fix: #2010 Improve PCM duration calculation and handle VAD truncation (#2009)

[ruskaruma]

Summary

In #2010 we noticed realtime PCM16 audio was being reported in microseconds. The root cause was our fallback branch in calculate_audio_length_ms, which divided by 24 * 2 and then by 1000, effectively treating every byte as a millisecond. I normalized the format first, split g711 from PCM handling, and introduced explicit constants for the sample rate and sample width so PCM16 now returns the true millisecond duration. This also covers empty buffers and any future lowercase/uppercase format variants.

To keep coverage honest, I updated the realtime playback tracker tests to assert against the corrected math using pytest.approx, and refreshed the manual tracker test to reflect the new millisecond totals.

Test plan

• uv run pytest tests/realtime/test_playback_tracker_manual_unit.py tests/realtime/test_playback_tracker.py tests/realtime/test_openai_realtime.py
• uv run pytest

Fixes #2010.

[ruskaruma]

Greetings @seratch, I’ve opened this PR referencing issue #2010. Please review the patch and let me know if it correctly addresses the problem or if any adjustments are needed.

[ruskaruma]

Greetings, @seratch .
I’ve made these changes and fixed the issues:

Updated calculate_audio_length_ms to return correct PCM16 durations (fixes #2010)

Made the VAD speech-start event emit conversation.item.truncate so the server properly records interruptions (fixes #2009)

Here’s the test plan:

uv run pytest tests/realtime/test_playback_tracker_manual_unit.py tests/realtime/test_playback_tracker.py tests/realtime/test_openai_realtime.py

uv run pytest

Let me know if you’d like me to tweak or add anything.