from faster_whisper import WhisperModel

# Load model
model = WhisperModel("Tnaot/whisper-large-v3-khmer-ct2", device="cuda", compute_type="int8_float16")

# Transcribe audio
segments, info = model.transcribe(
    "audio.wav",
    language="km",
    task="transcribe",
    beam_size=5,
    vad_filter=True
)

# Print results
for segment in segments:
    print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")

CLI Usage

faster-whisper audio.wav --model Tnaot/whisper-large-v3-khmer-ct2 --language km --task transcribe

Performance

Speed: 3-5x faster than original PyTorch model
Memory: ~50% less GPU memory usage
Quality: Same accuracy as the original model

Original Model

Fine-tuned model: Tnaot/whisper-large-v3-khmer-5hr
Base model: openai/whisper-large-v3

License

Apache 2.0 (same as OpenAI Whisper)

Citation

bibtex

@misc{whisper-large-v3-khmer-ct2,
  title={Whisper Large V3 Khmer - CTranslate2},
  author={Converted from Tnaot/whisper-large-v3-khmer-5hr},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/Tnaot/whisper-large-v3-khmer-ct2}}
}