Upload folder using huggingface_hub

metrics/samplerate_metric.py

@@ -25,26 +25,35 @@ def get_mic_sr(audio_path):
         # Compute magnitude spectrum
         S_full, phase = librosa.magphase(librosa.stft(y))
         
-        # Calculate power spectrum
-        power = S_full**2
+        # Max Hold Spectrum (Peak detection across time)
+        # Instead of average, we take the MAX magnitude at each frequency bin across all time frames.
+        # This aligns with "Max peaks on the spectrogram".
+        S_max = np.max(S_full, axis=1)
         
-        # Average power across time
-        avg_power = np.mean(power, axis=1)
-        
-        # Cumulative distribution of power
-        cumulative_power = np.cumsum(avg_power)
-        total_power = cumulative_power[-1]
-        
-        if total_power == 0:
+        # Normalize to Max Peak (0 Reference)
+        S_ref = np.max(S_max)
+        if S_ref == 0:
             return 0
             
-        # Find frequency where 99% of power is contained
-        threshold = 0.99 * total_power
-        idx = np.searchsorted(cumulative_power, threshold)
+        # Convert to dB relative to peak
+        S_db = librosa.amplitude_to_db(S_max, ref=S_ref)
+        
+        # Threshold: Find the highest frequency that is within X dB of the peak.
+        # Typical noise floor might be -80dB or -90dB.
+        # We look for the "last" bin that is ABOVE the threshold.
+        threshold_db = -80.0
         
-        # Convert bin index to frequency
         fft_freqs = librosa.fft_frequencies(sr=sr)
-        cutoff_freq = fft_freqs[idx]
+        
+        # Find indices where signal > threshold
+        valid_indices = np.where(S_db > threshold_db)[0]
+        
+        if len(valid_indices) == 0:
+            return 0
+            
+        # The highest index with significant energy
+        last_idx = valid_indices[-1]
+        cutoff_freq = fft_freqs[last_idx]
         
         # Effective SR is 2 * Cutoff
         effective_sr = int(cutoff_freq * 2)