PyTorch at Dolby Labs
summary
Dolby uses PyTorch to build audio-first models (patches get merged in days) and focuses on architectures designed for audio rather than spectrogram shortcuts. Key claims: sampleRNN style multi-rate models (MLP front end plus stacked GRU tiers) and WaveNet-class dilated convolutions handle the high dimensionality and multi-scale temporal dependencies. Dolby’s neural codec achieves perceptual quality at 6.4 kbps vs 16 kbps (comparable or better than SILK), and a voice‑conversion system conditions on content and target speaker embeddings (published at Interspeech 2018). Tools: PyTorch dynamic graphs.
AI-written summary · about
chapters (6)
- 0:00Introduction to audio AI
- 1:18Challenges in audio deep learning
- 3:15Breakthroughs in audio generation
- 4:45Deep learning speech coding
- 8:05Voice conversion applications
- 9:34Conclusion and future outlook
related talks
15mOptimizing inference for voice models in production
05mRTC @Scale, Audio ML Session
21mDeveloping Machine Learning Based Speech Enhancement Models for Teams and Skype
22mVocalPy: a core Python package for acoustic comms research
21mCan AI Disrupt Speech Compression?
22mWhy TTS Models Now Look Like LLMs
27mWhat happens when AI starts grocking streaming audio directly
14mLightning Talk: Enabling the Audio Modality for Language Models