lushy's picture
Update FastSpeech 2 code and training notebook
f4d4ab8 verified
Raw History Blame Contribute Delete
9.73 kB
import os
import random
import json
from pathlib import Path
import tgt
import librosa
import numpy as np
import pyworld as pw
from scipy.interpolate import interp1d
from sklearn.preprocessing import StandardScaler
from tqdm import tqdm
import audio as Audio
class Preprocessor:
def __init__(self, config):
self.config = config
self.in_dir = Path(config["path"]["raw_path"])
self.out_dir = Path(config["path"]["preprocessed_path"])
self.val_size = config["preprocessing"]["val_size"]
self.sampling_rate = config["preprocessing"]["audio"]["sampling_rate"]
self.hop_length = config["preprocessing"]["stft"]["hop_length"]
assert config["preprocessing"]["pitch"]["feature"] in ["phoneme_level", "frame_level"]
assert config["preprocessing"]["energy"]["feature"] in ["phoneme_level", "frame_level"]
self.pitch_phoneme_averaging = config["preprocessing"]["pitch"]["feature"] == "phoneme_level"
self.energy_phoneme_averaging = config["preprocessing"]["energy"]["feature"] == "phoneme_level"
self.pitch_normalization = config["preprocessing"]["pitch"]["normalization"]
self.energy_normalization = config["preprocessing"]["energy"]["normalization"]
self.STFT = Audio.stft.TacotronSTFT(
config["preprocessing"]["stft"]["filter_length"],
config["preprocessing"]["stft"]["hop_length"],
config["preprocessing"]["stft"]["win_length"],
config["preprocessing"]["mel"]["n_mel_channels"],
config["preprocessing"]["audio"]["sampling_rate"],
config["preprocessing"]["mel"]["mel_fmin"],
config["preprocessing"]["mel"]["mel_fmax"],
)
def build_from_path(self):
(self.out_dir / "mel").mkdir(parents=True, exist_ok=True)
(self.out_dir / "pitch").mkdir(parents=True, exist_ok=True)
(self.out_dir / "energy").mkdir(parents=True, exist_ok=True)
(self.out_dir / "duration").mkdir(parents=True, exist_ok=True)
print("Processing Data ...")
out = list()
n_frames = 0
pitch_scaler = StandardScaler()
energy_scaler = StandardScaler()
speakers = {}
search_dirs = [p for p in self.in_dir.iterdir() if p.is_dir()]
for i, speaker_dir in enumerate(tqdm(search_dirs)):
speaker = speaker_dir.name
speakers[speaker] = i
for wav_path in speaker_dir.iterdir():
if wav_path.suffix != ".wav":
continue
basename = wav_path.stem
tg_path = self.out_dir / "TextGrid" / speaker / f"{basename}.TextGrid"
if tg_path.exists():
ret = self.process_utterance(speaker, basename)
if ret is None:
continue
info, pitch, energy, n = ret
out.append(info)
if len(pitch) > 0:
pitch_scaler.partial_fit(pitch.reshape((-1, 1)))
if len(energy) > 0:
energy_scaler.partial_fit(energy.reshape((-1, 1)))
n_frames += n
print("Computing statistic quantities ...")
if hasattr(pitch_scaler, "mean_") and pitch_scaler.mean_ is not None and n_frames > 0:
pitch_mean = pitch_scaler.mean_[0] if self.pitch_normalization else 0
pitch_std = pitch_scaler.scale_[0] if self.pitch_normalization else 1
energy_mean = energy_scaler.mean_[0] if self.energy_normalization else 0
energy_std = energy_scaler.scale_[0] if self.energy_normalization else 1
pitch_min, pitch_max = self.normalize(self.out_dir / "pitch", pitch_mean, pitch_std)
energy_min, energy_max = self.normalize(self.out_dir / "energy", energy_mean, energy_std)
else:
print("\n[!] Warning: No audio samples matched with a corresponding TextGrid file.")
