Hviske-v2
| Property | Value |
|---|---|
| Developer | syv.ai |
| Parameter Count | 1.54B |
| Model Type | Speech Recognition |
| Base Architecture | Whisper v3 |
| Training Hardware | Single Nvidia A100 |
| Model Hub | HuggingFace |
What is hviske-v2?
Hviske-v2 is a cutting-edge Danish speech transcription model developed by syv.ai. It represents a significant improvement over its predecessors, achieving a Word Error Rate (WER) of 11.8% and Character Error Rate (CER) of 4.7%. The model is a fine-tuned version of Whisper v3, specifically optimized using Coral and Common Voice datasets for Danish language processing.
Implementation Details
The model was trained over a 10-day period using a single Nvidia A100 GPU. It builds upon the Whisper v3 architecture and has been specifically enhanced for Danish language processing. Implementation requires the transformers and datasets libraries, and can be deployed on both CPU and GPU environments.
- Built on Whisper v3 architecture with 1.54B parameters
- Supports both CPU and GPU execution with automatic hardware detection
- Implements float16 precision on GPU for optimal performance
- Includes built-in processor and feature extractor components
Core Capabilities
- 30% better WER than ROEST model
- 64% improvement over Hviske-v1
- Achieves 11.8% ± 0.3% WER on CoRal dataset
- Handles Danish speech recognition with state-of-the-art accuracy
Frequently Asked Questions
Q: What makes this model unique?
Hviske-v2 sets itself apart by achieving significantly better performance metrics compared to other Danish speech recognition models, with a 30% improvement in WER over ROEST and 64% improvement over its predecessor. It's specifically optimized for Danish language processing while maintaining the robust architecture of Whisper v3.
Q: What are the recommended use cases?
The model is ideal for Danish speech transcription tasks, particularly in applications requiring high accuracy. It can be integrated into various applications through the Transformers library and is suitable for both research and production environments. Users can try it freely through Ludwig.syv.ai.
