Voice Recorder / Journalist Recorder Hardware Design Guide
Core idea: A journalist voice recorder is engineered to capture speech instantly and reliably—clean mic/ADC audio, VAD/NR that doesn’t miss words, and storage that never corrupts files—even while charging and under real-world handling/ESD.
Success means measurable results: low noise floor, fast wake-to-record, stable long-take timing, and robust power-loss recovery so every press produces usable audio.
Scope, Use-Cases, and “What Good Looks Like”
Intent: Lock the engineering boundary and define measurable success criteria so every later chapter can be verified against the same yardstick.
Product Boundary
Primary Use-Cases (design must satisfy)
- Instant capture: pull out, press record, first syllable is preserved (no “missing first word”).
- Far/near speech: intelligible capture from 0.5–2 m, plus close-talk without clipping.
- Long standby + reliable wake: weeks of standby with predictable wake-to-record behavior.
- Field robustness: plug/unplug USB, ESD events, and low-battery conditions must not corrupt audio files.
- Record while charging: fast-charge and USB noise must not raise the audible noise floor beyond acceptable limits.
Five “Must-Hit” KPIs (each has a minimum evidence requirement)
Minimum evidence: (1) silent-room noise spectrum (charging vs not charging) and (2) a fixed speech sample SNR comparison at 0.5 m and 2 m.
User-visible failure: distant speech sounds “veiled,” consonants smear, NR makes “watery” artifacts.
Minimum evidence: GPIO marker + timestamped log or scope capture that brackets “button edge” to “first valid audio frame written.”
User-visible failure: first syllable clipped even though the user pressed record in time.
Minimum evidence: annotated speech clips → compare VAD segments vs ground truth (count missed starts/ends).
User-visible failure: quiet talk or trailing words disappear; recorder “cuts off sentences.”
Minimum evidence: >100 randomized power-cut/unplug tests with a pass metric: “playable,” “recoverable,” “lost,” “corrupted.”
User-visible failure: the interview exists but the file will not open—unacceptable.
Minimum evidence: charging vs battery-only noise spectra and spur thresholds (spurs near switching/SD burst frequencies).
User-visible failure: hiss rises when USB is plugged; periodic ticks during SD writes; “buzz” at fixed tones.
Evidence Chain (repeatable debug format)
Use the same 4-step structure in later chapters: Symptom → First 2 evidence points (probe/log) → Discriminator (what proves root cause) → First fix (fastest effective change).
Example (KPI-2 “missing first word”):
- Symptom: first syllable clipped after pressing record.
- First 2 evidence: (1) button-edge-to-audio-frame timestamp; (2) AFE output settling waveform during wake.
- Discriminator: if audio frames start late → wake/file-open bottleneck; if frames start on time but waveform ramps → AFE settle/HPF transient.
- First fix: add pre-roll buffer + shorten wake path; or add analog mute/soft-start window for AFE before writing “valid” frames.
Acoustic Front-End: Mic Choice, Biasing, and Analog Noise Budget
Intent: Make the recorder’s noise floor tangible—identify what sets intelligibility before DSP, and define how to prove the bottleneck with minimal measurements.
Mic Options (kept strictly recorder-scoped)
Focus criteria for journalist recorders: self-noise (far speech), bias/EMI sensitivity (USB/charging), handling/ESD robustness, and power budget.
- ECM: often strong subjective speech tone; bias network must be clean; can be sensitive to handling and layout noise if bias filtering is weak.
- Analog MEMS: consistent performance and easier mechanical integration; still needs careful bias/decoupling and RF/USB noise control.
- Digital MEMS: moves A/D closer to the mic; can improve analog noise susceptibility but introduces clock/data integrity concerns and power/EMI paths that must be contained (keep discussion at “integration implications,” not protocol deep-dive).
Biasing & Coupling: where pop/click and USB noise enter
- Mic bias node (thermal noise + ripple injection point)
- Preamp input node (ESD and touch/handling coupling point)
- AFE reference/analog rail (charger ripple + SD/CPU burst return path)
- Bias resistor is not “just a resistor”: its thermal noise can dominate the low-level floor if value is too high for the mic sensitivity and intended distance.
- RC filtering must be placed at the right physical point: the capacitor must be near the sensitive node; long routes turn the filter into an antenna.
- Pop/click root causes (recorder-typical): wake-up bias ramp, HPF settling transient, mux/AGC step changes, or ESD/USB attach events disturbing the bias/reference.
