ZLSCompLing/Whisper-Large-V3-Turbo-Luxembourgish-fp16

Whisper Large V3 Turbo - Luxembourgish (FP16)

An FP16 optimized version of a fine-tuned OpenAI's Whisper Large V3 Turbo for Luxembourgish speech recognition.

Model Description

This model was fine-tuned on approximately 75 hours of Luxembourgish speech data, curated to be representative of diverse speaking contexts within the available data landscape for the language. This FP16 version provides 50% smaller file size while maintaining transcription quality.

The model is provided in CTranslate2 format for efficient inference with faster-whisper and compatible libraries.

Model Details

• Base Model: openai/whisper-large-v3-turbo
• Language: Luxembourgish (lb)
• Format: CTranslate2 (FP16 precision)
• Model Size: 1.6GB (50% smaller than FP32)
• License: MIT

Intended Uses

• General automatic speech recognition (ASR) for Luxembourgish
• Audio and video transcription
• Speaker diarization pipelines (when combined with PyAnnote)

Usage

With faster-whisper

from faster_whisper import WhisperModel

model = WhisperModel("path/to/ct2_model_fp16", device="cuda")

segments, info = model.transcribe("audio.wav", language="lb")
for segment in segments:
    print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")

With WhisperX (for word-level timestamps and speaker diarization)

import whisperx

model = whisperx.load_model("path/to/ct2_model_fp16", device="cuda")
audio = whisperx.load_audio("audio.wav")
result = model.transcribe(audio, language="lb")

Technical Notes

When using this model with WhisperX and PyAnnote for speaker diarization, be aware of potential dependency conflicts between PyTorch, WhisperX, and PyAnnote. Depending on your setup, you may need both modern CUDA libraries and cuDNN 8 libraries installed.

This FP16 version provides significant storage and memory benefits while maintaining transcription quality. On older GPUs that don't support efficient FP16 computation, the model will automatically fall back to FP32 compute.

For reference, see the Sproochmaschinn project, which uses this model architecture in production.

Citation

If you use this model, please cite:

@misc{zls2025whisperlbfp16,
  title={Whisper Large V3 Turbo - Luxembourgish (FP16)},
  author={Zenter fir d'Lëtzebuerger Sprooch},
  year={2025},
  publisher={Hugging Face},
  note={FP16 optimized version for efficient inference}
}

Acknowledgments

Developed by Zenter fir d'Lëtzebuerger Sprooch.

This model powers Sproochmaschinn, a Luxembourgish speech processing platform.