macOS audio gaps under queue backpressure: sanitized v2.0.0 fork experiments #1117

Summary

A local macOS patch based on v2.0.0 still loses already-delivered microphone and system-audio spans under audio-queue backpressure. The level meter moves and the resulting MP4 contains audio, but decoded output has gaps during known test speech. This report is experimental evidence from a modified build, not a claim that current upstream main reproduces it.

Environment: macOS 26.6.2, Apple Silicon arm64, built-in microphone, full-display capture at 2880×1800, microphone + system audio + webcam. Local build: 2.0.1-hans.5; capture was stopped manually.

Measured result

• Audio duration: 186.219979 s, 48 kHz stereo. Video duration: 186.116667 s.
• Nine middle intervals contain exactly zero on both decoded channels for at least 20 ms.
• Longest middle zero interval: 140.397313–152.626667 s (12.229354 s).
• Known speech windows intersect gaps of 0.597958 s, 0.997354 s, and 2.914708 s.
• Native discarded-span diagnostics: microphone 14.669250 s, system 12.409417 s. These counters are not themselves proof of audible gaps; the independent PCM analysis above is.
• Maximum audioQueue wait: microphone 8.952128 s, system 8.951205 s.
• System callback lag stayed near 21 ms; microphone callback lag additionally reached 2.909080 s.

Experiments and limits

1. Independent AVCaptureSession microphone and explicit buffer conversion made microphone samples available, but did not establish reliable capture under load.
2. A one-second missing-data watermark retained a 350 ms synthetic late callback. Hardware runs of 832.44 s without webcam and 126.08 s with webcam had no middle zero interval >=20 ms. This did not generalize to the later 186 s run above.
3. Pre-anchor buffering removed the initial silent interval in the final test, but the middle gaps remained.

A separate apparent automatic-stop incident was caused by closing the captured Chrome window (SCStream -3815, source gone). Another short test hit disk full (AVFoundation -11807, POSIX 28). A 1.3 s take was tester-induced manual stop. None of these establishes a two-second product auto-stop regression.

Source-review hypothesis, not a verified fix

In the local AudioTrackMixer, the FIFO callback carries callbackTimelineTime, but drain advancement also reads execution-time clock(). With a long worker queue, an early queued callback can advance the missing-data watermark beyond later callbacks already waiting in the same FIFO. Those packets then arrive behind emittedEndFrame and are discarded. Timer ticks and timeline-start handling need the same event-clock ordering audit.

A useful regression would queue ~9 seconds of continuous, nonzero source packets while advancing the processing clock, replay the FIFO, and require zero discarded frames/no silent replacement. A producer-time timer snapshot should not overtake queued packet work. This fix has not been implemented or validated. Microphone producer stalls also remain a separate concern.

Expected behavior / reproduction aid

Already-delivered FIFO buffers should survive downstream encoding/queue stalls without silently being replaced by missing-data silence. Capture both sources plus webcam under concurrent load; export native timing diagnostics; decode the original MP4 and check actual speech windows, rather than relying on the level meter or dropped-span counters alone.

No personal recordings, full desktop diagnostics, API keys, or private paths are attached. Evidence: sanitized metrics, experiment report, and experimental mixer source. Related historical issue #709 had a different confirmed Int24 decoder cause; this report should not be treated as reopening that diagnosis.

Separate transcription experiment

The same fork contains an optional GMI integration whose historical attempts timed out or returned HTTP 500/429, including an explicit endpoint-overload response. Sanitized GMI experiment record. These cloud failures are distinct from capture loss and are not attributed to upstream Whisper. Local file exchange and other transcription routes are submitted separately.