Spaces:
Sleeping
Sleeping
Commit ·
f95b090
1
Parent(s): 04f3da2
Implement Conservative Consensus: SNR Analysis and Human Liveness Veto to eliminate false positives
Browse files- app/infer.py +109 -53
app/infer.py
CHANGED
|
@@ -29,6 +29,30 @@ class VoiceClassifier:
|
|
| 29 |
print(f"Error loading model: {e}")
|
| 30 |
self.model = None
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
def predict(self, waveform: torch.Tensor, language: str = "Unknown"):
|
| 33 |
if self.model is None:
|
| 34 |
return {"error": "Model not loaded"}
|
|
@@ -38,9 +62,12 @@ class VoiceClassifier:
|
|
| 38 |
wav_np = waveform.squeeze().cpu().numpy()
|
| 39 |
sr = 16000
|
| 40 |
|
| 41 |
-
# --- ADVANCED FEATURE EXTRACTION ---
|
| 42 |
t0 = time.time()
|
| 43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
# A. Pitch Analysis
|
| 45 |
f0, voiced_flag, voiced_probs = librosa.pyin(
|
| 46 |
wav_np, fmin=librosa.note_to_hz('C2'), fmax=librosa.note_to_hz('C7'), sr=sr
|
|
@@ -48,27 +75,31 @@ class VoiceClassifier:
|
|
| 48 |
f0_clean = f0[~np.isnan(f0)]
|
| 49 |
pitch_var = np.std(f0_clean) if len(f0_clean) > 0 else 0.0
|
| 50 |
|
| 51 |
-
# B. Spectral Flatness
|
| 52 |
flatness = np.mean(librosa.feature.spectral_flatness(y=wav_np))
|
| 53 |
|
| 54 |
-
# C. RMS Energy Variance
|
| 55 |
rms = librosa.feature.rms(y=wav_np)[0]
|
| 56 |
-
rms_var = np.std(rms) / (np.mean(rms) + 1e-6)
|
| 57 |
|
| 58 |
-
# D.
|
| 59 |
-
|
| 60 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
|
| 62 |
-
# --- TEMPORAL CONSISTENCY
|
| 63 |
-
chunk_size = 2 * sr
|
| 64 |
-
stride = 1 * sr
|
| 65 |
chunks = []
|
| 66 |
for i in range(0, len(wav_np) - chunk_size + 1, stride):
|
| 67 |
chunks.append(wav_np[i : i + chunk_size])
|
| 68 |
-
|
| 69 |
-
# If audio too short for stride, just use whole thing
|
| 70 |
-
if not chunks:
|
| 71 |
-
chunks = [wav_np]
|
| 72 |
|
| 73 |
chunk_probs = []
|
| 74 |
for chunk in chunks:
|
|
@@ -77,61 +108,86 @@ class VoiceClassifier:
|
|
| 77 |
with torch.no_grad():
|
| 78 |
outputs = self.model(**inputs)
|
| 79 |
probs = torch.softmax(outputs.logits, dim=-1)
|
| 80 |
-
chunk_probs.append(probs[0][0].item()) #
|
| 81 |
|
| 82 |
-
#
|
| 83 |
-
# We take the MAX probability across chunks to catch 'slips' in AI generation
|
| 84 |
prob_fake = np.max(chunk_probs)
|
| 85 |
-
prob_real = 1.0 - prob_fake
|
| 86 |
|
| 87 |
t1 = time.time()
|
| 88 |
-
print(f"DEBUG: Analysis took {t1 - t0:.3f}s.
|
| 89 |
-
print(f"DEBUG: Features -
|
| 90 |
|
| 91 |
-
# ---
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
#
|
| 99 |
-
|
| 100 |
-
if
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
#
|
| 108 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 109 |
is_english = language.lower() in ["english", "en"]
|
| 110 |
-
if
|
| 111 |
-
|
| 112 |
|
| 113 |
# Construct Explanation
|
| 114 |
if prediction == "AI_GENERATED":
|
| 115 |
-
reasons =
|
| 116 |
-
if
|
| 117 |
-
|
| 118 |
-
if not reasons: reasons.append("digital vocoder artifacts")
|
| 119 |
-
explanation = f"AI detected with {confidence*100:.1f}% confidence. Evidence: {', '.join(reasons)}."
|
| 120 |
else:
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
explanation = f"Audio likely Human ({confidence*100:.1f}%). Detected natural speech fluctuations despite localized artifacts."
|
| 125 |
|
| 126 |
return {
|
| 127 |
"prediction": prediction,
|
| 128 |
-
"probability_ai": float(f"{
|
| 129 |
"confidence": float(f"{confidence:.4f}"),
|
| 130 |
"features": {
|
| 131 |
"pitch_variance": float(f"{pitch_var:.2f}"),
|
| 132 |
-
"
|
| 133 |
-
"
|
| 134 |
-
"zcr_variance": float(f"{zcr_var:.4f}")
|
| 135 |
},
|
| 136 |
"explanation": explanation
|
| 137 |
}
|
|
|
|
| 29 |
print(f"Error loading model: {e}")
|
| 30 |
self.model = None
|
| 31 |
|
| 32 |
+
def calculate_snr(self, audio_np):
|
| 33 |
+
"""
|
| 34 |
+
Estimate Signal-to-Noise Ratio (SNR) in dB.
