MassivelyMultilingualTTS / Preprocessing /DescriptCodecAudioPreprocessor.py
Flux9665's picture
use explicit code instead of relying on release download
9e275b8
raw
history blame
3.41 kB
import torch
from torchaudio.transforms import Resample
class CodecAudioPreprocessor:
def __init__(self, input_sr, output_sr=16000, device="cpu"):
from dac.model import DAC
from dac.utils import load_model
self.device = device
self.input_sr = input_sr
self.output_sr = output_sr
self.resample = Resample(orig_freq=input_sr, new_freq=output_sr).to(self.device)
self.model = DAC()
self.model = load_model(model_type="16kHz", tag="0.0.5")
self.model.eval()
self.model.to(device)
def resample_audio(self, audio, current_sampling_rate):
if current_sampling_rate != self.input_sr:
print("warning, change in sampling rate detected. If this happens too often, consider re-ordering the audios so that the sampling rate stays constant for multiple samples")
self.resample = Resample(orig_freq=current_sampling_rate, new_freq=self.output_sr).to(self.device)
self.input_sr = current_sampling_rate
audio = torch.tensor(audio, device=self.device, dtype=torch.float32)
audio = self.resample(audio)
return audio
@torch.inference_mode()
def audio_to_codec_tensor(self, audio, current_sampling_rate):
if current_sampling_rate != self.output_sr:
audio = self.resample_audio(audio, current_sampling_rate)
elif type(audio) != torch.tensor:
audio = torch.tensor(audio, device=self.device, dtype=torch.float32)
return self.model.encode(audio.unsqueeze(0).unsqueeze(0))[0].squeeze()
@torch.inference_mode()
def audio_to_codebook_indexes(self, audio, current_sampling_rate):
if current_sampling_rate != self.output_sr:
audio = self.resample_audio(audio, current_sampling_rate)
elif type(audio) != torch.tensor:
audio = torch.tensor(audio, device=self.device, dtype=torch.float32)
return self.model.encode(audio.unsqueeze(0).unsqueeze(0))[1].squeeze()
@torch.inference_mode()
def indexes_to_codec_frames(self, codebook_indexes):
if len(codebook_indexes.size()) == 2:
codebook_indexes = codebook_indexes.unsqueeze(0)
return self.model.quantizer.from_codes(codebook_indexes)[1].squeeze()
@torch.inference_mode()
def indexes_to_audio(self, codebook_indexes):
return self.codes_to_audio(self.indexes_to_codec_frames(codebook_indexes))
@torch.inference_mode()
def codes_to_audio(self, continuous_codes):
z_q = 0.0
z_ps = torch.split(continuous_codes, self.model.codebook_dim, dim=0)
for i, z_p in enumerate(z_ps):
z_q_i = self.model.quantizer.quantizers[i].out_proj(z_p)
z_q = z_q + z_q_i
return self.model.decode(z_q.unsqueeze(0)).squeeze()
if __name__ == '__main__':
import soundfile
import time
with torch.inference_mode():
test_audio = "../audios/ry.wav"
wav, sr = soundfile.read(test_audio)
ap = CodecAudioPreprocessor(input_sr=sr)
indexes = ap.audio_to_codebook_indexes(wav, current_sampling_rate=sr)
print(indexes.shape)
t0 = time.time()
audio = ap.indexes_to_audio(indexes)
t1 = time.time()
print(audio.shape)
print(t1 - t0)
soundfile.write(file=f"../audios/ry_reconstructed_in_{t1 - t0}_descript.wav", data=audio, samplerate=16000)