pitch_min, pitch_max, pitch_mean, pitch_std = 0, 1, 0, 1
energy_min, energy_max, energy_mean, energy_std = 0, 1, 0, 1
with open(self.out_dir / "speakers.json", "w") as f:
f.write(json.dumps(speakers))
with open(self.out_dir / "stats.json", "w") as f:
stats = {
"pitch": [float(pitch_min), float(pitch_max), float(pitch_mean), float(pitch_std)],
"energy": [float(energy_min), float(energy_max), float(energy_mean), float(energy_std)],
}
f.write(json.dumps(stats))
print(f"Total time: {n_frames * self.hop_length / self.sampling_rate / 3600:.2f} hours")
random.shuffle(out)
out = [r for r in out if r is not None]
with open(self.out_dir / "train.txt", "w", encoding="utf-8") as f:
for m in out[self.val_size :]:
f.write(f"{m}\n")
with open(self.out_dir / "val.txt", "w", encoding="utf-8") as f:
for m in out[: self.val_size]:
f.write(f"{m}\n")
return out
def process_utterance(self, speaker, basename):
wav_path = self.in_dir / speaker / f"{basename}.wav"
text_path = self.in_dir / speaker / f"{basename}.lab"
tg_path = self.out_dir / "TextGrid" / speaker / f"{basename}.TextGrid"
textgrid = tgt.io.read_textgrid(str(tg_path))
phone, duration, start, end = self.get_alignment(textgrid.get_tier_by_name("phones"))
text = "{" + " ".join(phone) + "}"
if start >= end:
return None
wav, _ = librosa.load(str(wav_path))
wav = wav[int(self.sampling_rate * start) : int(self.sampling_rate * end)].astype(np.float32)
with open(text_path, "r", encoding="utf-8") as f:
raw_text = f.readline().strip("\n")
pitch, t = pw.dio(
wav.astype(np.float64),
self.sampling_rate,
frame_period=self.hop_length / self.sampling_rate * 1000,
)
pitch = pw.stonemask(wav.astype(np.float64), pitch, t, self.sampling_rate)
pitch = pitch[: sum(duration)]
if np.sum(pitch != 0) <= 1:
return None
mel_spectrogram, energy = Audio.tools.get_mel_from_wav(wav, self.STFT)
mel_spectrogram = mel_spectrogram[:, : sum(duration)]
energy = energy[: sum(duration)]
if self.pitch_phoneme_averaging:
nonzero_ids = np.where(pitch != 0)[0]
interp_fn = interp1d(
nonzero_ids,
pitch[nonzero_ids],
fill_value=(pitch[nonzero_ids[0]], pitch[nonzero_ids[-1]]),
bounds_error=False,
)
pitch = interp_fn(np.arange(0, len(pitch)))
pos = 0
for i, d in enumerate(duration):
pitch[i] = np.mean(pitch[pos : pos + d]) if d > 0 else 0
pos += d
pitch = pitch[: len(duration)]
if self.energy_phoneme_averaging:
pos = 0
for i, d in enumerate(duration):
energy[i] = np.mean(energy[pos : pos + d]) if d > 0 else 0
pos += d
energy = energy[: len(duration)]
np.save(self.out_dir / "duration" / f"{speaker}-duration-{basename}.npy", duration)
np.save(self.out_dir / "pitch" / f"{speaker}-pitch-{basename}.npy", pitch)
np.save(self.out_dir / "energy" / f"{speaker}-energy-{basename}.npy", energy)
np.save(self.out_dir / "mel" / f"{speaker}-mel-{basename}.npy", mel_spectrogram.T)
return (
"|".join([basename, speaker, text, raw_text]),
self.remove_outlier(pitch),
self.remove_outlier(energy),
mel_spectrogram.shape[1],
)
def get_alignment(self, tier):
sil_phones = ["sil", "sp", "spn"]
phones, durations = [], []
start_time, end_time, end_idx = 0, 0, 0
for t in tier._objects:
s, e, p = t.start_time, t.end_time, t.text
if phones == []:
if p in sil_phones:
continue
else:
start_time = s
if p not in sil_phones:
phones.append(p)
end_time = e
end_idx = len(phones)
else:
phones.append(p)
durations.append(
int(np.round(e * self.sampling_rate / self.hop_length) - np.round(s * self.sampling_rate / self.hop_length))
)
return phones[:end_idx], durations[:end_idx], start_time, end_time
def remove_outlier[T](self, values: T) -> T:
values = np.array(values)
p25 = np.percentile(values, 25)
p75 = np.percentile(values, 75)
lower = p25 - 1.5 * (p75 - p25)
upper = p75 + 1.5 * (p75 - p25)
normal_indices = np.logical_and(values > lower, values < upper)
return values[normal_indices]
def normalize(self, in_dir: Path, mean: float, std: float):
max_value = np.finfo(np.float64).min
min_value = np.finfo(np.float64).max
for file_path in in_dir.iterdir():
if file_path.is_file():
values = (np.load(file_path) - mean) / std
np.save(file_path, values)
max_value = max(max_value, max(values))
min_value = min(min_value, min(values))
return min_value, max_value