fix(speech): close temp file before transcribing to fix Windows EACCES in faster-whisper backend (#638)

faster-whisper's transcribe() was handed the path of a still-open NamedTemporaryFile; on Windows the open handle is exclusive, so PyAV's reopen failed with EACCES and local STT was broken. Switch to delete=False, close before transcribing, and unlink in a finally. The write is wrapped in the file's context manager so the handle closes even if write() raises, and unlink failures are logged at debug.

src/openjarvis/speech/faster_whisper.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import logging
+import os
 import tempfile
 from typing import List, Optional
 
@@ -105,11 +106,15 @@ class FasterWhisperBackend(SpeechBackend):
         try:
             model = self._ensure_model()
 
-            # Write audio to a temp file (faster-whisper needs a file path)
+            # Write audio to a temp file (faster-whisper needs a file path).
+            # delete=False + manual unlink: on Windows an open
+            # NamedTemporaryFile holds an exclusive handle, so PyAV's reopen
+            # of tmp.name inside model.transcribe() fails with EACCES.
             suffix = f".{format}" if not format.startswith(".") else format
-            with tempfile.NamedTemporaryFile(suffix=suffix, delete=True) as tmp:
-                tmp.write(audio)
-                tmp.flush()
+            tmp = tempfile.NamedTemporaryFile(suffix=suffix, delete=False)
+            try:
+                with tmp:
+                    tmp.write(audio)
 
                 kwargs = {}
                 if language:
@@ -117,6 +122,15 @@ class FasterWhisperBackend(SpeechBackend):
 
                 segments_iter, info = model.transcribe(tmp.name, **kwargs)
                 segments_list = list(segments_iter)
+            finally:
+                try:
+                    os.unlink(tmp.name)
+                except OSError as unlink_exc:
+                    logger.debug(
+                        "Could not remove temp audio file %s: %s",
+                        tmp.name,
+                        unlink_exc,
+                    )
         except Exception as exc:
             self._last_error = str(exc)
             raise

tests/speech/test_faster_whisper.py

@@ -53,6 +53,69 @@ def test_faster_whisper_transcribe():
         assert result.duration_seconds == 1.5
 
 
+def test_faster_whisper_transcribe_temp_file_reopenable_and_removed():
+    """The temp file must be closed before the model reads it, and gone after.
+
+    On Windows, an open NamedTemporaryFile holds an exclusive handle, so
+    PyAV's reopen of the path inside model.transcribe() fails with EACCES
+    unless the file is closed first. Opening the path inside the mocked
+    transcribe reproduces that failure mode on Windows.
+    """
+    import os
+
+    mock_info = MagicMock()
+    mock_info.language = "en"
+    mock_info.language_probability = 0.95
+    mock_info.duration = 1.5
+
+    seen = {}
+
+    def fake_transcribe(path, **kwargs):
+        seen["path"] = path
+        with open(path, "rb") as fh:
+            seen["content"] = fh.read()
+        return iter(()), mock_info
+
+    mock_model = MagicMock()
+    mock_model.transcribe.side_effect = fake_transcribe
+
+    with patch(
+        "openjarvis.speech.faster_whisper.WhisperModel",
+        return_value=mock_model,
+    ):
+        backend = FasterWhisperBackend(model_size="base", device="cpu")
+        backend.transcribe(b"fake audio bytes")
+
+    assert seen["content"] == b"fake audio bytes"
+    assert not os.path.exists(seen["path"])
+
+
+def test_faster_whisper_transcribe_removes_temp_file_on_error():
+    """The temp file is cleaned up even when transcription fails."""
+    import os
+
+    seen = {}
+
+    def fake_transcribe(path, **kwargs):
+        seen["path"] = path
+        raise RuntimeError("decode failed")
+
+    mock_model = MagicMock()
+    mock_model.transcribe.side_effect = fake_transcribe
+
+    with patch(
+        "openjarvis.speech.faster_whisper.WhisperModel",
+        return_value=mock_model,
+    ):
+        backend = FasterWhisperBackend(model_size="base", device="cpu")
+        with pytest.raises(RuntimeError, match="decode failed"):
+            backend.transcribe(b"fake audio bytes")
+
+    assert "path" in seen
+    assert not os.path.exists(seen["path"])
+    assert "decode failed" in (backend.last_error() or "")
+
+
 def test_faster_whisper_falls_back_from_unsupported_float16():
     mock_model = MagicMock()