SenseVoice-Small GGUF
A multilingual ASR model from Alibaba's Tongyi Lab, converted to GGUF format for broad compatibility.
Model Overview
SenseVoice-Small features:
- Parameters: Small variant
- Language Support: Multiple languages including Chinese, English, Cantonese, and more
- Quantization Format: GGUF (llama.cpp compatible)
- Architecture: FunASR-based speech recognition
- Features: High-accuracy recognition, real-time transcription, noise robustness
- Use Cases: Speech-to-text, subtitle generation, meeting transcription, multilingual transcription
Available Quantized Versions
| Filename | Quantization Type | File Size | Description |
|---|---|---|---|
| sensevoice-small-f16.gguf | F16 | ~449 MB | Full precision, highest accuracy |
| sensevoice-small-q5_k_m.gguf | Q5_K_M | ~164 MB | Balanced accuracy and speed |
| sensevoice-small-q8_0.gguf | Q8_0 | ~241 MB | High-accuracy quantization |
Usage
llama.cpp Command Line
./main -m sensevoice-small-f16.gguf -f audio.wav --output-txt
Python (llama-cpp-python)
from llama_cpp import Llama
# Load model
llm = Llama(
model_path="sensevoice-small-f16.gguf",
n_gpu_layers=-1, # Use GPU acceleration
)
# Perform speech recognition
result = llm.transcribe("audio.wav")
print(result["text"])
Performance Features
- High multilingual recognition accuracy
- Excellent Chinese recognition, supports Mandarin and Cantonese
- Noise resistance capabilities
- Supports real-time streaming transcription
Technical Specifications
- Architecture: Transformer-based encoder-decoder structure
- Sample Rate: 16kHz
- Input Format: WAV/PCM audio
- Output Format: Text
License
Apache 2.0 License
Acknowledgments
- Original Model: iic/SenseVoiceSmall
- Source: Alibaba DAMO Academy ModelScope