- ESD return path matters more than the TVS headline: clamp current should return to chassis/ground without crossing the preamp input/bias reference.
Noise Budget Worksheet (conceptual, measurable)
Two fast bottleneck checks (field-friendly):
- Gain-step test: increase analog gain by a known step. If noise rises proportionally, the floor is upstream (mic/bias/preamp). If noise barely changes, ADC/digital floor may dominate or NR is shaping the floor.
- Charging A/B test: record silence on battery, then on USB fast-charge. New narrowband spurs indicate coupling from charger/ground return rather than intrinsic mic/preamp noise.
Practical Targets (speech-first, not studio-first)
- Close-talk interview: prioritize headroom and stable limiter behavior—avoid harsh clipping and avoid aggressive AGC pumping.
- Tabletop / pocket pull-out: prioritize low noise floor and stable bias/reference—VAD should not gate out soft words; pre-roll must protect sentence starts.
- Handling & wind: analog HPF (or early DSP HPF) must remove sub-100 Hz rumble without thinning speech fundamentals.
Design Checklist (engineer-usable, recorder-scoped)
- Place mic bias RC filtering at the mic/AFE node (short loop, tight return).
- Keep mic input routing away from USB-C/charger switching nodes and SD clock lines.
- Define a controlled wake sequence: bias settle → AFE unmute → HPF/NR enable → write “valid frames.”
- Prevent gain-step clicks: ramp gain or apply a short crossfade window around changes.
- Choose bias resistor and preamp gain together to meet far-speech noise floor without sacrificing close-talk headroom.
- Use a dedicated analog LDO or filtered rail for the AFE if charging noise is a requirement (KPI-5).
- Provide an ESD path that returns clamp current without crossing AFE input/reference loops.
- Audit “quiet conditions”: SD write bursts, CPU wake bursts, and charger state changes—ensure they do not modulate bias/reference.
Preamp, PGA/AGC, Anti-Alias, ADC: Getting Clean Samples Without Clipping
Intent: Keep dynamic range for far/near speech while avoiding pumping artifacts and preserving consonants—by engineering gain staging, filtering, and “limiter vs AGC” placement as a system.
Signal Chain Decisions That Actually Matter
Think in three layers: (1) Analog headroom (preamp/PGA + HPF) → (2) Conversion quality (anti-alias + ADC) → (3) Level control policy (limiter/AGC order + time constants).
Preamp Topology Expectations (single-ended vs differential, recorder-scoped)
- Single-ended front-ends are simpler but more exposed to system noise coupling (USB attach, charger ripple, SD/CPU bursts) when the return path is noisy.
- Differential paths can better reject common-mode disturbances when routing, reference, and bias are engineered for high effective CMRR (practical benefit: fewer “charging-only” spurs reaching the ADC input).
- What to verify (minimum evidence): compare silent-record spectra on battery vs USB charging; then repeat during SD write bursts. New narrowband spurs imply coupling, not intrinsic mic noise.
PGA Step Size vs Speech Envelope: why coarse steps harm intelligibility
- Speech has fast envelope changes; large PGA steps can create audible level jumps that feel like pumping even when “AGC looks stable.”
- Coarse steps + frequent switching can smear consonants and exaggerate sibilants because the gain change overlaps the transient energy window.
- Practical mitigation: reduce switching frequency (wider hysteresis), ramp gain over short windows, or apply a brief crossfade/mute window around step transitions.
Analog HPF vs DSP HPF (handling/wind without destroying speech)
- Handling/wind rumble consumes headroom and forces limiter/AGC to act earlier, which increases pumping artifacts.
- Analog (or AFE-internal) HPF protects headroom before the ADC; DSP HPF cannot recover a transient that already clipped.
- DSP HPF stays valuable for scene tuning (handheld vs tabletop) after analog protection is in place.
Anti-Alias & ADC Knobs (speech-typical, no theory detours)
- Sample rate: choose a rate that comfortably covers speech bandwidth while controlling power; verify that the anti-alias strategy avoids high-frequency fold-back that “fizzes” consonants.
- ENOB/SNR vs power: align ADC noise with the front-end noise budget—if mic/bias/preamp dominate, overpaying for ADC SNR brings limited real-world improvement.
- Input range vs gain staging: match full-scale to prevent near-talk clipping while preserving far-talk resolution.
