Fix audio playback + persist logs across refreshes

Audio clips: Convert from float32 WAV to PCM int16 WAV for browser
compatibility. Float32 WAV causes playback to cut short or fail
in many browsers.

Logs: Save session history to JSON file on disk (DATA_DIR/logs/)
so logs survive page refreshes and container restarts. Load from
disk on init, clear button also wipes the file.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Dockerfile

@@ -20,7 +20,7 @@ RUN useradd -m -u 1000 user
 
 # Create model & data dirs under /home/user (NOT /app, which HF mounts over)
 RUN mkdir -p /home/user/models/vad /home/user/models/spkrec \
-    /home/user/data/db /home/user/data/clips \
+    /home/user/data/db /home/user/data/clips /home/user/data/logs \
     && chown -R user:user /home/user
 
 # Pre-download models AS the user so files are owned by user

app.py

@@ -86,12 +86,44 @@ def get_analyzer():
     return AudioAnalyzer()
 
 
+def _get_logs_file():
+    """Path to the persistent logs JSON file."""
+    data_dir = os.environ.get("DATA_DIR", "data")
+    logs_dir = os.path.join(data_dir, "logs")
+    os.makedirs(logs_dir, exist_ok=True)
+    return os.path.join(logs_dir, "analysis_logs.json")
+
+
+def _load_persisted_logs():
+    """Load log sessions from disk."""
+    import json as _json
+    path = _get_logs_file()
+    if os.path.exists(path):
+        try:
+            with open(path, 'r') as f:
+                return _json.load(f)
+        except Exception:
+            pass
+    return []
+
+
+def _save_persisted_logs(sessions):
+    """Save log sessions to disk."""
+    import json as _json
+    path = _get_logs_file()
+    try:
+        with open(path, 'w') as f:
+            _json.dump(sessions, f)
+    except Exception:
+        pass
+
+
 def _init_log_state():
-    """Initialize session state for analysis logs."""
+    """Initialize session state for analysis logs, loading from disk."""
     if 'analysis_logs' not in st.session_state:
         st.session_state['analysis_logs'] = {
             'current': [],
-            'sessions': []
+            'sessions': _load_persisted_logs()
         }
 
 
@@ -131,6 +163,7 @@ def _finalize_log_session(result=None):
     }
     logs['sessions'].insert(0, session)
     logs['current'] = []
+    _save_persisted_logs(logs['sessions'])
 
 
 def render_waveform(audio_data, sample_rate):
@@ -1123,6 +1156,7 @@ def render_logs_tab():
 
         if st.button("🗑️ Clear History", key="clear_log_history"):
             st.session_state['analysis_logs']['sessions'] = []
+            _save_persisted_logs([])
             st.rerun()
 
 

src/analyzer.py

@@ -646,11 +646,15 @@ class AudioAnalyzer:
         # Concatenate all clips
         clip = torch.cat(clips, dim=1)
 
+        # Convert to int16 PCM for browser compatibility
+        clip_np = clip.squeeze(0).numpy()
+        clip_int16 = np.clip(clip_np * 32767, -32768, 32767).astype(np.int16)
+
         # Save clip
+        import soundfile as sf
         clip_filename = f"{test_id}_{vp_id}_{total_duration:.1f}s.wav"
         clip_path = os.path.join(self.clips_dir, clip_filename)
-
-        torchaudio.save(clip_path, clip, sample_rate)
+        sf.write(clip_path, clip_int16, sample_rate, subtype='PCM_16')
 
         return clip_path