| Property | Value |
|---|---|
| Parameter Count | 315M |
| License | Apache 2.0 |
| Base Model | facebook/wav2vec2-xls-r-300m |
| Test WER | 25.845% |
| Test CER | 4.795% |
What is xls-r-uyghur-cv7?
XLS-R Uyghur CV7 is a specialized speech recognition model fine-tuned for the Uyghur language using Mozilla's Common Voice 7.0 dataset. Built upon Facebook's wav2vec2-xls-r-300m architecture, it represents a significant advancement in low-resource language ASR technology.
Implementation Details
The model employs a sophisticated training approach where the base XLS-R layers are frozen while fine-tuning a CTC/LM layer specifically for Uyghur speech recognition. The training procedure utilized a carefully crafted learning rate schedule with 2000 warmup steps and a maximum learning rate of 0.0001 across 18,500 total steps.
- Native AMP mixed precision training
- Adam optimizer with custom beta parameters
- Linear learning rate scheduler
- Batch size of 32 with gradient accumulation
Core Capabilities
- Automated speech recognition for Uyghur language
- Support for Perso-Arabic script transcription
- Optimized for broadcast content and general transcription
- 4.795% Character Error Rate on test set
Frequently Asked Questions
Q: What makes this model unique?
This model is specifically optimized for Uyghur language recognition, utilizing a sophisticated fine-tuning approach on the XLS-R architecture. It achieves impressive accuracy with a 25.845% Word Error Rate, making it particularly valuable for low-resource language processing.
Q: What are the recommended use cases?
The model is well-suited for draft video captioning and indexing of recorded broadcasts. However, it's not recommended for live captioning or accessibility purposes where perfect accuracy is crucial. Users should also respect privacy considerations when processing speech data.