Finetuned openai/whisper-small on 123 telugu training audio samples from AkCodes23/sarvam-tts-in-te-en.
This model was created from the Mozilla.ai Blueprint: speech-to-text-finetune.
Evaluation results on 8 audio samples of telugu:
Baseline model (before finetuning) on telugu
- Word Error Rate (Normalized): 109.73
- Word Error Rate (Orthographic): 124.413
- Character Error Rate (Normalized): 72.428
- Character Error Rate (Orthographic): 98.127
- Loss: 1.815
Finetuned model (after finetuning) on telugu
- Word Error Rate (Normalized): 66.126
- Word Error Rate (Orthographic): 90.141
- Character Error Rate (Normalized): 51.045
- Character Error Rate (Orthographic): 59.465
- Loss: 0.597