File	Size
added_tokens.json	0.0 MB
tokenizer_config.json	0.3 MB
special_tokens_map.json	0.0 MB
normalizer.json	0.1 MB
preprocessor_config.json	0.0 MB
config.json	0.0 MB
vocab.json	1.0 MB
vocabulary.json	1.0 MB
model.bin	2944.3 MB
merges.txt	0.5 MB

🔧 Installation & Usage

Installation

pip install faster-whisper

Basic Usage

python

from faster_whisper import WhisperModel

# Load the model
model = WhisperModel("2snem6/fastr-whisper-large-v3-edacc-commonvoice-l2arctic-v3")

# Transcribe audio
segments, info = model.transcribe("audio.wav")

print(f"Detected language: {info.language} (probability: {info.language_probability:.2f})")

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

Advanced Usage

python

# With custom parameters
segments, info = model.transcribe(
    "audio.wav",
    beam_size=5,
    language="en",  # Force English
    condition_on_previous_text=False,
    temperature=0.0
)

# Batch processing
audio_files = ["file1.wav", "file2.wav", "file3.wav"]
for audio_file in audio_files:
    segments, info = model.transcribe(audio_file)
    # Process segments...

Loading from Local Path

python

# If you've downloaded the model locally
model = WhisperModel("/path/to/downloaded/model")

⚡ Performance Comparison

FastWhisper (CTranslate2) vs Standard Transformers:

Metric	Standard Transformers	FastWhisper (CT2)	Improvement
Speed	1x	2-4x	2-4x faster
Memory	1x	0.5-0.8x	20-50% less
Model Size	1x	0.5-0.8x	20-50% smaller

Performance may vary depending on hardware and audio length.

🎯 Use Cases

This optimized model is perfect for:

Real-time transcription applications
Production deployments requiring fast inference
Resource-constrained environments
Batch processing of audio files
API services with high throughput requirements

📝 Technical Details

Conversion Process

This model was converted using the ct2-transformers-converter tool:

bash

ct2-transformers-converter \
    --model 2snem6/whisper-large-v3-edacc-commonvoice-l2arctic-v3 \
    --output_dir fastr-whisper-large-v3-edacc-commonvoice-l2arctic-v3 \
    --quantization float16 \
    --copy_files tokenizer.json preprocessor_config.json

Quantization

FLOAT16: Half-precision floating point for optimal speed/quality balance

🔄 Original Model

This is a converted version of a fine-tuned Whisper model. The original model 2snem6/whisper-large-v3-edacc-commonvoice-l2arctic-v3 was likely fine-tuned for specific:

Accents or dialects
Domain-specific vocabulary
Improved accuracy on certain audio types

Please refer to the original model card for training details and performance metrics.

📚 Citation

If you use this converted model, please cite both the original Whisper paper and the CTranslate2 library:

Original Whisper

bibtex

@misc{radford2022whisper,
  title={Robust Speech Recognition via Large-Scale Weak Supervision},
  author={Alec Radford and Jong Wook Kim and Tao Xu and Greg Brockman and Christine McLeavey and Ilya Sutskever},
  year={2022},
  eprint={2212.04356},
  archivePrefix={arXiv},
  primaryClass={eess.AS}
}

CTranslate2

bibtex

@misc{ctranslate2,
  title={CTranslate2: Fast inference with Transformers and OpenNMT models},
  author={Guillaume Klein},
  year={2020},
  url={https://github.com/OpenNMT/CTranslate2}
}

🤝 Contributing

Found an issue or want to improve this model?

Original model issues: Report to 2snem6/whisper-large-v3-edacc-commonvoice-l2arctic-v3
Conversion issues: Open an issue with details about the conversion process
Performance issues: Check the faster-whisper documentation

📄 License

This model inherits the license from the original model: Apache 2.0

Converted with ❤️ using CTranslate2 and faster-whisper