|
| 35 |
+
Assumes the quietest 10% of frames represent the noise floor.
|
| 36 |
+
"""
|
| 37 |
+
try:
|
| 38 |
+
# Frame-based RMS energy
|
| 39 |
+
rms = librosa.feature.rms(y=audio_np)[0]
|
| 40 |
+
if len(rms) < 10: return 50.0 # Too short, assume clean
|
| 41 |
+
|
| 42 |
+
# Sort RMS values to find noise floor
|
| 43 |
+
sorted_rms = np.sort(rms)
|
| 44 |
+
noise_len = max(1, int(0.1 * len(rms)))
|
| 45 |
+
noise_floor_rms = np.mean(sorted_rms[:noise_len]) + 1e-9
|
| 46 |
+
|
| 47 |
+
# Signal RMS (approximate as top 50% energy average)
|
| 48 |
+
signal_len = max(1, int(0.5 * len(rms)))
|
| 49 |
+
signal_rms = np.mean(sorted_rms[-signal_len:])
|
| 50 |
+
|
| 51 |
+
snr = 20 * np.log10(signal_rms / noise_floor_rms)
|
| 52 |
+
return snr
|
| 53 |
+
except Exception:
|
| 54 |
+
return 30.0 # Default to decent SNR if calculation fails
|
| 55 |
+
|
| 56 |
def predict(self, waveform: torch.Tensor, language: str = "Unknown"):
|
| 57 |
if self.model is None:
|
| 58 |
return {"error": "Model not loaded"}
|
|
|
|
| 62 |
wav_np = waveform.squeeze().cpu().numpy()
|
| 63 |
sr = 16000
|
| 64 |
|
|
|
|
| 65 |
t0 = time.time()
|
| 66 |
|
| 67 |
+
# --- SIGNAL QUALITY CHECKS ---
|
| 68 |
+
snr_db = self.calculate_snr(wav_np)
|
| 69 |
+
|
| 70 |
+
# --- ADVANCED FEATURE EXTRACTION ---
|
| 71 |
# A. Pitch Analysis
|
| 72 |
f0, voiced_flag, voiced_probs = librosa.pyin(
|
| 73 |
wav_np, fmin=librosa.note_to_hz('C2'), fmax=librosa.note_to_hz('C7'), sr=sr
|
|
|
|
| 75 |
f0_clean = f0[~np.isnan(f0)]
|
| 76 |
pitch_var = np.std(f0_clean) if len(f0_clean) > 0 else 0.0
|
| 77 |
|
| 78 |
+
# B. Spectral Flatness
|
| 79 |
flatness = np.mean(librosa.feature.spectral_flatness(y=wav_np))
|
| 80 |
|
| 81 |
+
# C. RMS Energy Variance
|
| 82 |
rms = librosa.feature.rms(y=wav_np)[0]
|
| 83 |
+
rms_var = np.std(rms) / (np.mean(rms) + 1e-6)
|
| 84 |
|
| 85 |
+
# D. Liveness (Pause) Detection
|
| 86 |
+
# Count distinct silent intervals (>0.1s)
|
| 87 |
+
silent_intervals = librosa.effects.split(wav_np, top_db=20, frame_length=2048, hop_length=512)
|
| 88 |
+
num_pauses = 0
|
| 89 |
+
if len(silent_intervals) > 1:
|
| 90 |
+
# Calculate gaps between speech segments
|
| 91 |
+
for i in range(len(silent_intervals)-1):
|
| 92 |
+
gap_samples = silent_intervals[i+1][0] - silent_intervals[i][1]
|
| 93 |
+
if gap_samples > sr * 0.1: # >100ms
|
| 94 |
+
num_pauses += 1
|
| 95 |
|
| 96 |
+
# --- TEMPORAL CONSISTENCY ---
|
| 97 |
+
chunk_size = 2 * sr
|
| 98 |
+
stride = 1 * sr
|
| 99 |
chunks = []
|
| 100 |
for i in range(0, len(wav_np) - chunk_size + 1, stride):
|
| 101 |
chunks.append(wav_np[i : i + chunk_size])
|
| 102 |
+
if not chunks: chunks = [wav_np]
|
|
|
|
|
|
|
|
|
|
| 103 |
|
| 104 |
chunk_probs = []
|
| 105 |
for chunk in chunks:
|
|
|
|
| 108 |
with torch.no_grad():
|
| 109 |
outputs = self.model(**inputs)
|
| 110 |
probs = torch.softmax(outputs.logits, dim=-1)
|
| 111 |
+
chunk_probs.append(probs[0][0].item()) # Prob fake
|
| 112 |
|
| 113 |
+
# Initial Raw Confidence (Max across chunks)
|
|
|
|
| 114 |
prob_fake = np.max(chunk_probs)
|
|
|
|
| 115 |
|
| 116 |
t1 = time.time()
|
| 117 |
+
print(f"DEBUG: Analysis took {t1 - t0:.3f}s. Raw prob_fake: {prob_fake:.4f}")
|
| 118 |
+
print(f"DEBUG: Features - SNR: {snr_db:.1f}dB, Pauses: {num_pauses}, PitchVar: {pitch_var:.1f}, Flatness: {flatness:.4f}")
