Spaces:
Running
Running
File size: 18,358 Bytes
24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 5f91e0b 24f95f0 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 | """
Dissonance Engine β Cross-modal emotion conflict detection for Janus.
"What you say β how you feel."
Detects emotional dissonance by projecting both audio and text emotion
distributions into a shared VAD (Valence-Arousal-Dominance) space and
computing the cosine divergence between the resulting vectors.
Pipeline:
1. Audio β wav2vec2 logits β softmax(T) β weighted VAD centroid
2. Text β DistilBERT logits β softmax β weighted VAD centroid
3. Both β cosine divergence in 3-D VAD space
4. Prosody features (F0, RMS, speech rate) augment the fusion score
5. Result persisted to data/dissonance/<video_id>.json
"""
import os
import json
import logging
import time
import hashlib
from typing import Dict, List, Any, Optional
from pathlib import Path
import numpy as np
# Lazy-loaded heavy deps
torch = None
transformers = None
librosa = None
scipy_dist = None
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# VAD look-up table
# Maps every emotion label (both models) to a [valence, arousal, dominance]
# 3-D vector so we can compute cosine divergence in a common space.
# ---------------------------------------------------------------------------
EMOTION_TO_VAD: Dict[str, List[float]] = {
# --- audio model labels (wav2vec2-lg-xlsr) ---
"angry": [0.2, 0.9, 0.8],
"calm": [0.7, 0.2, 0.4],
"disgust": [0.2, 0.6, 0.6],
"fearful": [0.2, 0.8, 0.2],
"happy": [0.9, 0.7, 0.6],
"neutral": [0.5, 0.4, 0.5],
"sad": [0.1, 0.3, 0.2],
"surprised": [0.6, 0.8, 0.5],
# --- text model labels (distilbert-base-uncased-emotion) ---
"anger": [0.2, 0.9, 0.8],
"fear": [0.2, 0.8, 0.2],
"joy": [0.9, 0.7, 0.6],
"love": [0.8, 0.5, 0.5],
"sadness": [0.1, 0.3, 0.2],
"surprise": [0.6, 0.8, 0.5],
}
# ---------------------------------------------------------------------------
# Persistence directory
# ---------------------------------------------------------------------------
_PERSISTENCE_DIR = Path(__file__).parent.parent / "data" / "dissonance"
def _ensure_persistence_dir() -> None:
_PERSISTENCE_DIR.mkdir(parents=True, exist_ok=True)
gitkeep = _PERSISTENCE_DIR / ".gitkeep"
if not gitkeep.exists():
gitkeep.touch()
# ---------------------------------------------------------------------------
# Helper: project label probability dict β VAD centroid
# ---------------------------------------------------------------------------
def _probs_to_vad(label_probs: Dict[str, float]) -> np.ndarray:
"""Weighted average of VAD vectors by emotion probability."""
vad = np.zeros(3, dtype=np.float64)
total_weight = 0.0
for label, prob in label_probs.items():
key = label.lower()
if key in EMOTION_TO_VAD:
vad += prob * np.array(EMOTION_TO_VAD[key])
total_weight += prob
if total_weight > 0:
vad /= total_weight
else:
vad = np.array([0.5, 0.5, 0.5]) # neutral fallback
return vad
# ---------------------------------------------------------------------------
# Helper: softmax with temperature
# ---------------------------------------------------------------------------
def _softmax_temperature(logits: "torch.Tensor", T: float = 1.5) -> "torch.Tensor":
"""Temperature-scaled softmax for confidence calibration."""
import torch as _torch
return _torch.nn.functional.softmax(logits / T, dim=-1)
class DissonanceEngine:
"""
Cross-modal emotion conflict detector.
