Spaces:
Build error
Build error
File size: 10,843 Bytes
7d5f092 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 | """FEATURE EXTRACTION LAYER
Parselmouth -> Formants, Pitch, Voice Quality
OpenSMILE -> 6373 features (eGeMAPSv02, ComParE_2016)
librosa -> Spectral, Rhythm, MFCC
"""
from __future__ import annotations
import logging
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any
import numpy as np
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Parselmouth: Formants, Pitch, Voice Quality
# ---------------------------------------------------------------------------
@dataclass
class FormantData:
f1_mean: float
f2_mean: float
f3_mean: float
f4_mean: float
f1_trajectory: list[float]
f2_trajectory: list[float]
f3_trajectory: list[float]
bandwidth_f1: float
bandwidth_f2: float
vowel_space_area: float
@dataclass
class PitchData:
mean_f0: float
min_f0: float
max_f0: float
std_f0: float
pitch_range: float
pitch_contour: list[float]
voiced_fraction: float
@dataclass
class VoiceQualityData:
hnr: float # Harmonics-to-noise ratio
jitter_local: float
jitter_rap: float
shimmer_local: float
shimmer_apq3: float
mean_intensity: float
intensity_std: float
spectral_tilt: float
cpp: float # Cepstral Peak Prominence
@dataclass
class ParselmouthFeatures:
formants: FormantData
pitch: PitchData
voice_quality: VoiceQualityData
def extract_parselmouth(audio_path: str | Path, gender: str = "neutral") -> ParselmouthFeatures:
"""Extract formants, pitch, and voice quality using Parselmouth (Praat)."""
import parselmouth
from parselmouth.praat import call
snd = parselmouth.Sound(str(audio_path))
# Gender-specific formant ceiling
ceiling_map = {"male": 5000, "female": 5500, "child": 6500, "neutral": 5500}
max_formant = ceiling_map.get(gender, 5500)
# -- Formants --
formant_obj = call(snd, "To Formant (burg)", 0.0, 5, max_formant, 0.025, 50.0)
num_frames = call(formant_obj, "Get number of frames")
f1_vals, f2_vals, f3_vals, f4_vals = [], [], [], []
bw1_vals, bw2_vals = [], []
for i in range(1, num_frames + 1):
t = call(formant_obj, "Get time from frame number", i)
for fnum, store in [(1, f1_vals), (2, f2_vals), (3, f3_vals), (4, f4_vals)]:
v = call(formant_obj, "Get value at time", fnum, t, "hertz", "Linear")
if not np.isnan(v):
store.append(v)
bw1 = call(formant_obj, "Get bandwidth at time", 1, t, "hertz", "Linear")
bw2 = call(formant_obj, "Get bandwidth at time", 2, t, "hertz", "Linear")
if not np.isnan(bw1):
bw1_vals.append(bw1)
if not np.isnan(bw2):
bw2_vals.append(bw2)
def safe_mean(arr: list[float]) -> float:
return float(np.mean(arr)) if arr else 0.0
# Vowel space area (triangle: F1/F2 of /i/, /a/, /u/ approximated from extremes)
f1_arr, f2_arr = np.array(f1_vals or [0]), np.array(f2_vals or [0])
if len(f1_arr) > 2 and len(f2_arr) > 2:
corners = np.array([
[np.min(f1_arr), np.max(f2_arr)], # /i/ region
[np.max(f1_arr), np.mean(f2_arr)], # /a/ region
[np.min(f1_arr), np.min(f2_arr)], # /u/ region
])
vsa = 0.5 * abs(
(corners[1, 0] - corners[0, 0]) * (corners[2, 1] - corners[0, 1])
- (corners[2, 0] - corners[0, 0]) * (corners[1, 1] - corners[0, 1])
)
else:
vsa = 0.0
formants = FormantData(
f1_mean=safe_mean(f1_vals),
f2_mean=safe_mean(f2_vals),
f3_mean=safe_mean(f3_vals),
f4_mean=safe_mean(f4_vals),
f1_trajectory=f1_vals[:100], # cap for JSON
f2_trajectory=f2_vals[:100],
f3_trajectory=f3_vals[:100],
bandwidth_f1=safe_mean(bw1_vals),
bandwidth_f2=safe_mean(bw2_vals),
vowel_space_area=float(vsa),