Limiter vs AGC: placement and failure signatures
Decision principle: protect against clipping first, then stabilize level with minimal audible modulation. Over-aggressive AGC is often worse than slightly uneven level in interviews.
- Run a 3-level speech script (quiet → normal → shout) at fixed distances.
- Record: (1) raw samples, (2) gain/limiter state logs, (3) battery vs charging comparison.
- Pass criteria: no hard clip in shout; no audible pumping in pauses; consonants remain crisp.
VAD That Doesn’t Miss Words: Features, Thresholds, and False-Reject Control
Intent: VAD is where journalist recorders win/lose usability—engineer it like a detector with evidence logs, not a marketing checkbox.
VAD Goals and the Minimum Feature Set
- Start/stop gating: decide when to open/close recording without chopping sentence starts/ends.
- Pre-roll safety net: keep a short buffer so “first syllable” survives wake + detection delay.
- Hangover control: keep recording briefly after speech ends to avoid cutting trailing words.
- Noise floor tracking: adapt thresholds as environments change (cafe, traffic, wind).
Detector Model (kept practical)
Use a small set of stable features: short-time energy (coarse gate) + spectral change (speech vs steady noise) + noise-floor tracker (environment adaptation). Keep the model explainable so logs can diagnose missed-word events.
Two Key Tunables: Attack & Hangover (trade-offs)
Pre-roll Buffer: the safety fuse (seconds vs RAM cost)
Pre-roll preserves the first syllable when wake time or detector attack is not zero. Buffer size scales with seconds × sample rate × bit depth × channels. Implement so buffered audio is committed only after a “speech confirmed” state, minimizing unnecessary file metadata churn.
Parameter Cheat-Sheet (symptom → evidence → first fix)
- Missed starts: check VAD state timestamps vs waveform onset → lower start threshold + add pre-roll.
- Cut tails: check hangover timer vs last voiced frames → increase hangover + soften stop threshold.
- Works in quiet, fails in cafe: compare noise-floor estimate drift → improve tracker smoothing; add spectral-change gate.
- Charging makes VAD worse: compare energy feature with USB spurs present → fix noise coupling upstream; update tracker to ignore narrowband spurs.
Noise Reduction (NR): Wind, Handling, HVAC Hum, and Café Clatter
Intent: Reduce real-world noise while preserving intelligibility. Define what to suppress, what must remain untouched, and how to detect artifacts with measurable evidence.
Noise Types and What They Break
- Wind & handling rumble (low-frequency, high energy): steals headroom and triggers limiter/AGC earlier.
- HVAC hum (narrowband, stationary): adds stable tones (fundamental + harmonics) that mask vowels.
- Indoor steady hiss (broadband stationary): reduces clarity in quiet words and endings.
- Café clatter (non-stationary transients): short bursts that spike loudness and confuse VAD/NR gates.
NR Stack for Recorders (pipeline mindset)
A recorder-friendly NR chain is typically: HPF (protect headroom) → Wind shaping (reduce low-frequency random energy) → Stationary suppression (steady hiss/hum) → Transient suppress (clatter/click bursts). Each stage must have explicit “speech preserve” guard rails.
Speech Preserve Guard Rails (what NOT to touch)
- Do not crush formants: preserve vowel structure to avoid “nasal / muffled” tone.
- Do not soften consonant edges: keep transient rise energy for /t k p s f/ clarity.
- Do not over-gate pauses: avoid “breathing / on-off background” that distracts listeners.
- Do not create narrowband sparkle: prevent musical noise during low-energy speech or silence.
Artifact Vocabulary (artifact → evidence → first knob)
Quick Validation (fast, recorder-realistic)
- A/B listening: same script at near and far distance; compare NR off vs on.
- Trend metrics: SNR improvement trend, pause stability (no step gating), consonant clarity check.
- Pass criteria: modest SNR gain is acceptable if intelligibility improves and artifacts remain absent.
Always-Ready Power: Standby Current, Wake Paths, and Button-to-Audio Latency
Intent: Standby life and instant capture are the product. Define power states, wake sources, and a measurable latency budget that protects the first syllable.
Power States (practical, recorder-specific)
- Deep sleep: always-on island only (RTC + button detect). Lowest standby current.
- VAD standby: low-power audio monitoring; noise-floor tracking may stay alive.
- Armed record: clocks and audio island near-ready for minimal capture latency.
- Active record: sustained audio + storage write; keep rails stable and quiet.
