continue fixing

sentence_api/audio_metrics.py

@@ -1,6 +1,6 @@
 from dataclasses import dataclass
 from pathlib import Path
-from typing import Tuple
+from typing import Optional, Tuple
 
 import numpy as np
 
@@ -11,6 +11,8 @@ FRAME_SAMPLES = 480
 # speech_frame_mask bridges gaps of up to 5 frames, so use 6 frames (~180 ms).
 MIN_PAUSE_FRAMES = 6
 ONSET_MARGIN_MS = 50
+START_LOOKBACK_MS = 300
+START_LOOKAHEAD_MS = 400
 
 
 class AudioAnalysisError(RuntimeError):
@@ -150,6 +152,57 @@ def refine_sentence_end_ms(
     return padded_end_ms
 
 
+def refine_sentence_start_ms(
+    samples: np.ndarray,
+    sample_rate: int = SAMPLE_RATE,
+    *,
+    raw_start_ms: int,
+    previous_end_ms: int,
+) -> int:
+    """Pull a sentence start forward to the actual speech onset.
+
+    Whisper's segment start can be earlier than the real speech onset (the tail
+    of the previous sentence is still audible at the beginning of the next
+    sentence). When the raw start lands inside continuous speech, this looks for
+    the last pause of at least MIN_PAUSE_FRAMES before the next speech run and
+    moves the start to that onset, so adjacent sentences share no audio.
+    """
+    if raw_start_ms <= previous_end_ms:
+        return max(previous_end_ms, raw_start_ms)
+    frame_ms = 1000 * max(1, int(round(sample_rate * 0.03))) / sample_rate
+    total_ms = samples.size / sample_rate * 1000
+    window_start_ms = max(0, raw_start_ms - START_LOOKBACK_MS)
+    window_end_ms = min(int(total_ms), raw_start_ms + START_LOOKAHEAD_MS)
+    start_sample = int(window_start_ms / 1000 * sample_rate)
+    end_sample = min(samples.size, int(window_end_ms / 1000 * sample_rate))
+    if end_sample - start_sample < max(1, int(round(sample_rate * 0.03))):
+        return raw_start_ms
+    try:
+        speech = speech_frame_mask(samples[start_sample:end_sample], sample_rate)
+    except AudioAnalysisError:
+        return raw_start_ms
+    onset_frame = int(round((raw_start_ms - window_start_ms) / frame_ms))
+    if (
+        0 <= onset_frame < len(speech)
+        and speech[onset_frame]
+        and (onset_frame == 0 or not speech[onset_frame - 1])
+    ):
+        return raw_start_ms  # already at a speech onset
+    boundary_ms: Optional[float] = None
+    silence_frames = 0
+    for index, is_speech in enumerate(speech):
+        if not is_speech:
+            silence_frames += 1
+            continue
+        if silence_frames >= MIN_PAUSE_FRAMES:
+            boundary_ms = window_start_ms + index * frame_ms
+        silence_frames = 0
+    if boundary_ms is None:
+        return raw_start_ms
+    boundary_ms = min(boundary_ms, raw_start_ms + START_LOOKAHEAD_MS)
+    return max(previous_end_ms, int(boundary_ms))
+
+
 def _bridge_false_runs(values: np.ndarray, max_frames: int) -> None:
     start = None
     for index, value in enumerate(values):