A server for Thai speech recognition using the Whisper-TH-Medium-Combined model from BioDat Lab. This server provides high-quality Thai speech-to-text transcription capabilities through a simple API built with FastAPI. The model supports both Thai and English, with optimization for Thai language audio transcription.
- Thai-optimized speech recognition using the Whisper-TH model
- Simple API for transcribing audio files
- Automatic model downloading from Hugging Face Hub
- Supports both CPU and GPU inference
- Cross-Origin Resource Sharing (CORS) enabled
- Batch processing for efficient transcription of longer audio files
git clone https://github.com/your-username/thonburian-stt.git
cd thonburian-sttpython -m venv venv
source venv/bin/activate # On Windows, use `venv\Scripts\activate`pip install -r requirements.txtpython server.pyThe server will start at http://0.0.0.0:8081.
Endpoint: GET /
Response:
{
"status": "Thonburian STT Server is running"
}Endpoint: POST /transcribe
Parameters:
file(file, required): The audio file to transcribelanguage(string, optional, default:th): The language code (e.g.,thfor Thai)task(string, optional, default:transcribe): The task type
Example Request:
curl -X POST "http://localhost:8081/transcribe" \
-H "accept: application/json" \
-H "Content-Type: multipart/form-data" \
-F "file=@your_audio_file.mp3" \
-F "language=th" \
-F "task=transcribe"Example Response:
{
"text": "สวัสดีครับ นี่คือการทดสอบระบบรู้จำเสียงภาษาไทย",
"processing_time": 2.34
}Endpoint: POST /stt
Parameters:
file(file, required): The audio file to transcribe- Request body (optional):
{ "language": "th", "task": "transcribe" }
If a CUDA-compatible GPU is available, the model will automatically use it for inference. Otherwise, it defaults to CPU.
device = 0 if torch.cuda.is_available() else "cpu"The required model files are automatically downloaded from Hugging Face and cached in the model_cache directory.
Logging is enabled by default and provides real-time updates on:
- Model loading status
- Transcription requests
- Processing time
- Errors and exceptions
The server uses the Whisper-TH model with the following optimizations:
- Chunk-based processing for handling longer audio files
- Batch processing for efficient inference
- Automatic device selection (GPU/CPU)
The server can process various audio formats, including:
- WAV
- MP3
- OGG
- FLAC
- M4A
- Add support for streaming audio input
- Implement custom vocabulary and language models
- Add speaker diarization capabilities
- Implement confidence scores for transcriptions