- USB mode: charging/data attach state; avoid injecting USB noise into audio island.
Wake Sources and Path Length
- Button wake: shortest path; target minimal latency to first valid samples.
- VAD trigger: requires continuous monitoring island; tune to avoid missed words.
- USB attach: largest state shift; define “record while charging” policy that preserves audio noise floor.
Latency Budget (segment-by-segment)
- Wake detect: interrupt → rails enable.
- Clock ready: oscillator/PLL stable.
- AFE settle: mic bias + ADC outputs valid samples.
- Record open: storage path ready and file opened.
- Buffer commit: pre-roll appended to protect first syllable.
Brownout-Safe Sequencing (always-on vs gated)
- Always-on candidates: RTC/timebase, button/wake logic, minimal PMIC state, low-rate event counters.
- Gated candidates: display/UI, high-speed storage interfaces when idle, USB PHY when detached.
- Pass criteria: after an unexpected drop, the next boot can detect incomplete sessions and avoid repeated corruption patterns.
USB-C Power Path & Fast Charge Without Recording Noise
Intent: Charging and USB switching inject noise. Engineer isolation and operating policy so “record while charging” does not ruin the noise floor.
Minimum USB-C Basics You Actually Need
- CC detect: determines attach and available default power.
- 5V default: the baseline input; the system must remain record-safe at 5V.
- Optional PD sink (concept): higher power input can change charger operating point and spur location; treat it as “more input headroom,” not a protocol tutorial.
Power-Path + Charger: System Power While Charging
A recorder typically needs a power-path that can run the system from USB while also charging the battery. The key risk is that switching current and cable events can couple into the audio analog island through shared rails, ground impedance, or reference contamination.
Three Main Coupling Loops (symptom → 2 probes → first fix)
-
Loop A: Charger ripple → analog rail/ground → AFE reference
Symptoms: stable spurs during charging; spur moves with charge mode/load.
Two probes: (1) ripple before/after analog LDO, (2) silent-recording spectrogram (charging on/off).
First fix: dedicate analog LDO + proper input filter; avoid sharing return paths with charger currents. -
Loop B: USB activity → digital burst current → ground bounce → gain chain
Symptoms: periodic buzz aligned with USB enumeration/traffic.
Two probes: (1) ground delta between digital and analog reference points, (2) audio noise vs “USB data active / charge-only” modes.
First fix: isolate digital bursts with ferrite/LC, tame return path, and apply record-mode policy. -
Loop C: Cable ESD/common-mode → protection return → ground partition breach
Symptoms: pops/clicks, sudden noise-floor shift after touch/cable swap.
Two probes: (1) ESD current return landing path check, (2) transient capture on rails + audio click counter.
First fix: correct TVS return landing (short, direct, away from analog reference); reinforce partition.
Mitigation Checklist (Hardware Isolation + Operating Policy)
- Analog LDO island: mic bias / AFE / ADC reference fed from a quiet rail.
- Ferrite/LC placement: block noise on the path from charger/digital rail toward analog island.
- Ground partition: single controlled connection point; keep charger high-current loops away from AFE reference.
- Port protection return: TVS return must land where it cannot modulate analog ground.
- Charge profile clamp: limit fast-charge aggressiveness during recording if it improves spur and click behavior.
- USB behavior: prefer charge-only during critical recording (reduce bus activity).
- Noisy blocks: postpone display refresh or heavy writes if they raise burst current during quiet speech.
- Validation loop: fixed test clip + charging modes A/B, measured against a spur threshold.
Storage Control: File Integrity, Journaling Strategy, and Power-Loss Protection
Intent: A journalist recorder that corrupts files is dead. Treat storage as a reliability subsystem: chunked writes, checkpoints, and a repeatable power-cut SOP with pass criteria.
Recording Write Pipeline (where corruption actually happens)
- Audio frames → RAM ring buffer: buffer first so capture starts before slower steps finish.
- Chunk writer: write fixed-size blocks with seq + CRC markers.
- Filesystem metadata: directory entries and headers are more fragile than raw data blocks.
- Media commit: verify “last-good checkpoint” can be found after an abrupt cut.
Safe Recording Behavior (chunk + checkpoint + marker)
- Chunked writes: lose at most the tail chunk instead of the entire file.
- Periodic checkpoints: update header/index at controlled intervals (not constantly).
- Recovery markers: store a last-good pointer (seq/offset) so recovery can truncate safely.
