neural vocoder

Neural vocoders convert acoustic features (mel spectrograms) back into high-fidelity audio waveforms. **Role in TTS pipeline**: Text leads to acoustic model leads to mel spectrogram leads to vocoder leads to audio waveform. Vocoder is final synthesis stage. **Why needed**: Mel spectrograms are compact representation, but contain no phase information needed for waveform. Vocoder reconstructs plausible phase and generates samples. **Key architectures**: **Autoregressive**: WaveNet (slow, high quality, sample-by-sample), WaveRNN. **Non-autoregressive**: HiFi-GAN (fast, excellent quality), UnivNet, Vocos. **GAN vocoders**: Generator produces waveform, discriminators judge quality. Multi-scale and multi-period discriminators. **Training**: Reconstruct original audio from mel spectrogram, GAN loss + feature matching + mel reconstruction. **Quality vs speed**: WaveNet: 1000x slower than real-time. HiFi-GAN: 1000x faster than real-time, comparable quality. **Universal vocoders**: Work across speakers/conditions vs speaker-specific. **Integration**: End-to-end models (VITS) combine acoustic model and vocoder. HiFi-GAN made high-quality neural TTS practical.