| Property | Value |
|---|---|
| License | Apache-2.0 |
| Author | BELLE-2 |
| Framework | PyTorch, Transformers |
| Task | Automatic Speech Recognition |
What is Belle-whisper-large-v3-zh?
Belle-whisper-large-v3-zh is an advanced Chinese speech recognition model that builds upon OpenAI's Whisper large-v3 architecture. Through comprehensive fine-tuning on major Chinese speech datasets, it achieves remarkable improvements in Chinese ASR performance, showing 24-65% better results compared to the base model across various benchmarks.
Implementation Details
The model underwent full fine-tuning using multiple high-quality Chinese speech datasets, including AISHELL-1, AISHELL-2, WenetSpeech, and HKUST. It operates at a 16KHz sample rate and leverages the Transformers library for easy deployment.
- Significant performance improvements on Chinese ASR benchmarks
- Specialized for complex acoustic environments
- Easy integration through Hugging Face Transformers pipeline
- Supports transcription tasks with Chinese language optimization
Core Capabilities
- Achieves 2.781% CER on AISHELL-1 test set
- Demonstrates exceptional performance in meeting scenarios (11.246% CER on WenetSpeech meeting)
- Handles diverse acoustic environments effectively
- Seamless integration with standard ASR pipelines
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its significantly improved performance on Chinese ASR tasks, particularly in challenging acoustic environments like meeting recordings. It offers a substantial improvement over the base whisper-large-v3 model, with error rate reductions of up to 65% in some scenarios.
Q: What are the recommended use cases?
The model is ideal for Chinese speech recognition tasks, particularly in scenarios requiring high accuracy such as meeting transcription, general speech-to-text conversion, and applications requiring robust performance in various acoustic conditions. It's especially effective for complex audio environments where traditional ASR systems might struggle.