)
# -- Pitch --
pitch_obj = call(snd, "To Pitch", 0.0, 75, 600)
f0_values = [
call(pitch_obj, "Get value at time", t, "hertz", "Linear")
for t in np.arange(0, snd.duration, 0.01)
]
f0_clean = [v for v in f0_values if not np.isnan(v) and v > 0]
total_frames_pitch = len(f0_values)
voiced_frames = len(f0_clean)
pitch = PitchData(
mean_f0=safe_mean(f0_clean),
min_f0=float(min(f0_clean)) if f0_clean else 0.0,
max_f0=float(max(f0_clean)) if f0_clean else 0.0,
std_f0=float(np.std(f0_clean)) if f0_clean else 0.0,
pitch_range=(max(f0_clean) - min(f0_clean)) if f0_clean else 0.0,
pitch_contour=f0_clean[:200],
voiced_fraction=voiced_frames / total_frames_pitch if total_frames_pitch > 0 else 0.0,
)
# -- Voice Quality --
point_process = call(snd, "To PointProcess (periodic, cc)", 75, 600)
jitter_local = call(point_process, "Get jitter (local)", 0, 0, 0.0001, 0.02, 1.3)
jitter_rap = call(point_process, "Get jitter (rap)", 0, 0, 0.0001, 0.02, 1.3)
shimmer_local = call([snd, point_process], "Get shimmer (local)", 0, 0, 0.0001, 0.02, 1.3, 1.6)
shimmer_apq3 = call([snd, point_process], "Get shimmer (apq3)", 0, 0, 0.0001, 0.02, 1.3, 1.6)
harmonicity = call(snd, "To Harmonicity (cc)", 0.01, 75, 0.1, 1.0)
hnr = call(harmonicity, "Get mean", 0, 0)
intensity_obj = call(snd, "To Intensity", 75, 0.0, "yes")
mean_intensity = call(intensity_obj, "Get mean", 0, 0, "dB")
std_intensity = call(intensity_obj, "Get standard deviation", 0, 0)
# Spectral tilt (slope of long-term spectrum)
spectrum = call(snd, "To Spectrum", "yes")
ltas = call(spectrum, "To Ltas (1-to-1)")
low_energy = call(ltas, "Get mean", 0, 1000, "dB")
high_energy = call(ltas, "Get mean", 1000, 4000, "dB")
spectral_tilt = low_energy - high_energy if not (np.isnan(low_energy) or np.isnan(high_energy)) else 0.0
# CPP approximation via power cepstrum
try:
pc = call(snd, "To PowerCepstrogram", 60, 0.002, 5000, 50)
cpps = call(pc, "Get CPPS", "no", 0.02, 0.0005, 60, 330, 0.05, "parabolic", 0.001, 0, "Exponential decay", "Robust slow")
cpp_val = cpps if not np.isnan(cpps) else 0.0
except Exception:
cpp_val = 0.0
voice_quality = VoiceQualityData(
hnr=hnr if not np.isnan(hnr) else 0.0,
jitter_local=jitter_local if not np.isnan(jitter_local) else 0.0,
jitter_rap=jitter_rap if not np.isnan(jitter_rap) else 0.0,
shimmer_local=shimmer_local if not np.isnan(shimmer_local) else 0.0,
shimmer_apq3=shimmer_apq3 if not np.isnan(shimmer_apq3) else 0.0,
mean_intensity=mean_intensity if not np.isnan(mean_intensity) else 0.0,
intensity_std=std_intensity if not np.isnan(std_intensity) else 0.0,
spectral_tilt=float(spectral_tilt),
cpp=float(cpp_val),
)
return ParselmouthFeatures(formants=formants, pitch=pitch, voice_quality=voice_quality)
# ---------------------------------------------------------------------------
# librosa: Spectral, Rhythm, MFCC
# ---------------------------------------------------------------------------
@dataclass
class LibrosaFeatures:
mfcc_mean: list[float]
mfcc_std: list[float]
spectral_centroid_mean: float
spectral_bandwidth_mean: float
spectral_rolloff_mean: float
spectral_contrast_mean: list[float]
spectral_flatness_mean: float
zero_crossing_rate_mean: float
rms_mean: float
rms_std: float
tempo: float
chroma_mean: list[float]
mel_spectrogram_db: list[list[float]] # downsampled for visualization
def extract_librosa(audio_path: str | Path) -> LibrosaFeatures:
"""Extract spectral, rhythm, and MFCC features using librosa."""
