TurjoDutta5555/whisper-small-0DB

🎤 Whisper Small Fine-tuned Model (0dB Noise) => High Noise

This model is a fine-tuned version of openai/whisper-small for speech recognition under noisy (10dB) conditions.

---

📌 Model Details

• Base Model: openai/whisper-small
• Noise Condition: 0dB
• Epochs: 10
• Total Steps: 790
• Best Checkpoint: checkpoint-700

---

🏆 Best Evaluation Results

The best model was selected at step 700:

• WER (Word Error Rate): 77.7023
• CER (Character Error Rate): 34.8762
• Eval Loss: 0.5518

---

📊 Training Summary

Metric	Value
Final Training Loss	~0.0673
Best Step	700
Total Steps	790
Epochs	10

---

📉 Evaluation History (Key Steps)

Step	WER	CER	Eval Loss
100	93.33	47.17	0.7525
200	83.26	37.40	0.4301
300	80.05	34.88	0.3891
400	78.75	34.78	0.4148
500	78.13	33.95	0.4685
600	78.69	34.83	0.5106
700	77.70	34.88	0.5518

---

⚙️ Training Hyperparameters

• Learning Rate: 1e-5
• Batch Size: 32
• Optimizer: AdamW (fused)
• LR Scheduler: Linear
• Warmup Steps: 100
• Epochs: 10
• Mixed Precision: AMP
• Gradient Accumulation: 2

---

🎯 Intended Use

• Speech-to-text transcription
• Noisy audio (0dB) ASR experiments (HIGH NOISE )
• Research / academic use

---

⚠️ Limitations

• Performance decreases on very noisy / unseen domains
• Not fully optimized for production use
• Dataset diversity affects generalization

---

🧠 Notes

• Best model selected from checkpoint-700
• Training continued till 790 steps but best performance was earlier
• Metrics may vary depending on dataset split