Skip to content

FreGrad

A lightweight frequency-aware diffusion vocoder

Visit official site

What FreGrad is

The vocoder is described as lightweight and fast, using a frequency-aware diffusion approach to generate realistic audio.

What you can do with it

  • Train audio generation models
  • Apply the vocoder to speech synthesis
  • Use it in audio processing plugins

Reported figures

  • Training is stated as 3.7 times faster
  • Inference is stated as 2.2 times faster
  • The model is stated at 1.78 million parameters

Pros & cons

✓ What we like

  • Training is stated as 3.7 times faster
  • Inference is stated as 2.2 times faster

! What to watch out for

  • The stated model size is 1.78 million parameters
  • Gains are reported against a baseline model

FAQ

What is in the framework?

Discrete wavelet transform, frequency-aware dilated convolution and quality enhancement techniques.

What speed figures are listed?

Training 3.7 times faster and inference 2.2 times faster than a baseline.

Last reviewed: 2026-11-15

More AI music creation tools

View all →

How we review