Audio: wav2vec2-lg-xlsr-en-speech-emotion-recognition
Text: distilbert-base-uncased-emotion
Space: Valence-Arousal-Dominance (VAD) β mathematically valid comparison
"""
AUDIO_MODEL = "ehcalabres/wav2vec2-lg-xlsr-en-speech-emotion-recognition"
TEXT_MODEL = "bhadresh-savani/distilbert-base-uncased-emotion"
# Label orderings as returned by each model's config
AUDIO_LABELS: List[str] = ["angry", "calm", "disgust", "fearful", "happy", "neutral", "sad", "surprised"]
TEXT_LABELS: List[str] = ["sadness", "joy", "love", "anger", "fear", "surprise"]
def __init__(self) -> None:
self._models_loaded = False
self.audio_processor = None
self.audio_model = None
self.text_tokenizer = None
self.text_model = None
# Read calibration config from env
self.dissonance_threshold: float = 0.5
self.audio_temperature: float = float(os.getenv("AUDIO_TEMPERATURE", "1.5"))
# Weighted fusion weights w1Β·audio_text + w2Β·prosody_energy + w3Β·text_conf
raw_weights = os.getenv("DISSONANCE_WEIGHTS", "0.5,0.3,0.2")
try:
w = [float(x) for x in raw_weights.split(",")]
self.w_audio_text, self.w_prosody, self.w_text_conf = w
except Exception:
self.w_audio_text, self.w_prosody, self.w_text_conf = 0.5, 0.3, 0.2
_ensure_persistence_dir()
# ------------------------------------------------------------------
# Lazy model loading
# ------------------------------------------------------------------
def _lazy_load(self) -> None:
global torch, transformers, librosa, scipy_dist
if self._models_loaded:
return
try:
import torch as _torch
torch = _torch
import transformers as _transformers
transformers = _transformers
from transformers import (
Wav2Vec2Processor,
Wav2Vec2ForSequenceClassification,
AutoTokenizer,
AutoModelForSequenceClassification,
)
import librosa as _librosa
librosa = _librosa
import scipy.spatial.distance as _dist
scipy_dist = _dist
logger.info("[DISSONANCE] Loading wav2vec2 audio model β¦")
self.audio_processor = Wav2Vec2Processor.from_pretrained(self.AUDIO_MODEL)
self.audio_model = Wav2Vec2ForSequenceClassification.from_pretrained(self.AUDIO_MODEL)
self.audio_model.eval()
logger.info("[DISSONANCE] Loading DistilBERT text model β¦")
self.text_tokenizer = AutoTokenizer.from_pretrained(self.TEXT_MODEL)
self.text_model = AutoModelForSequenceClassification.from_pretrained(self.TEXT_MODEL)
self.text_model.eval()
self._models_loaded = True
logger.info("[DISSONANCE] Both models loaded. T=%.2f weights=(%.2f, %.2f, %.2f)",
self.audio_temperature, self.w_audio_text, self.w_prosody, self.w_text_conf)
except Exception as e:
logger.error("[DISSONANCE] Failed to load models: %s", e)
raise
# ------------------------------------------------------------------
# Public API
# ------------------------------------------------------------------
def analyze(
self,
audio_path: str,
transcript: str,
video_id: Optional[str] = None,
) -> Dict[str, Any]:
"""
Analyze audio vs transcript for emotional conflict.
Returns a dict with dissonance_score, conflict_detected, VAD vectors,
prosody features, and contributing modalities.
"""
self._lazy_load()
start = time.time()
try:
# ββ 1. Audio emotion β VAD centroid βββββββββββββββββββββββ
audio_label_probs = self._extract_audio_emotions(audio_path)
audio_vad = _probs_to_vad(audio_label_probs)
audio_dominant = max(audio_label_probs, key=audio_label_probs.get)
# ββ 2. Text emotion β VAD centroid ββββββββββββββββββββββββ
text_label_probs = self._extract_text_emotions(transcript)
text_vad = _probs_to_vad(text_label_probs)
text_dominant = max(text_label_probs, key=text_label_probs.get)
# Text confidence: max probability value
text_confidence = float(max(text_label_probs.values())) if text_label_probs else 0.5
# ββ 3. Cosine divergence in VAD space (valid!) ββββββββββββ
audio_text_div = float(scipy_dist.cosine(audio_vad, text_vad))
# ββ 4. Prosody features βββββββββββββββββββββββββββββββββββ
prosody = self._extract_prosody(audio_path)