Brownout Strategy (what must happen in order)
- Early warning: brownout interrupt / power-good change triggers “safe close” path.
- Stop adding risk: pause high-risk metadata churn before doing final flush.
- Flush + marker: commit last chunk, then write recovery marker / checkpoint.
- Atomic step: optional atomic rename or journal entry to protect directory integrity.
Power-Cut SOP (repeatable validation + pass criteria)
- Test matrix: cut power during (1) steady recording, (2) near checkpoint, (3) heavy-write moment, (4) record-while-charge.
- Repeat: random cut timing, repeated N times, including quick reboot cycles.
- Pass criteria: file opens on device/PC, directory remains intact, tail loss bounded, recovery never loops.
Clocking & Audio Quality Under Low Power: Jitter, Drift, and Sample-Rate Stability
Intent: Keep it practical: what clock issues become audible in a recorder, how to validate fast, and how to stabilize clock domains under charger/CPU/storage aggressors.
What Users Hear (3 symptom classes)
- Pitch drift / long-take sync error: slow sample-rate offset (ppm) and temperature drift accumulate over minutes to hours.
- Sidebands / “grain” in quiet passages: short-term clock jitter or supply-modulated jitter creates spurs near aggressor frequencies.
- Warble / unstable tone (rare but real): PLL operating at the edge of lock or clock-domain switching during state changes.
Symptoms ↔ Likely Clock Cause (fast discriminator)
- Only during charging: suspect charger switching modulating the sampling clock path or its supply/reference.
- Only during SD heavy write: suspect rail droop/ground bounce disturbing PLL/clock buffers during write bursts.
- Only when CPU/UI is active: suspect burst current coupling into clock rails or reference points.
- Only over long duration: suspect temperature drift of the timebase and long-term ppm offset.
Validation Fast Pack (minimal tools, quick confidence)
- Record 30–120 minutes in a stable environment.
- Compare duration vs a known reference or repeated marker events.
- Pass if drift stays within a defined bound for the product target.
- Make a “quiet” capture (no speech) and inspect a spectrogram.
- Toggle charging, USB activity, and SD writing states.
- Spurs that align with state transitions usually indicate a coupling loop.
Mitigation (placement, isolation, and policy)
- Clock placement: keep the sampling timebase close to the AFE/ADC; shorten sensitive loops.
- Power isolation: feed clock/PLL rails from a cleaner island; block aggressor bursts with local decoupling and isolation elements.
- State stability: avoid frequent clock/PLL transitions during recording; lock behavior should remain stable across load changes.
- Record-mode policy: schedule SD writes and CPU bursts to reduce modulation during quiet passages.
EMC/ESD & Ruggedness: Survive Bags, Cables, and Static
Intent: Design robustness, not hope: map ESD entry points, control return paths, and validate with audio continuity + file integrity acceptance tests.
Threat Map (real entry points)
- USB-C shell & cable: frequent touch and insertion events.
- Buttons / seams / metal chassis: direct finger discharge and edge coupling.
- Mic jack / external mic cable (if present): long-line antenna behavior.
- Headphone/monitor port (if present): another cable-coupled entry path.
How ESD/EFT Becomes User-Facing Failures
- Pop/click in audio: return current modulates analog reference or injects into sensitive input structures.
- MCU reset / hang: discharge couples into reset rails, power-good lines, or causes rail droop.
- SD corruption: disturbance during write windows breaks metadata or interrupts commit steps.
Practical Protections (placement strategy, not parts dumping)
- TVS landing: the return path should be short, direct, and kept away from the analog reference region.
- Partitioning: keep high-energy returns out of mic AFE reference loops; control the single connection point.
- Line conditioning: add common-mode elements only where long cables create common-mode injection risk.
- Reset robustness: ensure reset and power-good signals are not fragile to fast transients.
Acceptance Tests (audio continuity + file integrity)
- ESD contact/air: hit USB shell, buttons, seams, and cable-connected ports in multiple states (idle, recording, record-while-charge, heavy-write).
- Pass: recording continues (or bounded loss), file opens on device/PC, directory remains intact.
- Audio quality: clicks/pops remain under a defined event rate / peak bound for the product target.
Validation & Field Debug Playbook: Symptom → Evidence → Isolate → Fix
Intent: A repeatable routine that maps directly to this product’s subsystems (Mic AFE / ADC / VAD / NR / Storage / Power / USB-C / Clock / EMC). Each symptom is handled with two first measurements, a clear discriminator, and a “first fix”.
