File size: 8,923 Bytes
6e802a4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 | # score_naturalness_panel.py: UTMOS and spectral / F0 diagnostics over a folder of wavs.
from __future__ import annotations
import argparse
import json
import sys
from pathlib import Path
import numpy as np
REPO_ROOT = Path(__file__).resolve().parent.parent
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
def hf_percent(wav, sr, cutoff_hz=6000.0):
x = np.asarray(wav, dtype=np.float32)
if x.ndim > 1:
x = x.mean(axis=1)
n = 2048
if len(x) < n:
return 0.0
win = np.hanning(n)
mags = np.stack([np.abs(np.fft.rfft(x[i:i + n] * win))
for i in range(0, len(x) - n, n // 2)])
p = (mags ** 2).mean(axis=0)
freqs = np.fft.rfftfreq(n, 1.0 / sr)
return float(p[freqs >= cutoff_hz].sum() / (p.sum() + 1e-12) * 100.0)
def f0_track(wav, sr):
import librosa
x = np.asarray(wav, dtype=np.float32)
if x.ndim > 1:
x = x.mean(axis=1)
hop = int(sr * 0.010)
f0, _, _ = librosa.pyin(x, fmin=60, fmax=400, sr=sr, hop_length=hop,
frame_length=2048)
return f0, 1.0 / 0.010
def flutter_percent(f0, fs, fast_hz=8.0):
v = np.isfinite(f0)
if not v.any():
return float("nan")
runs, start = [], None
for i, ok in enumerate(v):
if ok and start is None:
start = i
elif not ok and start is not None:
runs.append((start, i)); start = None
if start is not None:
runs.append((start, len(v)))
lo, hi = max(runs, key=lambda r: r[1] - r[0])
seg = f0[lo:hi]
if len(seg) < int(0.4 * fs):
return float("nan")
seg = seg - np.polyval(np.polyfit(np.arange(len(seg)), seg, 1),
np.arange(len(seg)))
spec = np.abs(np.fft.rfft(seg * np.hanning(len(seg)))) ** 2
freqs = np.fft.rfftfreq(len(seg), 1.0 / fs)
band = freqs >= 1.0
fast = freqs > fast_hz
return float(spec[fast].sum() / (spec[band].sum() + 1e-12) * 100.0)
def am_probe(wav, sr, band_lo=4.0, band_hi=30.0):
x = np.asarray(wav, dtype=np.float32)
if x.ndim > 1:
x = x.mean(axis=1)
hop = int(sr * 0.010)
nfr = len(x) // hop
if nfr < 60:
return float("nan"), float("nan")
rms = np.sqrt(np.mean(
x[: nfr * hop].reshape(nfr, hop) ** 2, axis=1))
fs = 100.0
thr = rms.max() * (10.0 ** (-30.0 / 20.0))
active = rms > thr
runs, start = [], None
for i, ok in enumerate(active):
if ok and start is None:
start = i
elif not ok and start is not None:
runs.append((start, i)); start = None
if start is not None:
runs.append((start, len(active)))
if not runs:
return float("nan"), float("nan")
lo, hi = max(runs, key=lambda r: r[1] - r[0])
seg = rms[lo:hi]
if len(seg) < int(0.5 * fs):
return float("nan"), float("nan")
seg = seg - np.polyval(np.polyfit(np.arange(len(seg)), seg, 1),
np.arange(len(seg)))
spec = np.abs(np.fft.rfft(seg * np.hanning(len(seg)))) ** 2
freqs = np.fft.rfftfreq(len(seg), 1.0 / fs)
denom = spec[freqs >= 1.0].sum() + 1e-12
am_pct = float(spec[(freqs >= band_lo) & (freqs <= band_hi)].sum()
/ denom * 100.0)
pk_mask = (freqs >= 2.0) & (freqs <= 40.0)
peak_hz = float(freqs[pk_mask][np.argmax(spec[pk_mask])])
return am_pct, peak_hz
def low_am_percent(wav, sr, f_lo=100.0, f_hi=500.0, m_lo=4.0, m_hi=7.5,
n_fft=1024, hop=256):
x = np.asarray(wav, dtype=np.float32)
if x.ndim > 1:
x = x.mean(axis=1)
if len(x) < n_fft * 4:
return float("nan")
win = np.hanning(n_fft)
frames = np.lib.stride_tricks.sliding_window_view(x, n_fft)[::hop] * win
S = np.abs(np.fft.rfft(frames, axis=1))
freqs = np.fft.rfftfreq(n_fft, 1.0 / sr)
fs_env = sr / hop
full = S.sum(axis=1)
thr = full.max() * (10.0 ** (-30.0 / 20.0))
active = full > thr
runs, start = [], None
for i, ok in enumerate(active):
if ok and start is None:
start = i
elif not ok and start is not None:
runs.append((start, i)); start = None
if start is not None:
runs.append((start, len(active)))
if not runs:
return float("nan")
lo, hi = max(runs, key=lambda r: r[1] - r[0])