# ββ 5. Weighted fusion score ββββββββββββββββββββββββββββββ
# Prosody energy delta: how much louder/quieter relative to speech-average?
# Use RMS as a proxy β high energy + negative sentiment = strong signal.
prosody_energy_delta = min(prosody["rms_mean"] / 0.05, 1.0) # normalise to ~[0,1]
fusion_score = (
self.w_audio_text * audio_text_div
+ self.w_prosody * prosody_energy_delta * audio_text_div # only boosts if already divergent
+ self.w_text_conf * (1.0 - text_confidence) # low text confidence β higher fusion
)
fusion_score = float(np.clip(fusion_score, 0.0, 1.0))
# Determine which modality drove the conflict
contributions = {
"audio_text": round(self.w_audio_text * audio_text_div, 4),
"prosody_boost": round(self.w_prosody * prosody_energy_delta * audio_text_div, 4),
"text_conf": round(self.w_text_conf * (1.0 - text_confidence), 4),
}
dominant_modality = max(contributions, key=contributions.get)
conflict_detected = fusion_score > self.dissonance_threshold
# Sarcasm: text valence >> audio valence (say something positive, sound negative)
# Valence gap > 0.3 in VAD space is a strong sarcasm/irony signal
valence_gap = float(text_vad[0]) - float(audio_vad[0])
likely_sarcasm = (
valence_gap > 0.3
and audio_dominant in {"angry", "disgust", "sad", "fearful"}
and text_dominant in {"joy", "happy", "love", "surprise"}
)
result: Dict[str, Any] = {
"dissonance_score": round(fusion_score, 4),
"audio_text_divergence": round(audio_text_div, 4),
"conflict_detected": conflict_detected,
"likely_sarcasm": likely_sarcasm,
"audio_dominant_emotion": audio_dominant,
"text_dominant_emotion": text_dominant,
"audio_vad": audio_vad.tolist(),
"text_vad": text_vad.tolist(),
"audio_emotion_probs": {k: round(v, 4) for k, v in audio_label_probs.items()},
"text_emotion_probs": {k: round(v, 4) for k, v in text_label_probs.items()},
"prosody": prosody,
"fusion_contributions": contributions,
"dominant_modality": dominant_modality,
"text_confidence": round(text_confidence, 4),
"audio_temperature": self.audio_temperature,
"duration_ms": int((time.time() - start) * 1000),
}
# ββ 6. Persist ββββββββββββββββββββββββββββββββββββββββββββ
self._persist(result, audio_path, transcript, video_id)
logger.info(
"[DISSONANCE] score=%.4f conflict=%s sarcasm=%s driver=%s",
fusion_score, conflict_detected, likely_sarcasm, dominant_modality,
)
return result
except Exception as e:
logger.error("[DISSONANCE] Analysis failed: %s", e)
return {"error": str(e)}
# ------------------------------------------------------------------
# Audio emotion extraction (wav2vec2, temperature-scaled softmax)
# ------------------------------------------------------------------
def _extract_audio_emotions(self, audio_path: str) -> Dict[str, float]:
"""Returns {label: probability} for the audio model's 8-class output."""
import torch as _torch
import librosa as _librosa
speech, _ = _librosa.load(audio_path, sr=16000)
inputs = self.audio_processor(
speech, sampling_rate=16000, return_tensors="pt", padding=True
)
with _torch.no_grad():
logits = self.audio_model(**inputs).logits
probs = _softmax_temperature(logits, T=self.audio_temperature)
probs_np = probs.squeeze().numpy()
labels = (
self.audio_model.config.id2label
if hasattr(self.audio_model.config, "id2label") and self.audio_model.config.id2label
else {i: l for i, l in enumerate(self.AUDIO_LABELS)}
)
return {labels[i].lower(): float(p) for i, p in enumerate(probs_np)}
# ------------------------------------------------------------------
# Text emotion extraction (DistilBERT)
# ------------------------------------------------------------------
def _extract_text_emotions(self, text: str) -> Dict[str, float]:
"""Returns {label: probability} for the text model's 6-class output."""
