personal memory agent
Treat marginal VAD speech (<5s) with no transcription as silence
When VAD detects a small amount of speech but the STT backend returns
zero statements, log it as marginal detection instead of raising a
RuntimeError. This avoids noisy error alerts for segments that are
just ambient noise with brief blips.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
+14
-5
+14
-5
observe/transcribe/main.py
+14
-5
observe/transcribe/main.py
···
504
504
505
505
# Sanity check: if VAD detected speech but we got no statements, something is wrong
506
506
if vad_result.has_speech and not statements:
507
-
raise RuntimeError(
508
-
f"VAD detected {vad_result.speech_duration:.1f}s of speech "
509
-
f"(from {vad_result.duration:.1f}s total) but transcription produced "
510
-
f"0 statements. This indicates a transcription failure, not silence."
511
-
)
507
+
if vad_result.speech_duration < 5.0:
508
+
# Marginal speech detection — treat as silence rather than failure.
509
+
# VAD occasionally flags brief noise as speech; if the STT backend
510
+
# can't produce anything from it, that's expected, not an error.
511
+
logging.info(
512
+
f"VAD detected {vad_result.speech_duration:.1f}s of marginal speech "
513
+
f"but transcription produced 0 statements — treating as silence"
514
+
)
515
+
else:
516
+
raise RuntimeError(
517
+
f"VAD detected {vad_result.speech_duration:.1f}s of speech "
518
+
f"(from {vad_result.duration:.1f}s total) but transcription produced "
519
+
f"0 statements. This indicates a transcription failure, not silence."
520
+
)
512
521
513
522
# Load config for preserve_all setting
514
523
config = get_config()