fixed a gain

sentence_api/audio_metrics.py

@@ -7,6 +7,10 @@ import numpy as np
 
 SAMPLE_RATE = 16_000
 FRAME_SAMPLES = 480
+# A silence run at least this long (in 30 ms frames) separates two sentences.
+# speech_frame_mask bridges gaps of up to 5 frames, so use 6 frames (~180 ms).
+MIN_PAUSE_FRAMES = 6
+ONSET_MARGIN_MS = 50
 
 
 class AudioAnalysisError(RuntimeError):
@@ -108,6 +112,44 @@ def speech_frame_mask(samples: np.ndarray, sample_rate: int = SAMPLE_RATE) -> np
     return speech
 
 
+def refine_sentence_end_ms(
+    samples: np.ndarray,
+    sample_rate: int = SAMPLE_RATE,
+    *,
+    raw_end_ms: int,
+    padded_end_ms: int,
+) -> int:
+    """Pull a padded sentence end back to just before the next sentence's speech.
+
+    Whisper's timestamps are not always aligned with the real audio: the next
+    segment's start can be later than the actual speech onset, so a fixed
+    end-padding may occasionally run into the next sentence's beginning. This
+    finds the first silence run of at least MIN_PAUSE_FRAMES inside the padded
+    region and stops the sentence just before the speech that follows it.
+    """
+    if padded_end_ms <= raw_end_ms:
+        return padded_end_ms
+    frame_ms = 1000 * max(1, int(round(sample_rate * 0.03))) / sample_rate
+    start_sample = max(0, int(raw_end_ms / 1000 * sample_rate))
+    end_sample = min(samples.size, int(padded_end_ms / 1000 * sample_rate))
+    if end_sample <= start_sample:
+        return padded_end_ms
+    try:
+        speech = speech_frame_mask(samples[start_sample:end_sample], sample_rate)
+    except AudioAnalysisError:
+        return padded_end_ms
+    silence_frames = 0
+    for index, is_speech in enumerate(speech):
+        if not is_speech:
+            silence_frames += 1
+            continue
+        if silence_frames >= MIN_PAUSE_FRAMES:
+            onset_ms = int(raw_end_ms + index * frame_ms)
+            return max(raw_end_ms, min(padded_end_ms, onset_ms - ONSET_MARGIN_MS))
+        silence_frames = 0
+    return padded_end_ms
+
+
 def _bridge_false_runs(values: np.ndarray, max_frames: int) -> None:
     start = None
     for index, value in enumerate(values):