import librosa
y, sr = librosa.load(str(audio_path), sr=22050)
# MFCC
mfcc = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)
mfcc_mean = np.mean(mfcc, axis=1).tolist()
mfcc_std = np.std(mfcc, axis=1).tolist()
# Spectral features
cent = librosa.feature.spectral_centroid(y=y, sr=sr)
bw = librosa.feature.spectral_bandwidth(y=y, sr=sr)
rolloff = librosa.feature.spectral_rolloff(y=y, sr=sr)
contrast = librosa.feature.spectral_contrast(y=y, sr=sr)
flatness = librosa.feature.spectral_flatness(y=y)
zcr = librosa.feature.zero_crossing_rate(y)
rms = librosa.feature.rms(y=y)
# Tempo
tempo_val, _ = librosa.beat.beat_track(y=y, sr=sr)
tempo_scalar = float(tempo_val[0]) if hasattr(tempo_val, '__len__') else float(tempo_val)
# Chroma
chroma = librosa.feature.chroma_stft(y=y, sr=sr)
chroma_mean = np.mean(chroma, axis=1).tolist()
# Mel spectrogram (downsampled for JSON transport)
mel = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=64)
mel_db = librosa.power_to_db(mel, ref=np.max)
step = max(1, mel_db.shape[1] // 100)
mel_down = mel_db[:, ::step].tolist()
return LibrosaFeatures(
mfcc_mean=mfcc_mean,
mfcc_std=mfcc_std,
spectral_centroid_mean=float(np.mean(cent)),
spectral_bandwidth_mean=float(np.mean(bw)),
spectral_rolloff_mean=float(np.mean(rolloff)),
spectral_contrast_mean=np.mean(contrast, axis=1).tolist(),
spectral_flatness_mean=float(np.mean(flatness)),
zero_crossing_rate_mean=float(np.mean(zcr)),
rms_mean=float(np.mean(rms)),
rms_std=float(np.std(rms)),
tempo=tempo_scalar,
chroma_mean=chroma_mean,
mel_spectrogram_db=mel_down,
)
# ---------------------------------------------------------------------------
# OpenSMILE: 6373 features (ComParE_2016)
# ---------------------------------------------------------------------------
@dataclass
class OpenSmileFeatures:
feature_set: str
feature_count: int
features: dict[str, float]
def extract_opensmile(audio_path: str | Path, feature_set: str = "eGeMAPSv02") -> OpenSmileFeatures | None:
"""Extract acoustic features using openSMILE."""
try:
import opensmile
feature_sets = {
"eGeMAPSv02": opensmile.FeatureSet.eGeMAPSv02,
"ComParE_2016": opensmile.FeatureSet.ComParE_2016,
}
fs = feature_sets.get(feature_set, opensmile.FeatureSet.eGeMAPSv02)
smile = opensmile.Smile(feature_set=fs, feature_level=opensmile.FeatureLevel.Functionals)
df = smile.process_file(str(audio_path))
features = {col: float(df[col].iloc[0]) for col in df.columns}
return OpenSmileFeatures(
feature_set=feature_set,
feature_count=len(features),
features=features,
)
except ImportError:
logger.warning("opensmile not installed, skipping")
return None
except Exception as exc:
logger.warning("openSMILE extraction failed: %s", exc)
return None
|