Spaces:
Sleeping
Sleeping
File size: 7,889 Bytes
ac9876a f8c1205 ac9876a f8c1205 ac9876a f8c1205 ac9876a f8c1205 ac9876a 91c5b8e ac9876a 91c5b8e ac9876a 91c5b8e f8c1205 91c5b8e f8c1205 91c5b8e ac9876a f8c1205 ac9876a f8c1205 ac9876a f8c1205 ac9876a f8c1205 ac9876a f8c1205 ac9876a f8c1205 ac9876a f8c1205 ac9876a f8c1205 ac9876a f8c1205 ac9876a f8c1205 ac9876a f8c1205 ac9876a | 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 | """
Scoring Engine
Advanced heuristic and logic layer for combining ML predictions with expert rules.
Implements a "Signal Detection" approach using Evidence Accumulation.
"""
import numpy as np
from typing import Dict, Tuple
class ScoringEngine:
"""
Bio-inspired Signal Detection System.
Accumulates evidence for "Busy" vs "Not Busy".
"""
def __init__(self):
# Evidence Weights (Log-Odds contributions)
# Positive = Evidence for BUSY
# Negative = Evidence for NOT BUSY
self.WEIGHTS = {
'explicit_busy': 6.0, # Strongest signal
'explicit_free': -4.0, # Strong negative signal (if we had it)
'traffic_noise': 3.0, # Strong context
'office_noise': 1.0, # Weak context
'rushed_speech': 1.5, # Medium context
'short_answers': 1.2, # Medium context
'deflection': 2.0, # Medium-Strong context
'latency': 0.5, # Weak context
'ml_model_factor': 0.5, # Multiplier for ML log-odds (reduced)
'emotion_stress': 2.5, # Paper 1: Strong indicator of busy state
'emotion_energy': 0.8, # Medium indicator
}
def _sigmoid(self, x: float) -> float:
"""Convert log-odds to probability (0-1)"""
return 1.0 / (1.0 + np.exp(-x))
def _logit(self, p: float) -> float:
"""Convert probability to log-odds (-inf to +inf)"""
p = np.clip(p, 0.01, 0.99) # Avoid inf
return np.log(p / (1.0 - p))
def calculate_score(
self,
audio_features: Dict[str, float],
text_features: Dict[str, float],
ml_probability: float
) -> Tuple[float, Dict]:
"""
Calculate Busy Score using Evidence Accumulation.
"""
evidence = 0.0
positive_evidence = 0.0
negative_evidence = 0.0
details = []
def add_evidence(points: float, label: str) -> None:
nonlocal evidence, positive_evidence, negative_evidence
evidence += points
if points >= 0:
positive_evidence += points
else:
negative_evidence += points
details.append(label)
# Check if user explicitly invited conversation (intent overrides context)
explicit_free = text_features.get('t0_explicit_free', 0.0)
intent_overrides_context = explicit_free > 0.5
# --- 1. Text Evidence (Intent) ---
# Explicit Busy
explicit = text_features.get('t1_explicit_busy', 0.0)
if explicit > 0.5:
points = self.WEIGHTS['explicit_busy'] * explicit
add_evidence(points, f"Explicit Intent (+{points:.1f})")
# Explicit Free (negative evidence)
if explicit_free > 0.5:
points = self.WEIGHTS['explicit_free'] * explicit_free
add_evidence(points, f"Explicit Free ({points:.1f})")
# Short Answers (Brevity) - only counts when there's other busy evidence
short_ratio = text_features.get('t3_short_ratio', 0.0)
if short_ratio > 0.3:
deflection = text_features.get('t6_deflection', 0.0)
time_pressure = text_features.get('t5_time_pressure', 0.0)
busy_context = (explicit > 0.5) or (deflection > 0.1) or (time_pressure > 0.1)
if intent_overrides_context:
points = self.WEIGHTS['short_answers'] * short_ratio * 0.4
add_evidence(points, f"Brief Responses (+{points:.1f}, reduced - user invited talk)")
elif busy_context:
