DeepSpeech: A Journey to <10% Word Error Rate
Abstract
Talk on DeepSpeech and the journey to achieving less than 10% word error rate in automatic speech recognition.
Key points:
- Mozilla DeepSpeech architecture and training
- Challenges in achieving low word error rates
- On-premise ASR deployment for privacy
- Real-world deployment at Fireflies.ai serving 10M+ users
Takeaway: Open-source ASR can achieve production-quality results while keeping speech data on-premise for privacy.
