Whisper Large-v2 Hausa
| Property | Value |
|---|---|
| License | Apache 2.0 |
| Dataset | Mozilla Common Voice 11.0 |
| WER Score | 37.41% |
| Training Steps | 1000 |
What is whisper-large-v2-ha?
Whisper Large-v2 Hausa is a specialized automatic speech recognition (ASR) model fine-tuned from OpenAI's Whisper Large-v2 architecture specifically for the Hausa language. This model represents a significant step forward in African language processing, achieving a Word Error Rate (WER) of 37.41% on the Mozilla Common Voice 11.0 Hausa dataset.
Implementation Details
The model utilizes a transformer-based architecture and was trained using PyTorch with specific hyperparameters including a learning rate of 1e-05, batch sizes of 32 for training and 16 for evaluation, and implements a linear learning rate scheduler with 100 warmup steps. The training process was conducted using multi-GPU infrastructure and optimized using Adam optimizer.
- Training conducted over 1000 steps with 12.06 epochs
- Implements linear learning rate scheduling
- Utilizes Transformers 4.26.0 and PyTorch 1.13.0
- Multi-GPU distributed training support
Core Capabilities
- Specialized Hausa language speech recognition
- Batch processing capability for both training and inference
- Integration with TensorBoard for monitoring
- Supports inference endpoints for deployment
Frequently Asked Questions
Q: What makes this model unique?
This model is specifically optimized for Hausa language speech recognition, filling a crucial gap in African language ASR technology. Its relatively strong WER of 37.41% makes it particularly useful for Hausa speech processing tasks.
Q: What are the recommended use cases?
The model is ideal for Hausa speech transcription tasks, academic research in African languages, and applications requiring Hausa language processing. It's particularly suitable for batch processing scenarios given its evaluation batch size of 16.
