Fast Neural Vocoders: WaveGlow and HiFi-GAN
What you'll learn: How flow-based and GAN-based vocoders achieve real-time, high-quality speech synthesis by dramatically improving on WaveNet's speed.
The Speed Problem
You've learned that WaveNet produces excellent audio quality but is painfully slow—it generates samples one at a time autoregressively. For real-time speech synthesis (like voice assistants), this is impractical. WaveGlow and HiFi-GAN solve this by enabling parallel generation.
WaveGlow: Flow-Based Vocoder
WaveGlow uses normalizing flows to model the distribution of audio waveforms. Think of it as a reversible pipeline: during training, it transforms real audio into simple noise; during inference, it reverses the process, converting random noise into speech all at once rather than sample-by-sample.
Key advantages:
- Parallel generation: produces thousands of samples simultaneously
- Exact likelihood training: learns the true distribution of audio
- No quality loss: matches WaveNet quality at 1000× speed
HiFi-GAN: GAN-Based Vocoder
HiFi-GAN takes a different approach using Generative Adversarial Networks. It learns to upsample mel-spectrograms directly into waveforms through adversarial training, with discriminators ensuring both audio quality and perceptual realism.
Key features:
- Multi-scale discriminators: evaluate audio at different resolutions
- Even faster than WaveGlow: achieves real-time synthesis on CPUs
- Lightweight: smaller models, faster inference
Think of HiFi-GAN as a skilled artist who learns shortcuts—it doesn't model every detail explicitly but generates perceptually convincing audio through learned patterns.
Which to Choose?
WaveGlow guarantees exact likelihood modeling (more principled), while HiFi-GAN prioritizes speed and efficiency (more practical for production).
Key Takeaway: WaveGlow and HiFi-GAN enable real-time speech synthesis by replacing WaveNet's autoregressive generation with parallel approaches—flows for exact modeling or GANs for ultra-fast generation.