUs
#       FROM pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime
#    for Volta GPUs (3080, V100)
#       FROM pytorch/pytorch:2.5.1-cuda11.8-cudnn9-runtime
# 2. Use tencentyun pypi mirror (in mainland China)
#       RUN pip install --no-cache-dir -r requirements.txt --index-url http://mirrors.tencentyun.com/pypi/simple/ --trusted-host mirrors.tencentyun.com
docker build -t fireredasr-service .

  3. Run the container:

    docker run -d \
  -p 8000:8000 \
  --gpus '"cuda:0"' \
  -v $(pwd)/models:/app/models \
  fireredasr-service

API Usage

Transcribe Audio

Endpoint: POST /audio/transcriptions

Parameters:

  • file: Audio file (multipart/form-data)
  • media_type: Response format (application/json or text/plain)

Example using curl:

curl -X POST \
  http://localhost:8000/audio/transciptions \
  -H "Content-Type: multipart/form-data" \
  -F file="@audio.wav"

Response:

{
  "sentences": [{
    "confidence": 0.8,
    "text": "transcribed text here"
    }],
  "wav_file": "audio.wav"
}

Environment Variables

  • FIREREDASR_PATH: Path to FireRedASR installation (default: /app/fireredasr)
  • MODEL_DIR: Path to model directory (default: /app/models)