MIC3 Was an Electrical Speaker Reference, Not Just Another Microphone

This was not primarily a firmware problem or a hardware problem. It was an interface problem: AEC debugging initially suffered because the channel expected to carry a far-end reference appeared almost dead.

The board context for this series is LOUP's Minewing V1.6 ESP32-S3 hardware: ESP32-S3 N16R8-class memory configuration, ES8311 playback, ES7210 capture, AXP2101 power management, e-paper display, physical controls and factory/recovery interfaces. I use those identifiers only where they help explain the engineering boundary; the larger lesson is about how firmware, schematic and physical assembly have to agree.

The evidence for this case was specific: Minewing schematic evidence connected SPK_OUTP/SPK_OUTN into the ES7210 MIC3 reference network; after the packing fix, pre3 became the live reference and downstream AEC reference amplitude rose from near-zero to meaningful values. I treat that as evidence from this board/revision and investigation, not as a universal statement about every ESP32-S3 design.

The result I retained was equally narrow: The hardware reference was proven alive and the fault moved to software interpretation/packing rather than the MIC3 circuitry.

How I framed the problem

I wrote down the physical signal or state first, then asked which document and which firmware path claimed to own it. The problem was AEC debugging initially suffered because the channel expected to carry a far-end reference appeared almost dead. Runtime/document evidence showed Minewing schematic evidence connected SPK_OUTP/SPK_OUTN into the ES7210 MIC3 reference network; after the packing fix, pre3 became the live reference and downstream AEC reference amplitude rose from near-zero to meaningful values. Because An electrical reference is a deliberate feedback path from speaker output into the ADC; its semantics differ from an acoustic microphone even when both arrive through the same codec., I refused to accept Treating all ES7210 inputs as interchangeable microphone channels. as proof. The useful outcome was The hardware reference was proven alive and the fault moved to software interpretation/packing rather than the MIC3 circuitry.

Audio hardware sits across several domains at once. Digital I2S can be correct while analog gain is wrong. The ADC can answer over I2C while its physical reference channel is misinterpreted. A speaker can reproduce every sample and still sound thin because the amplifier, driver and enclosure do not support the expected acoustic response. The useful unit of debugging is therefore the complete signal path, not the codec part number.

The practical rule that came out of the case was: AEC channel names must be grounded in the schematic and runtime signal, not only the codec register map. I prefer a rule like that over a one-off patch because it changes the next bring-up decision before another board is modified.

Evidence matrix

I keep this table because board bring-up narratives become unreliable very quickly. A working prototype encourages retrospective certainty: once the device boots, it is easy to rewrite every earlier guess as if it had been obvious. The matrix preserves the difference between what the board actually demonstrated and what I merely considered plausible.

The boundary I wanted to prove

PCM -> ES8311 -> amplifier -> speaker -> enclosure/room
                              ^             |
                              |             v
                    electrical ref       microphones
                              \----------> ES7210 -> DMA/DSP

For this layer I wanted at least these checks before changing the design:

• codec presence and clocks
• physical channel/reference routing
• analog enable/power path
• known electrical stimulus or capture
• acoustic result after digital path is proven

The important part is ordering. I do not start with the last item just because firmware is the easiest thing for me to edit. If the rail is absent, a driver rewrite is irrelevant. If the exact part differs from the assumed part, a timing tweak may only hide the mismatch. If the physical channel is wrong, the DSP can be perfectly stable while processing the wrong signal.

For this case, An electrical reference is a deliberate feedback path from speaker output into the ADC; its semantics differ from an acoustic microphone even when both arrive through the same codec. That mechanism defines which measurement belongs before the patch and which measurement should change afterward.

Investigation method

I try to separate presence, configuration and function. Presence means the part answers or the net exists. Configuration means register state, pin mux and power state are what I intended. Function means a real signal traverses the subsystem. A component can pass the first two and fail the third. That hierarchy prevented several I2C and display checks from being promoted to false product acceptance.

The shortcut I deliberately avoided here was Treating all ES7210 inputs as interchangeable microphone channels. That shortcut is attractive because it converts a cross-disciplinary problem into something one person can edit immediately. It is also how firmware becomes a compensation layer for an electrical problem that nobody has actually measured.