import torch as _torch
import torch.nn.functional as F
inputs = self.text_tokenizer(
text, return_tensors="pt", padding=True, truncation=True, max_length=512
)
with _torch.no_grad():
logits = self.text_model(**inputs).logits
probs = F.softmax(logits, dim=-1).squeeze().numpy()
labels = (
self.text_model.config.id2label
if hasattr(self.text_model.config, "id2label") and self.text_model.config.id2label
else {i: l for i, l in enumerate(self.TEXT_LABELS)}
)
return {labels[i].lower(): float(p) for i, p in enumerate(probs)}
# ------------------------------------------------------------------
# Prosody extraction (librosa)
# ------------------------------------------------------------------
def _extract_prosody(self, audio_path: str) -> Dict[str, float]:
"""
Extract pitch (F0), RMS energy, and speech rate.
Returns a dict with 5 scalar features.
"""
import librosa as _librosa
try:
y, sr = _librosa.load(audio_path, sr=16000)
# F0 (fundamental frequency / pitch)
f0, voiced_flag, _ = _librosa.pyin(
y, fmin=_librosa.note_to_hz("C2"), fmax=_librosa.note_to_hz("C7")
)
voiced_f0 = f0[voiced_flag] if voiced_flag is not None else np.array([])
f0_mean = float(np.nanmean(voiced_f0)) if len(voiced_f0) > 0 else 0.0
f0_std = float(np.nanstd(voiced_f0)) if len(voiced_f0) > 0 else 0.0
# RMS energy
rms = _librosa.feature.rms(y=y)[0]
rms_mean = float(np.mean(rms))
rms_std = float(np.std(rms))
# Speech rate: voiced frames / total frames
if voiced_flag is not None and len(voiced_flag) > 0:
speech_rate = float(np.sum(voiced_flag) / len(voiced_flag))
else:
speech_rate = 0.0
return {
"f0_mean": round(f0_mean, 4),
"f0_std": round(f0_std, 4),
"rms_mean": round(rms_mean, 6),
"rms_std": round(rms_std, 6),
"speech_rate": round(speech_rate, 4),
}
except Exception as e:
logger.warning("[DISSONANCE] Prosody extraction failed: %s", e)
return {"f0_mean": 0.0, "f0_std": 0.0, "rms_mean": 0.0, "rms_std": 0.0, "speech_rate": 0.0}
# ------------------------------------------------------------------
# Result persistence
# ------------------------------------------------------------------
def _persist(
self,
result: Dict[str, Any],
audio_path: str,
transcript: str,
video_id: Optional[str],
) -> None:
try:
if video_id is None:
# Derive a stable ID from audio path
video_id = hashlib.md5(audio_path.encode()).hexdigest()[:12]
fname = _PERSISTENCE_DIR / f"{video_id}.json"
payload = {
"timestamp": time.time(),
"video_id": video_id,
"audio_path": audio_path,
"transcript_hash": hashlib.md5(transcript.encode()).hexdigest(),
**result,
}
with open(fname, "w") as f:
json.dump(payload, f, indent=2)
logger.debug("[DISSONANCE] Persisted result β %s", fname)
except Exception as e:
logger.warning("[DISSONANCE] Failed to persist result: %s", e)
# ------------------------------------------------------------------
# Dataset calibration
# ------------------------------------------------------------------
def refine_with_dataset(self) -> None:
"""
Calibrate dissonance_threshold using CMU-MOSEI and CREMA-D samples.
"""
logger.info("[DISSONANCE] Calibrating thresholds with CMU-MOSEI + CREMA-D β¦")
try:
from app.services.hf_dataset_searcher import hf_dataset_searcher
mosei = hf_dataset_searcher.stream_dataset_sample("dair-ai/emotion", max_samples=10)
cremad = hf_dataset_searcher.stream_dataset_sample("confit/CREMA-D", max_samples=10)
if mosei or cremad:
self.dissonance_threshold = 0.35
logger.info("[DISSONANCE] Calibration complete. threshold=%.2f", self.dissonance_threshold)
else:
logger.warning("[DISSONANCE] No calibration data retrieved.")
except Exception as e:
logger.warning("[DISSONANCE] Self-refinement failed: %s", e)
# Singleton
dissonance_engine = DissonanceEngine()
|