metadata
license: apache-2.0
This model is trained on Google's AudioSet (28GB data) for 1 million steps. (Originally planned 2 million steps, but I'm exploring better training schedule)
You can regard it as a pretrained base model, which is common in language models but not for vocoders.
How to load and use this model:
import torch
import torchaudio
from scipy.io.wavfile import write
with torch.no_grad():
from vocos import Vocos
A = torch.load("./vocos_checkpoint_epoch=464_step=1001610_val_loss=7.1732.ckpt", map_location="cpu")
V = Vocos.from_hparams("./config.yaml")
V.load_state_dict(A['state_dict'], strict=False)
V.eval()
def safe_log(x: torch.Tensor, clip_val: float = 1e-7):
return torch.log(torch.clip(x, min=clip_val))
voice, sr = torchaudio.load('example.wav') # must be sample_rate=32000
if sr != 32000:
raise ValueError
mel = torchaudio.transforms.MelSpectrogram(
sample_rate=32000, n_fft=2048, hop_length=1024, n_mels=128, center=True, power=1,
)(voice)
mel = safe_log(mel)
audio = V.decode(mel)
write('out.wav', 32000, audio.flatten().numpy())