- Level 0 (always): known-good USB-C cable + stable 5V adapter, a PC to play files, and a simple spectrogram/FFT tool.
- Level 1 (field engineering): DMM + USB power meter (in-line) + portable scope (1–2 channels).
- Level 2 (lab confirmation): audio analyzer / deeper FFT, programmable supply (brownout scripting), controlled ESD source for repeatability.
Tip: every test run should include an event marker (button tap or short beep) so audio artifacts can be aligned to USB attach / SD write / state changes.
Quick Jump (Top Symptoms)
Example MPNs (reference parts for fast prototyping)
These are example part numbers used as anchors when building and debugging. Final selection depends on noise targets, IO count, footprint, and supply chain.
Symptom: “First syllable missing” (start-of-speech clipped)
Most likely subsystems: VAD attack / pre-roll buffer / wake latency / AFE settle / file open latency
- First 2 measurements
- Event-align: add a marker (button tap/beep) at record start; measure time to first valid samples in the waveform.
- State timing: log timestamps for wake → clocks stable → AFE enabled → file open → encoder start (or the closest available counters).
- Discriminator
- If the clipped portion is a fixed time (e.g., always ~150–300 ms), it usually points to wake/AFE/file-open latency.
- If it happens only with VAD enabled and varies with background noise, it usually points to VAD attack / threshold / pre-roll too short.
- First fix (fast)
- Firmware/params: enable pre-roll (0.3–1.0 s typical), reduce VAD attack time, add hangover, and start capturing immediately into a ring buffer before file commit.
- Hardware/layout: ensure mic bias + AFE rails settle quickly; avoid gating the analog reference rail in “armed record”. Consider keeping AFE bias alive in standby if product allows.
- Example MPN anchors
- Mic preamp/AGC reference: MAX9814 (helps validate AGC/pumping vs start clipping)
- Codec/ADC reference: TLV320AIC3254 or PCM1863A
- Low-noise always-on analog rail: TPS7A20 (analog island LDO)
Symptom: “Hiss increases when charging” (record-while-charge noise)
Most likely subsystems: charger ripple → ground/rail coupling → AFE/ADC clock/reference; USB activity bursts
- First 2 measurements
- A/B FFT: capture a quiet segment on battery vs charging; compare spurs and wideband floor.
- Rail probe: scope the analog rail (AFE/ADC supply) and charger node; note ripple frequencies and load-step bursts.
- Discriminator
- If noise shows fixed spurs that match a switching frequency/harmonics, suspect charger coupling.
- If noise increases mainly during USB activity (file transfer / UI), suspect digital burst coupling into analog reference/clock.
- First fix (fast)
- Firmware/policy: add a dedicated “record-while-charge” mode (limit CPU bursts, schedule storage writes, reduce display refresh if present).
- Hardware/layout: split analog supply via low-noise LDO; add isolation element between charger/system rails; shorten return loops; keep charger hot loops away from AFE/clock.
- Example MPN anchors
- Charger: BQ25895 or BQ25601D
- Power-path / source selection: TPS2121
- Analog LDO: TPS7A20 (quiet rail for AFE/clock reference)
- USB ESD/EMI helper: TPD4E05U06 + (USB2 CM choke) ACM2012-900-2P
Symptom: “Random stop” / “File won’t open”
Most likely subsystems: brownout during write, file commit strategy, storage ESD/EMC disturbance, reset robustness
- First 2 measurements
- Power-cut script: repeat controlled power interruptions during recording (especially during SD writes); track corruption rate.
- Flags/logs: capture reset/brownout indicators and the last “file-close marker” status (or last-good checkpoint index).
- Discriminator
- If failures cluster around write windows and a rail droop is visible, it’s likely power integrity / brownout.
- If the device keeps running but the file is broken, it’s likely commit/journaling strategy (metadata not protected) or media/connector issues.
- First fix (fast)
- Firmware: use chunked writes + periodic recovery markers; delay metadata updates until checkpoints; atomic rename for finalization; keep a small “last-good index”.
- Hardware: improve rail hold-up for the storage rail; separate storage supply from noisy loads; protect SDIO lines; add ESD arrays at the connector.