points = self.WEIGHTS['short_answers'] * short_ratio
add_evidence(points, f"Brief Responses (+{points:.1f})")
else:
details.append("Brief Responses (ignored - no busy evidence)")
# Deflection / Time Pressure
deflection = text_features.get('t6_deflection', 0.0)
if deflection > 0.1:
points = self.WEIGHTS['deflection'] * deflection
add_evidence(points, f"Deflection (+{points:.1f})")
# --- 2. Audio Evidence (Context) ---
# Traffic Noise (reduced when user explicitly invites talk)
traffic = audio_features.get('v2_noise_traffic', 0.0)
if traffic > 0.5:
points = self.WEIGHTS['traffic_noise'] * traffic
if intent_overrides_context:
points *= 0.3 # Strong availability signal overrides traffic context
add_evidence(points, f"Traffic Context (+{points:.1f}, reduced - user invited talk)")
else:
add_evidence(points, f"Traffic Context (+{points:.1f})")
# Speech Rate
rate = audio_features.get('v3_speech_rate', 0.0)
if rate > 3.5: # Fast speech
points = self.WEIGHTS['rushed_speech']
add_evidence(points, f"Rushed Speech (+{points:.1f})")
elif rate < 1.0: # Very slow speech (might be distracted)
pass # Neutral for now
# Energy/Pitch (Stress)
pitch_std = audio_features.get('v5_pitch_std', 0.0)
if pitch_std > 80.0: # High variation
add_evidence(0.5, "Voice Stress (+0.5)")
# --- 2b. Emotion Evidence (if present) ---
emotion_stress = audio_features.get('v11_emotion_stress', 0.0)
if emotion_stress > 0.6:
points = self.WEIGHTS['emotion_stress'] * emotion_stress
add_evidence(points, f"Emotional Stress (+{points:.1f})")
emotion_energy = audio_features.get('v12_emotion_energy', 0.0)
if emotion_energy > 0.7:
points = self.WEIGHTS['emotion_energy'] * emotion_energy
add_evidence(points, f"High Energy (+{points:.1f})")
# --- 3. Machine Learning Evidence (Baseline) ---
# Convert Model Probability to Log-Odds Evidence
ml_evidence = self._logit(ml_probability)
weighted_ml_evidence = ml_evidence * self.WEIGHTS['ml_model_factor']
add_evidence(weighted_ml_evidence, f"ML Baseline ({weighted_ml_evidence:+.1f})")
# --- 4. Final Calculation ---... more value ot the voice features, especially the emotional ones
# Sigmoid converts total evidence back to 0-1 probability
final_score = self._sigmoid(evidence)
breakdown = {
'total_evidence': evidence,
'positive_evidence': positive_evidence,
'negative_evidence': negative_evidence,
'details': details,
'ml_contribution': weighted_ml_evidence
}
return final_score, breakdown
def get_confidence(self, score: float, breakdown: Dict) -> float:
"""
Calculate confidence based on EVIDENCE MAGNITUDE.
Strong evidence (positive or negative) = High Confidence.
Zero evidence = Low Confidence.
"""
positive_evidence = breakdown.get('positive_evidence', 0.0)
negative_evidence = abs(breakdown.get('negative_evidence', 0.0))
total_strength = positive_evidence + negative_evidence
conflict = min(positive_evidence, negative_evidence)
conflict_ratio = conflict / (total_strength + 1e-6)
# We model confidence as a sigmoid of absolute evidence
# |evidence| = 0 -> Confidence = 0.5 (Unsure) ?? No, 0.0 (Total Guess)
# But standard sigmoid(0) is 0.5.
# We want 0->0, values->1.
# Use a scaling factor. Evidence of > 3.0 is very strong.
# tanh is good: tanh(0)=0, tanh(3)≈0.995
base_confidence = np.tanh(total_strength / 2.0)
confidence = base_confidence * (1.0 - conflict_ratio)
return float(confidence)
|