xlsr_kurmanji_kurdish
| Property | Value |
|---|---|
| Base Model | facebook/wav2vec2-xls-r-300m |
| Training Data | Common Voice 7.0 - Kurmanji Kurdish |
| Best WER | 38.86% |
| Author | Akashpb13 |
| Framework | PyTorch 1.10.0 |
What is xlsr_kurmanji_kurdish?
xlsr_kurmanji_kurdish is a specialized speech recognition model fine-tuned specifically for the Kurmanji Kurdish language. Built upon Facebook's wav2vec2-xls-r-300m architecture, this model represents a significant step forward in Kurdish language processing technology. The model achieved a Word Error Rate (WER) of 38.86% through careful optimization and extensive training on curated Common Voice datasets.
Implementation Details
The model was developed using a comprehensive training approach, utilizing multiple datasets from Common Voice 7.0. The training process involved careful data curation, where only entries with more upvotes than downvotes were considered, and duplicates were removed. The training was conducted using a 90-10 split of the combined datasets, with sophisticated hyperparameter optimization.
- Learning rate: 0.000096 with cosine restart scheduling
- Batch size: 16 (train and eval)
- Training epochs: 100
- Gradient accumulation steps: 16
- Mixed precision training with Native AMP
Core Capabilities
- Specialized Kurdish speech recognition
- Robust performance across various speech patterns
- Optimized for production deployment
- Compatible with Common Voice dataset evaluation
Frequently Asked Questions
Q: What makes this model unique?
This model is specifically optimized for Kurmanji Kurdish speech recognition, utilizing a carefully curated dataset and achieving significant performance improvements through extensive training iterations, as evidenced by the progression from initial WER of 100% to final WER of 38.86%.
Q: What are the recommended use cases?
The model is ideal for Kurdish speech recognition tasks, particularly in applications requiring transcription of Kurmanji Kurdish speech. It's suitable for both academic research and practical applications, though users should be aware of the current WER limitations.
