rfwave-libritts-24k

RFWave LibriTTS 24 kHz

Mel-spectrogram-conditioned 24 kHz vocoder from the RFWave family. It reconstructs a 24 kHz mono waveform from 100-band mel spectrograms with a rectified-flow backbone and an STFT output head.

⚠️ Code compatibility — read this first

These weights must be used together with the RFWave code at commit 28698f06b188e49bd2e112f8e9e67e777292a323 (2024-11-01, "sample steps according to straightness"). The module layout and state-dict keys correspond to that revision; other revisions may fail to load with strict=True, or may silently mismatch.

git clone https://github.com/bfs18/rfwave.git
cd rfwave && git checkout 28698f06b188e49bd2e112f8e9e67e777292a323

config.yaml in this repo is the config from that revision.

This is the mel-conditioned 24 kHz model from the rfwave-dev-infill-stft-true experiment (weights exported 2024-06-02). Its EnCodec-conditioned counterpart is laupeng1989/rfwave-encodec-24k.

Files

Architecture

From config.yaml:

• Experiment: rfwave.experiment_reflow_subband.VocosExp
• Backbone: VocosRFBackbone — dim=512, intermediate_dim=1536, num_layers=8, num_bands=8, input_channels=100, output_channels=160, Fourier features with pe_scale=1000
• Conditioning: MelSpectrogramFeatures — n_fft=1024, hop_length=256, n_mels=100, padding=center, sample_rate=24000
• Head: RFSTFTHead — n_fft=1024, hop_length=256, padding=center
• Sampling rate: 24 000 Hz, mono
• Training: rectified flow, p_uncond=0.1, inference guidance 2.0

Training data: LibriTTS at 24 kHz.

Usage

import torch

ckpt = torch.load("last.ckpt", map_location="cpu", weights_only=True)
state_dict = ckpt["state_dict"]          # 104 tensors

# instantiate rfwave.experiment_reflow_subband.VocosExp from config.yaml (commit 28698f0),
# then:
# model.load_state_dict(state_dict, strict=True)

License

BSD, matching the RFWave project.