|
| 119 |
|
| 120 |
+
# --- CONSERVATIVE CONSENSUS LOGIC ---
|
| 121 |
+
|
| 122 |
+
# 1. Initialize Flags (Relaxed thresholds)
|
| 123 |
+
ai_flags = []
|
| 124 |
+
human_flags = []
|
| 125 |
+
|
| 126 |
+
# AI Indicators
|
| 127 |
+
if pitch_var < 10.0: ai_flags.append("Low pitch variance") # Relaxed from 15
|
| 128 |
+
if flatness < 0.002: ai_flags.append("Unnatural spectral flatness") # Relaxed from 0.005
|
| 129 |
+
if rms_var < 0.1: ai_flags.append("Robotic volume consistency")
|
| 130 |
+
|
| 131 |
+
# Human Indicators (VETO Power)
|
| 132 |
+
if snr_db < 15.0: human_flags.append("High Background Noise")
|
| 133 |
+
if num_pauses >= 2: human_flags.append("Natural breathing pauses")
|
| 134 |
+
if pitch_var > 35.0: human_flags.append("High expressive variation")
|
| 135 |
+
|
| 136 |
+
# 2. Apply Penalties / Vetoes
|
| 137 |
+
confidence_penalty = 1.0
|
| 138 |
+
|
| 139 |
+
# VETO 1: NOISE
|
| 140 |
+
# If noisy, the model's "Fake" detection is untrustworthy. Cap it.
|
| 141 |
+
if snr_db < 15.0:
|
| 142 |
+
print("DEBUG: Low SNR detected. Applying penalty.")
|
| 143 |
+
confidence_penalty *= 0.6 # Reduce confidence by 40%
|
| 144 |
+
|
| 145 |
+
# VETO 2: LIVENESS
|
| 146 |
+
if num_pauses >= 2 and prob_fake < 0.95:
|
| 147 |
+
print("DEBUG: Natural pauses detected. Applying penalty.")
|
| 148 |
+
confidence_penalty *= 0.8 # Reduce confidence by 20%
|
| 149 |
+
|
| 150 |
+
# Apply penalty to the probability of being fake
|
| 151 |
+
prob_fake_adjusted = prob_fake * confidence_penalty
|
| 152 |
+
|
| 153 |
+
# 3. Final Decision
|
| 154 |
+
# We demand HIGHER evidence for AI (Conservatism)
|
| 155 |
+
threshold = 0.65 # Default threshold (requires >65% certainty)
|
| 156 |
+
|
| 157 |
+
# If we have AI flags, we can trust the model more (lower threshold)
|
| 158 |
+
if len(ai_flags) >= 2:
|
| 159 |
+
threshold = 0.55
|
| 160 |
+
|
| 161 |
+
if prob_fake_adjusted > threshold:
|
| 162 |
+
prediction = "AI_GENERATED"
|
| 163 |
+
confidence = prob_fake_adjusted
|
| 164 |
+
else:
|
| 165 |
+
prediction = "HUMAN"
|
| 166 |
+
confidence = 1.0 - prob_fake_adjusted
|
| 167 |
+
|
| 168 |
+
# 4. Language Awareness Dampening
|
| 169 |
is_english = language.lower() in ["english", "en"]
|
| 170 |
+
if prediction == "AI_GENERATED" and not is_english:
|
| 171 |
+
confidence *= 0.9 # Extra caution for non-English
|
| 172 |
|
| 173 |
# Construct Explanation
|
| 174 |
if prediction == "AI_GENERATED":
|
| 175 |
+
reasons = ai_flags
|
| 176 |
+
if not reasons: reasons.append("high confidence from deepfake classifier")
|
| 177 |
+
explanation = f"AI detected ({confidence*100:.1f}%). Indicators: {', '.join(reasons)}."
|
|
|
|
|
|
|
| 178 |
else:
|
| 179 |
+
reasons = human_flags
|
| 180 |
+
if not reasons: reasons.append("insufficient evidence of synthesis")
|
| 181 |
+
explanation = f"Verified Human ({confidence*100:.1f}%). Evidence: {', '.join(reasons)}."
|
|
|
|
| 182 |
|
| 183 |
return {
|
| 184 |
"prediction": prediction,
|
| 185 |
+
"probability_ai": float(f"{prob_fake_adjusted:.4f}"),
|
| 186 |
"confidence": float(f"{confidence:.4f}"),
|
| 187 |
"features": {
|
| 188 |
"pitch_variance": float(f"{pitch_var:.2f}"),
|
| 189 |
+
"snr_db": float(f"{snr_db:.1f}"),
|
| 190 |
+
"pauses": num_pauses
|
|
|
|
| 191 |
},
|
| 192 |
"explanation": explanation
|
| 193 |
}
|