fff

sentence_api/audio_metrics.py

@@ -57,12 +57,38 @@ def analyze_audio(path: Path) -> AudioMetrics:
 
 
 def analyze_samples(samples: np.ndarray, sample_rate: int = SAMPLE_RATE) -> AudioMetrics:
+    if samples.ndim != 1:
+        samples = samples.reshape(-1)
+    speech = speech_frame_mask(samples, sample_rate)
+    speech_indexes = np.flatnonzero(speech)
+    if speech_indexes.size == 0:
+        raise AudioAnalysisError("No usable speech was detected in the recording.")
+
+    recording_duration_ms = max(1, int(round(samples.size / sample_rate * 1000)))
+    window = max(1, int(round(sample_rate * 0.03)))
+    frame_ms = window / sample_rate * 1000
+    speech_duration_ms = max(1, int(round(speech.sum() * frame_ms)))
+    first = int(speech_indexes[0])
+    last = int(speech_indexes[-1])
+    internal_frames = max(1, last - first + 1)
+    internal_silence_frames = int((~speech[first : last + 1]).sum())
+    internal_silence_ms = int(round(internal_silence_frames * frame_ms))
+    internal_pause_ratio = internal_silence_frames / internal_frames
+    return AudioMetrics(
+        recording_duration_ms=recording_duration_ms,
+        speech_duration_ms=speech_duration_ms,
+        internal_silence_ms=internal_silence_ms,
+        internal_pause_ratio=round(internal_pause_ratio, 4),
+    )
+
+
+def speech_frame_mask(samples: np.ndarray, sample_rate: int = SAMPLE_RATE) -> np.ndarray:
+    """Per-frame speech mask (30 ms frames) using the same RMS threshold as analyze_samples."""
     if samples.ndim != 1:
         samples = samples.reshape(-1)
     if samples.size == 0 or sample_rate <= 0:
         raise AudioAnalysisError("The uploaded audio is empty.")
 
-    recording_duration_ms = max(1, int(round(samples.size / sample_rate * 1000)))
     window = max(1, int(round(sample_rate * 0.03)))
     complete_frames = int(np.ceil(samples.size / window))
     padded = np.pad(samples, (0, complete_frames * window - samples.size))
@@ -79,24 +105,7 @@ def analyze_samples(samples: np.ndarray, sample_rate: int = SAMPLE_RATE) -> Audi
     # Treat very short gaps inside a word as speech, then reject short clicks.
     _bridge_false_runs(speech, max_frames=5)
     _remove_true_runs(speech, max_frames=2)
-    speech_indexes = np.flatnonzero(speech)
-    if speech_indexes.size == 0:
-        raise AudioAnalysisError("No usable speech was detected in the recording.")
-
-    frame_ms = window / sample_rate * 1000
-    speech_duration_ms = max(1, int(round(speech.sum() * frame_ms)))
-    first = int(speech_indexes[0])
-    last = int(speech_indexes[-1])
-    internal_frames = max(1, last - first + 1)
-    internal_silence_frames = int((~speech[first : last + 1]).sum())
-    internal_silence_ms = int(round(internal_silence_frames * frame_ms))
-    internal_pause_ratio = internal_silence_frames / internal_frames
-    return AudioMetrics(
-        recording_duration_ms=recording_duration_ms,
-        speech_duration_ms=speech_duration_ms,
-        internal_silence_ms=internal_silence_ms,
-        internal_pause_ratio=round(internal_pause_ratio, 4),
-    )
+    return speech
 
 
 def _bridge_false_runs(values: np.ndarray, max_frames: int) -> None: