EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·25d ago·5 min read
Canto: A speech model built for the real world
.jpg&w=3840&q=75)
TL;DR
Canto, a new speech recognition model from Wispr Advanced Interfaces Lab, excels in real-world dictation scenarios, achieving the lowest word error rate among competitors. It is designed to handle challenging audio conditions, making it suitable for everyday use.
✦ Why It Matters
Consider using Canto for applications requiring accurate speech recognition in noisy environments.
Key Takeaways
- Evaluate Canto for real-time dictation applications.
- Utilize user corrections to improve model accuracy.
- Explore reinforcement learning techniques for speech models.
- Focus on training models for challenging audio conditions.
How It Works
Canto employs a two-stage training process: first, it learns to transcribe audio using supervised fine-tuning, then it uses reinforcement learning to refine its transcription quality based on feedback from generated outputs.