Update app.py

app.py

@@ -42,16 +42,60 @@ diarize_pipeline = DiarizationPipeline.from_pretrained(
 )
 print("Speaker-diarization model ready!")
 
+MIN_NOISE_BLOCK_SEC = 1.0  # حداقل طول یه بلوک بی‌گوینده تا «نویز» حساب بشه
 
-def realign_word_timestamps(waveform_16k, words, whisper_start, whisper_end):
+
+def get_speaker_at(diarization, time_sec):
+    for turn, _, speaker in diarization.itertracks(yield_label=True):
+        if turn.start <= time_sec <= turn.end:
+            return speaker
+    return None
+
+
+def find_leading_noise_cutoff(diarization, whisper_start, search_back_sec=4.0):
+    """
+    برمی‌گرده به عقب از شروعِ حدسیِ Whisper، دنبال یه بلوک نویز/بی‌گوینده‌ی
+    طولانی (حداقل MIN_NOISE_BLOCK_SEC) که دقیقاً قبل از این جمله باشه.
+    اگه پیدا کرد، زمان پایانِ اون بلوک نویز رو برمی‌گردونه (یعنی «از اینجا به بعد
+    اعتماد کن»)؛ اگه پیدا نکرد، None برمی‌گردونه (یعنی چیزی رو نبُر).
+    """
+    window_start = max(0, whisper_start - search_back_sec)
+    step = 0.05
+    t = whisper_start
+    noise_end = None
+    noise_start = None
+
+    while t > window_start:
+        speaker = get_speaker_at(diarization, t)
+        if speaker is None:
+            if noise_end is None:
+                noise_end = t
+            noise_start = t
+        else:
+            break
+        t -= step
+
+    if noise_start is not None and noise_end is not None:
+        block_len = noise_end - noise_start
+        if block_len >= MIN_NOISE_BLOCK_SEC:
+            return noise_end
+    return None
+
+
+def realign_word_timestamps(waveform_16k, words, whisper_start, whisper_end, cutoff_start=None):
     """
     برای یه جمله، به‌جای زمان‌بندی حدسی خودِ Whisper، با یه مدل تخصصی جدا
     (wav2vec2 از طریق torchaudio) دقیقاً چک می‌کنه هر کلمه از کجای صدا گفته شده.
+    اگه cutoff_start داده بشه (یعنی یه بلوک نویز قبل از جمله شناسایی شده)،
+    جستجو از همون نقطه شروع میشه، نه از قبل‌ترش.
     اگه به هر دلیلی هم‌ترازی جواب نداد، همون زمان‌بندی قبلی Whisper دست‌نخورده می‌مونه.
     """
     try:
-        margin = 1
-        start_sample = max(0, int((whisper_start - margin) * 16000))
+        margin = 0.5
+        raw_start = whisper_start - margin
+        effective_start = max(raw_start, cutoff_start) if cutoff_start is not None else raw_start
+
+        start_sample = max(0, int(effective_start * 16000))
         end_sample = int((whisper_end + margin) * 16000)
         clip = waveform_16k[:, start_sample:end_sample]
 
@@ -88,8 +132,8 @@ def realign_word_timestamps(waveform_16k, words, whisper_start, whisper_end):
             if len(matched) < n_chars:
                 new_words.append(w)
                 continue
-            w_start = matched[0].start * ratio + (whisper_start - margin)
-            w_end = matched[-1].end * ratio + (whisper_start - margin)
+            w_start = matched[0].start * ratio + effective_start
+            w_end = matched[-1].end * ratio + effective_start
             new_words.append({"word": w["word"], "start": round(max(0, w_start), 3), "end": round(max(0, w_end), 3)})
             span_idx += n_chars
 
@@ -99,17 +143,6 @@ def realign_word_timestamps(waveform_16k, words, whisper_start, whisper_end):
         return words
 
 
-def get_speaker_at(diarization, time_sec):
-    """
-    برای یه لحظه‌ی مشخص از صدا، می‌گه اون لحظه مال کدوم گوینده‌ست (SPEAKER_00, SPEAKER_01, ...).
-    اگه هیچ گوینده‌ای پیدا نشد، None برمی‌گردونه.
-    """
-    for turn, _, speaker in diarization.itertracks(yield_label=True):
-        if turn.start <= time_sec <= turn.end:
-            return speaker
-    return None
-
-
 @app.get("/health")
 async def health():
     return {"status": "ok", "message": "SubFlow backend is running"}
@@ -165,7 +198,7 @@ async def transcribe(websocket: WebSocket):
         print("Step 5: transcription done")
 
         await websocket.send_json({"status": "aligning"})
-        print("Step 5.5: loading audio for word-alignment")
+        print("Step 5.5: loading audio + running speaker diarization")
 
         align_waveform, sr = torchaudio.load(audio_to_transcribe)
         if sr != 16000:
@@ -173,8 +206,8 @@ async def transcribe(websocket: WebSocket):
         if align_waveform.shape[0] > 1:
             align_waveform = align_waveform.mean(dim=0, keepdim=True)
 
-        print("Step 5.6: running speaker diarization")
         diarization = diarize_pipeline(audio_to_transcribe)
+        print("Step 5.6: diarization done, now aligning words per segment")
 
         segments = []
         for seg in segments_raw:
@@ -183,8 +216,12 @@ async def transcribe(websocket: WebSocket):
                 continue
             word_dicts = [{"word": w.word, "start": w.start, "end": w.end} for w in words]
 
+            cutoff = find_leading_noise_cutoff(diarization, word_dicts[0]["start"])
+            if cutoff is not None:
+                print(f"  [noise-trim] '{seg.text.strip()[:40]}...' cutoff moved from {word_dicts[0]['start']:.2f}s to {cutoff:.2f}s")
+
             realigned = realign_word_timestamps(
-                align_waveform, word_dicts, word_dicts[0]["start"], word_dicts[-1]["end"]
+                align_waveform, word_dicts, word_dicts[0]["start"], word_dicts[-1]["end"], cutoff_start=cutoff
             )
 
             mid_time = (realigned[0]["start"] + realigned[-1]["end"]) / 2