comma four: microphone pop fixes and audio quality investigation

[sshane]

Comma four desk recordings exposed two microphone firmware defects: an uninitialized output buffer produces startup junk, and independent DFSDM/I2S clocks cause repeated 512-sample blocks about every 10.53 seconds. This initializes/clears the buffers and adds continuous clock matching with cubic interpolation and a startup fade.

This is a hackathon work-in-progress draft, including tools, listening reports and plots so we can document the investigation as we go. The muffled sound / weak upper-frequency music content is still unresolved, and the original driving/motion complaint has not been reproduced under normal on-road load.

Actual pop before/after

The fresh baseline repeats 512 samples at 10.521 s; the candidate does not. These are separate real acoustic recordings, not simulated distortion.

Quality investigation

The native digital transport passed a bit-exact 528,000-sample test through 22 kHz. The cubic resampler adds gradual high-frequency loss and residual distortion; it does not explain an abrupt cutoff by itself.

The built-in c4 speaker produced detectable tones through 22 kHz, but it has different acoustic/mechanical coupling than an external phone. The iPhone sweep recovers 16–17 kHz; 18–20 kHz remain close to the noise floor. The new music recording shows the same weak upper band. Raising file amplitude produces less than proportional recorded tone growth, so source/capture nonlinearity is another open question.

Validation and open work

• Fresh-reset 300-second instrumented run: 28,127 output blocks, zero re-centering, repeats or capture overruns; max mic IRQ 531 µs against a 10.67 ms block period.
• Final uninstrumented firmware through actual micd/loggerd with CPU load: 1,200 messages / 960,000 PCM samples matched the rlog byte-for-byte.
• Saved final ARM build, intended MISRA checks and actual-C UBSan clock-offset simulations pass. New phone music and level captures have zero overruns.
• Pending: normal on-road workload/motion validation, end-to-end latency, higher-quality interpolation, and separating external speaker response from mic/enclosure response. The companion openpilot PR covers native-rate capture and compressed logging; the firmware PR itself does not change micd or AAC.

Full investigation notebook · All plots · Embedded listening reports — download and open locally · Hackathon tools

Companion openpilot PR: commaai/openpilot#38967 — 48 kHz mono capture and one 48 kbps AAC encoder shared by qcamera and compressed rlog audio. Experimental EQ/denoise remains offline.