streaming-vocos

Streaming Vocos is a wrapper of Vocos. It supports streaming reconstruction of audio from mel-spectrogram or EnCodec tokens.

Usage

import torch
import torchaudio
from streaming_vocos import StreamingVocos
from wavesurfer import display

waveform, sample_rate = torchaudio.load("data/test_24k.wav")

• From mel-spectrogram

vocos = StreamingVocos("mel")
features = vocos.feature_extractor(waveform)

def audio_generator():
    for feature in torch.unbind(features, dim=2):
        for chunk in vocos.streaming_decode(feature[:, :, None]):
            yield chunk
    yield vocos.decode_caches()

display(audio_generator(), rate=sample_rate, verbose=True)

• From EnCodec tokens

vocos = StreamingVocos("encodec")
codes = vocos.get_encodec_codes(waveform)

def audio_generator():
    for code in torch.unbind(codes, dim=2):
        for chunk in vocos.streaming_decode_codes(code[:, :, None]):
            yield chunk
    yield vocos.decode_caches()

display(audio_generator(), rate=sample_rate, verbose=True)