WaveGlow: Generative Modeling for Audio Synthesis
summary
Waveglow is a flow-based vocoder that converts mel-spectrograms into raw waveforms by using invertible affine coupling layers and 1×1 invertible convolutions. It maximizes exact data likelihood (not GAN or MSE losses), stacks about a dozen flow blocks with gated convolutions, runs on GPUs, and produces parallel high-quality audio at roughly 10 kHz and ~20 seconds of output per second of GPU time.
AI-written summary · about
chapters (16)
- 0:00Audio synthesis intro
- 1:09Classical speech methods
- 3:24Parametric limits
- 4:56Modern spectrograms
- 9:03Generative model classes
- 10:11GAN pros and cons
- 11:16VAE interpretability
- 12:17Flow-based uniqueness
- 14:09Why generative models
- 18:28Normalizing flows
- 22:45WaveGlow architecture
- 27:16Loss and training
- 29:32WaveGlow vs others
- 32:30WaveGlow performance
- 34:22Flow efficiency
- 35:28Q&A
related talks
30mSound Modelling - parametric methods and deep learning...
58mMusic Generation via Masked Acoustic Token Modeling
23mFLUX, Open Research, and the Future of Visual AI
09mSynthesizing Video and Turning it into Music!
01h 04mKeynote 3. Interactive generative models for music
01h 03mGenerative AI and What Is Meaningful Sperm Whale Communication
21mCan AI Disrupt Speech Compression?
28mReal-Time Data for Generative Feedback Loop (GFL) Automation