if hi - lo < int(0.5 * fs_env):
return float("nan")
env = S[lo:hi, (freqs >= f_lo) & (freqs < f_hi)].sum(axis=1)
env = env - np.polyval(np.polyfit(np.arange(len(env)), env, 1),
np.arange(len(env)))
spec = np.abs(np.fft.rfft(env * np.hanning(len(env)))) ** 2
mf = np.fft.rfftfreq(len(env), 1.0 / fs_env)
denom = spec[mf >= 1.0].sum() + 1e-12
return float(spec[(mf >= m_lo) & (mf < m_hi)].sum() / denom * 100.0)
def flatmod_percent(wav, sr, m_lo=12.0, m_hi=25.0, f_lo=300.0, f_hi=4000.0,
n_fft=1024, hop=256):
x = np.asarray(wav, dtype=np.float32)
if x.ndim > 1:
x = x.mean(axis=1)
if len(x) < n_fft * 4:
return float("nan")
win = np.hanning(n_fft)
frames = np.lib.stride_tricks.sliding_window_view(x, n_fft)[::hop] * win
S = np.abs(np.fft.rfft(frames, axis=1)) ** 2 + 1e-12
freqs = np.fft.rfftfreq(n_fft, 1.0 / sr)
sel = (freqs >= f_lo) & (freqs < f_hi)
Sb = S[:, sel]
flat = np.exp(np.mean(np.log(Sb), axis=1)) / np.mean(Sb, axis=1)
fs_env = sr / hop
full = S.sum(axis=1)
thr = full.max() * (10.0 ** (-30.0 / 20.0))
active = full > thr
runs, start = [], None
for i, ok in enumerate(active):
if ok and start is None:
start = i
elif not ok and start is not None:
runs.append((start, i)); start = None
if start is not None:
runs.append((start, len(active)))
if not runs:
return float("nan")
lo, hi = max(runs, key=lambda r: r[1] - r[0])
seg = np.log(flat[lo:hi] + 1e-9)
if len(seg) < int(0.5 * fs_env):
return float("nan")
seg = seg - np.polyval(np.polyfit(np.arange(len(seg)), seg, 1),
np.arange(len(seg)))
spec = np.abs(np.fft.rfft(seg * np.hanning(len(seg)))) ** 2
mf = np.fft.rfftfreq(len(seg), 1.0 / fs_env)
denom = spec[mf >= 1.0].sum() + 1e-12
return float(spec[(mf >= m_lo) & (mf < m_hi)].sum() / denom * 100.0)
def jitter_percent(f0):
v = np.isfinite(f0)
d = []
for i in range(1, len(f0)):
if v[i] and v[i - 1]:
d.append(abs(f0[i] - f0[i - 1]) / f0[i - 1])
return float(np.mean(d) * 100.0) if d else float("nan")
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--dirs", nargs="+", required=True,
help="directories of .wav files; each becomes one row")
ap.add_argument("--out", default=None, help="write per-file + summary JSON")
ap.add_argument("--no_utmos", action="store_true")
ap.add_argument("--device", default="cpu")
args = ap.parse_args()
import soundfile as sf
scorer = None
if not args.no_utmos:
from eval.naturalness import NaturalnessScorer
scorer = NaturalnessScorer(device=args.device)
all_rows = {}
print(f"{'dir':28s} {'n':>3s} {'utmos':>6s} {'hf%':>6s} {'flut%':>6s} "
f"{'jit%':>6s} {'am%':>6s} {'amPk':>5s} {'lowAM':>6s} {'flatM':>6s}")
for d in args.dirs:
wavs = sorted(Path(d).glob("*.wav"))
rows = []
for w in wavs:
audio, sr = sf.read(str(w), dtype="float32")
f0, fs = f0_track(audio, sr)
am_pct, am_peak = am_probe(audio, sr)
row = {"file": w.name,
"hf_pct": round(hf_percent(audio, sr), 3),
"flutter_pct": round(flutter_percent(f0, fs), 2),
"jitter_pct": round(jitter_percent(f0), 3),
"am_pct": round(am_pct, 2),
"am_peak_hz": round(am_peak, 2),
"lowam_pct": round(low_am_percent(audio, sr), 2),
"flatmod_pct": round(flatmod_percent(audio, sr), 2),
"dur_s": round(len(audio) / sr, 2)}
if scorer is not None:
row["utmos"] = round(scorer.score(audio, sr)["utmos"], 3)
rows.append(row)
all_rows[d] = rows
def m(k):
vals = [r[k] for r in rows if r.get(k) == r.get(k)]
return sum(vals) / len(vals) if vals else float("nan")
print(f"{Path(d).name:28s} {len(rows):3d} "
f"{m('utmos') if scorer else float('nan'):6.2f} "
f"{m('hf_pct'):6.2f} {m('flutter_pct'):6.1f} {m('jitter_pct'):6.2f} "
f"{m('am_pct'):6.1f} {m('am_peak_hz'):5.1f} {m('lowam_pct'):6.1f} "
f"{m('flatmod_pct'):6.1f}")
if args.out:
Path(args.out).write_text(json.dumps(all_rows, indent=2))
print(f"-> {args.out}")
if __name__ == "__main__":
main()
|