metadata
license: cc-by-4.0
language:
- ur
base_model:
- nvidia/parakeet-ctc-0.6b
pipeline_tag: automatic-speech-recognition
U-CTC
U-CTC is an Urdu automatic speech recognition (ASR) model based on the Parakeet-CTC-0.6B architecture. It has been fine-tuned on ~21 hours of Urdu speech data using the NVIDIA NeMo framework. The model is optimized for CTC-based transcription of spoken Urdu.
Model Summary
- Model Name: U-CTC
- Base Architecture: Parakeet-CTC-0.6B
- Framework: NVIDIA NeMo
- Language: Urdu
- Model Type: Conformer Encoder + CTC Decoder
- Loss Function: CTC Loss
- Hardware: Trained on NVIDIA RTX 3090
Training Configuration
| Setting | Value |
|---|---|
| Epochs | 69 |
| Max Steps | 14,800 |
| Optimizer | AdamW |
| Learning Rate | 0.001 |
| Betas | (0.9, 0.98) |
| Weight Decay | 0.001 |
| Scheduler | CosineAnnealing |
| Warmup Steps | 15,000 |
| Min LR | 0.0001 |
Dataset
The model was trained and evaluated on a manually curated Urdu speech dataset:
| Split | Files | Duration |
|---|---|---|
| Train | 9,425 | 10.87 h |
| Validation | 4,056 | 5.22 h |
| Test | 4,056 | 5.22 h |
- Total audio hours: ~21.3 hours
- Samples skipped due to CTC alignment failure: ~2.57%
- Average AM sequence length: 50.39
- Average target sequence length: 30.51
- AM-to-target length ratio: ~1.83
Performance
- Best Validation WER: 21.00%
Sample Predictions
| Reference Sentence | U-CTC Output |
|---|---|
| پاکستان اور زمبابوے کے درمیان ون ڈے سیریز جمعہ سے شروع ہوگی | پاکستان اسزموبے تنا ونڈے سیریز جما کے رو ہوگی |
| بی بی سی نے بہت دیر کردی یہ چیز دکھانے میں | بیسینہ بات اغیر کردی ی ھی اس سکھایں |
| ٹھنڈی ٹھنڈی ہوا | ٹندی ٹھنڈی ہوہا |
| ایک اینڈ تو ولیمسن سنبھالے ہوئے تھے | یہ ک اندوسا تو محالے ہوئے تھے |