What firmware can prove—and what it cannot

Firmware can prove that it configured a peripheral, observed an I2C ACK, selected a pin mux, received DMA data, read a status bit or saw a button transition. Those are useful facts. They are not substitutes for physical measurements when the disputed state exists outside the MCU.

For MIC3 Was an Electrical Speaker Reference, Not Just Another Microphone, the relevant distinction is that An electrical reference is a deliberate feedback path from speaker output into the ADC; its semantics differ from an acoustic microphone even when both arrive through the same codec. A log can expose the software side of that relationship, but the electrical/mechanical side still needs the appropriate observation point.

This matters most when a diagnostic success is weaker than the product claim. An I2C scan cannot prove microphone quality. A BUSY transition cannot prove display alignment or long-term FPC reliability. A GPIO write cannot prove the amplifier enable pin actually changed if an expander or transistor sits between them. A factory programming command cannot prove traceability unless the result is tied to the unit identity.

I therefore write two columns in bring-up notes: “software evidence” and “physical evidence.” A fix is stronger when both point at the same mechanism.

Acceptance test I would keep

I prefer a fixture-friendly acceptance test over a one-off engineering ritual. Anything that needs a probe point should have an accessible point or a factory diagnostic. Anything that relies on component identity should appear in the BOM/revision record. Anything that can regress in firmware should have a boot or factory log marker that is cheap enough to retain.

For this case the acceptance target is derived from the retained result: The hardware reference was proven alive and the fault moved to software interpretation/packing rather than the MIC3 circuitry. The test should prove that result directly rather than infer it from a neighboring signal.

What this changes before PCB release

The lesson is not only about debugging the current EVT. It changes the release package. AEC channel names must be grounded in the schematic and runtime signal, not only the codec register map.

For a board revision, I want the schematic revision, BOM identity, power-tree assumptions, pin map, factory/recovery interfaces and firmware hardware contract to move together. If one changes, the others should either change or explicitly state why they do not. This is especially important around programmable parts such as the PMIC and around signals whose semantics are created jointly by analog routing and software mapping.

I also want unresolved questions to remain visible. “Works on EVT” should not silently close an electrical-margin question, a tactile-control mismatch or an acoustic uncertainty. A production decision needs evidence appropriate to the risk. That may be a reset-time voltage analysis, a fixture measurement, a component supplier confirmation, an enclosed-device acoustic test or a repeated assembly trial.

The factory benefits from the same clarity. A deterministic test path reduces rework and makes a failed unit diagnosable instead of merely rejected.

What I would change on the next board

I would make more of these boundaries explicit before layout. Each programmable power rail would have a table with voltage, owner, default state and test point. Boot-sensitive GPIOs would be reviewed as a separate checklist before peripheral placement is frozen. Codec and ADC channel mapping would be documented from schematic net to DMA representation. Recovery pads would be designed with the fixture, not added after the board already existed.

For human-interface parts, I would require an exact supplier variant and physical sample whenever the requirement contains a tactile or acoustic adjective. “Detented,” “loud,” “clear,” “thin,” “clicky” and “stable” cannot be accepted from a symbol or generic family datasheet alone.

For MIC3 Was an Electrical Speaker Reference, Not Just Another Microphone, I would carry forward the mechanism directly: An electrical reference is a deliberate feedback path from speaker output into the ADC; its semantics differ from an acoustic microphone even when both arrive through the same codec. That turns this incident into a design-review question instead of another bring-up surprise.

The rule I kept

AEC channel names must be grounded in the schematic and runtime signal, not only the codec register map.

The retained result from this case was: The hardware reference was proven alive and the fault moved to software interpretation/packing rather than the MIC3 circuitry.

That is how I now approach PCB bring-up. I do not ask firmware to compensate for an unmeasured electrical problem, and I do not ask hardware engineers to redesign a circuit because a software label looked wrong. I locate the boundary, choose an observation point that can actually see it, reconcile the authoritative artifacts, and only then change the layer that owns the failure.

The process feels slower than immediately editing code. Across multiple board revisions it is much faster, because every confirmed boundary becomes reusable evidence for the next failure.