- Example MPN anchors
- Storage line ESD: TPD4E1U06
- Checkpoint memory (optional): MB85RS2MT (SPI FRAM for last-good markers)
- Charger/power: BQ25895 + power mux TPS2121
Symptom: “VAD misses quiet speaker” (false reject)
Most likely subsystems: VAD threshold vs noise tracking, attack/hangover, pre-roll length, NR interaction
- First 2 measurements
- State trace: log VAD state transitions (Idle/Pre-roll/Record/Hangover) alongside a known quiet speech test clip.
- Noise floor capture: measure the estimated noise floor over time (or proxy: short-time energy distribution) in the target environment.
- Discriminator
- If VAD fails only in certain noise types (HVAC hum / café clatter), the noise estimator is drifting or features are biased.
- If VAD improves immediately by raising gain but adds pumping, the issue is often gain staging / coarse PGA steps before detection.
- First fix (fast)
- Parameters: lower start threshold slightly, increase hangover, and ensure pre-roll covers the expected onset; tune noise floor update rate.
- System: ensure the detection path sees a stable, unclipped signal; avoid NR over-suppressing consonant edges before VAD features are computed.
- Example MPN anchors
- Front-end gain staging reference: MAX9814 (AGC behavior baseline)
- ADC/codec reference: TLV320AIC3254 / PCM1863A
Symptom: “Clicks during SD writes” (artifact aligned to write bursts)
Most likely subsystems: storage burst current → ground bounce → AFE reference/clock modulation; insufficient analog isolation
- First 2 measurements
- Event align: mark write bursts (log “write start/end” or flash an LED) and align clicks to those events in the waveform.
- Rail probe: scope analog rail and storage rail during writes; look for correlated dips/spikes.
- Discriminator
- If clicks line up with write bursts and a rail spike is present, it is a power/return-path coupling problem (not NR/VAD).
- If clicks occur even when not writing, suspect ESD/EMI ingress or mechanical contact issues.
- First fix (fast)
- Firmware: batch writes into fewer bursts; avoid writes during quiet passages; add buffering and schedule commits.
- Hardware/layout: isolate analog rail with a quiet LDO; keep SD high-current return away from AFE ref; add line ESD near the connector.
- Example MPN anchors
- Analog LDO: TPS7A20
- SD line ESD: TPD4E1U06
- Power-path: TPS2121
Symptom: “Pop/click or reset after ESD” (USB shell / buttons / seams)
Most likely subsystems: ESD entry + return path crossing analog island; reset/power-good fragility; storage disturbance during write
- First 2 measurements
- Controlled hit map: apply repeatable discharge to USB shell, buttons, seams, and cable ports while recording; note whether audio pops, resets, or file damage occurs.
- Return-path observation: correlate failure with touch point + cable presence; failures that depend on cable/hand position usually indicate a return-path problem.
- Discriminator
- If the unit resets but audio path is otherwise clean, focus on reset/power-good/transient immunity.
- If pops occur without reset, focus on analog reference injection (return crossing AFE ref).
- If files corrupt, focus on storage write window protection (commit markers + rail hold-up).
- First fix (fast)
- Hardware/layout: move TVS to create a short local return to chassis/ground landing; keep discharge currents out of analog reference loops; add ESD arrays to exposed ports; minimize loop area.
- Firmware: during recording, reduce state transitions; ensure recovery markers allow partial file recovery after transient events.
- Example MPN anchors
- USB ESD array: TPD4E05U06 / USB-C array: TPD4E02B04
- SD line ESD: TPD4E1U06
- USB2 CM choke (when needed): ACM2012-900-2P
Decision Tree (text SOP)
- Start: choose the symptom group → run the two “first measurements”.
- Q1: Does the issue appear only while charging? → Yes: isolate to charger/USB coupling (rail FFT + layout loops).
- Q2: Do artifacts align with SD write bursts? → Yes: isolate to storage burst coupling (rails + batching policy).
- Q3: Any reset/brownout flags around the event? → Yes: isolate to power sequencing/hold-up (brownout scripting).
- Q4: Is audio clipped/saturated near start or loud speech? → Yes: isolate to AFE headroom/gain staging (preamp/PGA/limiter).
- Q5: Does VAD state show false reject under quiet speech? → Yes: isolate to VAD thresholds/noise tracking (attack/hangover/pre-roll).
- Q6: Do pops/resets occur on touch/ESD points? → Yes: isolate to ESD return path (TVS landing + chassis return).
- End: apply the “first fix” (params/policy) → re-run the same measurement to confirm closure.
