kashif/soundstream_mel_decoder

A SoundStream decoder to reconstruct audio from a mel-spectrogram.

Overview

This model is a SoundStream decoder which inverts mel-spectrograms computed with the specific hyperparameters defined in the example below. This model was trained on music data and used in Multi-instrument Music Synthesis with Spectrogram Diffusion (ISMIR 2022).

A typical use-case is to simplify music generation by predicting mel-spectrograms (instead of a raw waveform), and then use this model to reconstruct audio.

If you use it, please consider citing:

@article{zeghidour2021soundstream,
  title={Soundstream: An end-to-end neural audio codec},
  author={Zeghidour, Neil and Luebs, Alejandro and Omran, Ahmed and Skoglund, Jan and Tagliasacchi, Marco},
  journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing},
  volume={30},
  pages={495--507},
  year={2021},
  publisher={IEEE}
}

Example Use

from diffusers import OnnxRuntimeModel


SAMPLE_RATE = 16000
N_FFT = 1024
HOP_LENGTH = 320
WIN_LENGTH = 640
N_MEL_CHANNELS = 128
MEL_FMIN = 0.0
MEL_FMAX = int(SAMPLE_RATE // 2)
CLIP_VALUE_MIN = 1e-5
CLIP_VALUE_MAX = 1e8

mel = ...

melgan = OnnxRuntimeModel.from_pretrained("kashif/soundstream_mel_decoder")

audio = melgan(input_features=mel.astype(np.float32))