fix mic input: always downmix to mono 16 kHz (stereo recordings crashed NeMo)

app.py

@@ -20,30 +20,48 @@ def get_model():
     return MODEL
 
 
-def _ensure_wav(path):
-    """Mic recordings arrive as webm/mp4; libsndfile can't read those — convert."""
+def _prepare(path):
+    """Normalize any input to mono 16 kHz PCM and return (path, duration_seconds).
+
+    Browser microphone recordings arrive as stereo (and sometimes as webm/mp4 that
+    libsndfile cannot open at all). The model takes a mono signal only: feeding it a
+    two-channel file raises "Output shape expected = (batch, time)" inside NeMo.
+    """
+    import subprocess
+    import tempfile
+
     import soundfile as sf
+
     try:
-        sf.info(path)
-        return path
+        info = sf.info(path)
+        needs_convert = info.channels != 1 or info.samplerate != 16000
+        duration = info.duration
     except Exception:
-        import subprocess
-        import tempfile
-        out = tempfile.NamedTemporaryFile(suffix=".wav", delete=False).name
-        subprocess.run(
-            ["ffmpeg", "-y", "-loglevel", "error", "-i", path, "-ar", "16000", "-ac", "1", out],
-            check=True,
+        needs_convert = True
+        duration = None
+
+    if not needs_convert:
+        return path, duration
+
+    out = tempfile.NamedTemporaryFile(suffix=".wav", delete=False).name
+    proc = subprocess.run(
+        ["ffmpeg", "-y", "-loglevel", "error", "-i", path,
+         "-ar", "16000", "-ac", "1", "-c:a", "pcm_s16le", out],
+        capture_output=True, text=True,
+    )
+    if proc.returncode != 0:
+        raise gr.Error(
+            "Could not read that audio (ffmpeg: %s). Try a WAV or MP3 file."
+            % (proc.stderr.strip().splitlines() or ["unknown error"])[-1]
         )
-        return out
+    return out, sf.info(out).duration
 
 
 def transcribe(audio_path):
     if not audio_path:
         return "(no audio)"
-    import soundfile as sf
-    audio_path = _ensure_wav(audio_path)
-    info = sf.info(audio_path)
-    if info.duration > 60:
+    audio_path, duration = _prepare(audio_path)
+    if duration is not None and duration > 60:
         return "Please keep clips under 60 seconds for this CPU demo."
     model = get_model()
     out = model.transcribe([audio_path], batch_size=1, verbose=False)[0]