Vocoder
A model that converts a spectrogram or other audio representation into a waveform - turning abstract audio features into actual sound.
Think of it like
A printer for sound: you describe the pattern (spectrogram), and the printer produces the actual sound waves.
Example
Text-to-speech pipeline: text -> phonemes -> mel-spectrogram (Tacotron) -> waveform (vocoder like WaveGlow or HiFi-GAN).
How it actually works
Spectrograms compress audio (lose phase, time resolution), but humans need waveforms to hear. Vocoders reconstruct waveforms from spectrograms. Early: Griffin-Lim (heuristic, low quality). Modern: WaveNet (autoregressive, high quality, slow), WaveGlow (flow-based, fast), HiFi-GAN (GAN-based, fastest and best quality). Challenges: phase reconstruction, generalization to unseen speakers, real-time synthesis.
For product teams
Critical for TTS quality; faster vocoders enable real-time speech synthesis on mobile.
For engineers
Input: mel-spectrogram. Output: 16 or 24 kHz waveform. Models: HiFi-GAN, WaveGlow, Glow-TTS. Metric: MOS (Mean Opinion Score).
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome