ex3ndr / supervoice-vocoder

Production-ready vocoder using BigVSAN

✨ SuperVoice Vocoder

Easy to use SOTA vocoder for speech synthesis from Mel Spectograms.

Description

This repository contains easy to use SOTA vocoder that is uses BigVSAN original pre-trained weights with removed weight normalization and most of the code, but built to be used in a plug and play fashion. This model is tailored to a specific Mel-Spectogram parameters which is most common for voice synthesis tasks.

• Model: BigVSAN
• Sample rate: 24000 Hz
• Mel Spectogram:
  • Mel Number: 100
  • Number of FFT: 1024
  • Hop Length: 256
  • Window Length: 1024
  • Norm: slaney
  • Scale: slaney
  • Power: 1.0 aka amplitude spectogram
  • Center: true
  • Padding: reflect

Evaluation

To evaluate model you can use evaluation notebook which can run anywhere where torch and torchaudio are installed.

How to use

This model is available using Torch Hub:

# Load model
bigvsan = torch.hub.load(repo_or_dir='ex3ndr/supervoice-vocoder', model='bigvsan')

# Source mel spectogram
spec = torch.randn(100, 1234)

# Synthesized audio
audio = model.generate(spec)

License

MIT