hr16/ljspeech-sidon

This is a preprocessed dataset made by applying Sidon: Fast and Robust Open-Source Multilingual Speech Restoration for Large-scale Dataset Cleansing (https://huggingface.co/spaces/sarulab-speech/sidon_demo_beta) on LJSpeech-1.1 dataset

Format: Following Stylish-TTS

Files:

• ljspeech-sidon-48khz.zip
• ljspeech-sidon-24khz.zip

Reproduction code:

%cd /content
!sudo apt install aria2 -y
!rm LJSpeech-1.1.tar.bz2
!aria2c -x 16 https://data.keithito.com/data/speech/LJSpeech-1.1.tar.bz2
!tar -xf LJSpeech-1.1.tar.bz2
!uv pip install "misaki[en] @ git+https://github.com/Fannovel16/misaki" --system

import numpy as np
import torch
import torchaudio
import transformers
from huggingface_hub import hf_hub_download
import soundfile as sf
from misaki.en import G2P

g2p = G2P()
fe_path = hf_hub_download("sarulab-speech/sidon-v0.1", filename="feature_extractor_cuda.pt")
decoder_path = hf_hub_download("sarulab-speech/sidon-v0.1", filename="decoder_cuda.pt")

preprocessor = transformers.SeamlessM4TFeatureExtractor.from_pretrained(
    "facebook/w2v-bert-2.0"
)


fe = torch.jit.load(fe_path,map_location='cuda').to('cuda')
decoder = torch.jit.load(decoder_path,map_location='cuda').to('cuda')

def denoise_speech(audio):
    if audio is None:
        return None

    sample_rate, waveform = audio
    waveform = 0.9 * (waveform / np.abs(waveform).max())
    target_n_samples = int(48_000/sample_rate* waveform.shape[0])
    # Ensure waveform is a tensor
    if not isinstance(waveform, torch.Tensor):
        waveform = torch.tensor(waveform, dtype=torch.float32)

    # If stereo, convert to mono
    if waveform.ndim > 1 and waveform.shape[0] > 1:
        waveform = torch.mean(waveform, dim=1)

    # Add a batch dimension
    waveform = waveform.view(1, -1)
    wav = torchaudio.functional.highpass_biquad(waveform, sample_rate, 50)
    wav_16k = torchaudio.functional.resample(wav, sample_rate, 16_000)
    restoreds = []
    features =[]
    feature_cache = None
    wav_16k = torch.nn.functional.pad(wav_16k,(0,24000))
    for chunk in wav_16k.view(-1).split(16000 * 96):
        inputs = preprocessor(
            torch.nn.functional.pad(chunk, (160, 160)), sampling_rate=16_000, return_tensors="pt"
        ).to('cpu')
        with torch.inference_mode():
            feature = fe(inputs["input_features"].to("cuda"))["last_hidden_state"]
            if feature_cache is not None:
                feature = torch.cat([feature_cache,feature],dim=1)
            restoreds.append(decoder(feature.transpose(1,2)).view(-1)[:-960])
            feature_cache = feature[:,-1:]
        
    restored_wav = torch.cat(restoreds,dim=0)

    return 48_000, restored_wav.cpu().numpy()[:target_n_samples]

!rm -r /content/ljspeech-sidon-48khz /content/ljspeech-sidon-24khz
!mkdir /content/ljspeech-sidon-48khz
!mkdir /content/ljspeech-sidon-48khz/wavs
!mkdir /content/ljspeech-sidon-24khz
!mkdir /content/ljspeech-sidon-24khz/wavs

from pathlib import Path
from tqdm import tqdm
import soundfile as sf
import librosa
train_list, val_list = [], []
for line in tqdm(Path("/content/LJSpeech-1.1/metadata.csv").read_text().splitlines()):
    fid, _, text = line.split('|')
    book_id = int(fid.split('-')[0].removeprefix('LJ'))
    phoneme = g2p(text)[0]
    line = f"{fid}.wav|{phoneme}|0|{text}"
    
    waveform, sr = sf.read(f"/content/LJSpeech-1.1/wavs/{fid}.wav")
    sr, waveform = denoise_speech((sr, waveform))
    waveform_24khz = librosa.resample(waveform, orig_sr=sr, target_sr=24_000)
    sf.write(f"/content/ljspeech-sidon-48khz/wavs/{fid}.wav", waveform, sr)
    sf.write(f"/content/ljspeech-sidon-24khz/wavs/{fid}.wav", waveform_24khz, 24_000)
    if book_id <= 40:
        train_list.append(line)
    else:
        val_list.append(line)
train_list = '\n'.join(train_list)
val_list = '\n'.join(val_list)
Path("/content/ljspeech-sidon-48khz/train-list.txt").write_text(train_list)
Path("/content/ljspeech-sidon-48khz/val-list.txt").write_text(val_list)
Path("/content/ljspeech-sidon-24khz/train-list.txt").write_text(train_list)
Path("/content/ljspeech-sidon-24khz/val-list.txt").write_text(val_list)