transcript updates

cht/config.py

@@ -26,5 +26,7 @@ AUDIO_SAFETY_MARGIN = 2       # seconds safety margin (matches scene detector)
 WAVEFORM_BUCKET_MS = 50       # milliseconds per waveform peak bucket
 
 # Transcription
-WHISPER_MODEL = "small"        # "small" for speed, "medium" for accuracy
+WHISPER_MODEL = "medium"       # "small" for speed, "medium" for accuracy
 WHISPER_DEVICE = "cuda"        # "cuda" or "cpu"
+TRANSCRIBE_MIN_CHUNK_S = 5     # minimum seconds of audio before transcribing
+TRANSCRIBE_LINES_PER_GROUP = 3 # whisper segments grouped per transcript ID (1-5)

cht/stream/manager.py

@@ -39,6 +39,16 @@ def list_sessions():
     return sessions
 
 
+def delete_sessions(session_ids):
+    """Delete session directories by ID."""
+    import shutil
+    for sid in session_ids:
+        path = SESSIONS_DIR / sid
+        if path.exists() and path.is_dir():
+            shutil.rmtree(path)
+            log.info("Deleted session: %s", sid)
+
+
 class StreamManager:
     def __init__(self, session_id=None):
         if session_id is None:

cht/transcriber/engine.py

@@ -59,32 +59,75 @@ class TranscriberEngine:
             return []
         self._ensure_model()
         try:
-            kwargs = {"beam_size": 5, "vad_filter": True}
+            kwargs = {
+                "beam_size": 5,
+                "vad_filter": True,
+                "condition_on_previous_text": True,
+            }
             if self.language:
                 kwargs["language"] = self.language
+            # Feed last transcript text as context for better continuity
+            if self._segments:
+                kwargs["initial_prompt"] = self._segments[-1].text
             segments_iter, info = self._model.transcribe(str(wav_path), **kwargs)
         except Exception as e:
             log.error("Whisper transcription failed: %s", e)
             return []
 
+        # Group whisper segments: new T-ID every N lines or on silence gap (>1s)
+        from cht.config import TRANSCRIBE_LINES_PER_GROUP
+        lines_per_group = TRANSCRIBE_LINES_PER_GROUP
+        SILENCE_GAP_S = 1.0
+
+        raw_segs = []
+        for seg in segments_iter:
+            text = seg.text.strip()
+            if text:
+                raw_segs.append((time_offset + seg.start, time_offset + seg.end, text))
+
         new_segments = []
         with self._lock:
             if self._stopped:
                 return []
-            for seg in segments_iter:
-                text = seg.text.strip()
-                if not text:
-                    continue
+
+            group_start = None
+            group_end = None
+            group_lines = []
+            prev_end = None
+
+            def _flush():
+                nonlocal group_start, group_end, group_lines
+                if not group_lines:
+                    return
                 tid = f"T{self._next_id:04d}"
                 self._next_id += 1
                 entry = TranscriptSegment(
                     id=tid,
-                    start=time_offset + seg.start,
-                    end=time_offset + seg.end,
-                    text=text,
+                    start=group_start,
+                    end=group_end,
+                    text=" ".join(group_lines),
                 )
                 self._segments.append(entry)
                 new_segments.append(entry)
+                group_lines = []
+                group_start = None
+                group_end = None
+
+            for start, end, text in raw_segs:
+                # Silence gap → flush current group
+                if prev_end is not None and start - prev_end > SILENCE_GAP_S:
+                    _flush()
+
+                if group_start is None:
+                    group_start = start
+                group_end = end
+                group_lines.append(text)
+                prev_end = end
+
+                if len(group_lines) >= lines_per_group:
+                    _flush()
+
+            _flush()
 
         return new_segments