DelosLogic/accento-v2.0

Accento V2.0 - Trinidadian Creole English ASR

Accento V2.0 is a fine-tuned Whisper Large V3 Turbo model optimized for Trinidadian Creole English.

Performance

• WER: 19.94% (with beam_size=3)
• CER: 10.00%
• 54% better than base Whisper
• 27% better than Accento V1.0

Usage

from accento import AccentoTranscriber

# Auto-downloads from Hugging Face if not found locally
transcriber = AccentoTranscriber(model_path="models/accento-v2.0")
result = transcriber.transcribe("audio.wav")
print(result.text)

Technical Details

• Base: Whisper Large V3 Turbo (809M params)
• Method: LoRA (rank=32, alpha=64)
• Adapters: ~106M parameters
• Training: 179 labeled samples + iterative training + model soups

License

MIT License