Request a Quote
FAQs (Accordion ×12): Evidence-first Fixes Without Scope Creep
Rule: Each answer gives two first checks, a clear discriminator, and a first fix. No cloud backend, no multi-track field recorder scope.
1 First word is clipped when I press record — wake latency or AFE settling?
First check: align an event marker and measure button-to-first-valid-samples; then log wake → clocks stable → AFE enabled → file open timing. If the missing chunk is a fixed 150–300 ms, it is latency/settling; if it varies with noise, it is VAD attack/pre-roll. First fix: start a 0.3–0.8 s ring-buffer pre-roll and keep the analog rail on (e.g., TPS7A20).
2 VAD misses quiet speakers but records noise — which 2 thresholds to tune first?
First check: record VAD state transitions with a quiet speech clip; then review the estimated noise floor (or energy histogram) in the target environment. Tune (1) start threshold vs tracked noise floor and (2) hangover/end threshold. If VAD chatters, noise tracking is too fast; if it never triggers, start threshold is too high. First fix: slow noise update, lower start threshold 1–2 dB, add 200–500 ms hangover, and verify gain staging (e.g., MAX9814 baseline).
3 Hiss gets worse only when charging — ground coupling or charger ripple? What to measure?
First check: A/B FFT on battery vs charging; then scope the analog rail and charger switch node to spot ripple/spurs. If spurs sit at switching frequency/harmonics, it is charger ripple coupling; if noise rises during USB activity bursts, it is digital/ground return injection. First fix: isolate the analog island with a quiet LDO (e.g., TPS7A20), keep charger hot-loop away, and use a robust power-path (e.g., TPS2121) with a charger like BQ25895.
4 Clicks every few seconds — SD write bursts or clock spur?
First check: event-align clicks to SD write timestamps (or LED/log); then run an FFT to see whether the artifact is a spur or a transient pop. If clicks line up with writes and rails dip, it is storage burst coupling; if a fixed spur remains even without writes, it is clock/switching interference. First fix: batch writes, increase buffering, isolate storage/analog returns, and protect SD IO near the connector (e.g., TPD4E1U06).
5 Record stops early, file corrupted — brownout or FS metadata update?
First check: read reset/brownout flags around the stop; then do a controlled power-cut test during writes and measure corruption rate. If rail droops coincide with stops, it is brownout/hold-up; if the device stays alive but the file is broken, it is commit/journaling strategy. First fix: write in chunks with periodic recovery markers and atomic finalize; optionally store last-good checkpoints in FRAM (e.g., MB85RS2MT) and harden power-path (e.g., TPS2121).
6 Wind noise overwhelms speech outdoors — which filter order works best for intelligibility?
First check: confirm wind energy dominates below ~200–300 Hz via a spectrogram; then verify the AFE is not clipping on gusts. Best order is: mechanical windscreen → analog/digital HPF → wind detector → gentle suppression, keeping speech formants intact. If suppression comes before HPF, NR often pumps and smears consonants. First fix: place HPF early, cap low-band attenuation, and keep mid-band untouched; validate with A/B and word intelligibility scoring.
Mapped to: H2-5
7 Speech sounds “watery” after NR — which knob reduces musical noise first?
First check: view a spectrogram for “random tonal bins” and listen to quiet speech tails; then compare noise-only segments to detect over-subtraction. If watery artifacts worsen in low SNR, the suppression floor is too aggressive and smoothing is too light. First fix: raise the suppression floor, increase time/frequency smoothing, and slow attack/release; add comfort noise if available. If NR runs before VAD features, verify VAD is not being starved by over-suppression.
Mapped to: H2-5
8 Loud laughter clips but normal speech is fine — gain staging or limiter location?
First check: probe AFE output and inspect ADC codes for hard clipping; then repeat with a fixed loud input to see whether clipping happens pre-ADC or post-ADC. If clipping is already on the analog waveform, reduce analog gain or increase headroom; if only in the digital path, move/retune the limiter earlier. First fix: lower preamp/PGA, use a soft-knee limiter before compression, and validate with shout/laugh bursts (e.g., PCM1863A as a clean ADC reference).
Mapped to: H2-3
9 Long recordings drift in pitch/time — clock drift or sample-rate mismatch? Quick test?
First check: record a long reference tone and compare pitch at start vs end; then compare file duration to RTC timestamps. If pitch shifts, the audio clock is drifting; if duration mismatch grows without pitch change, sample-rate metadata or resampling is inconsistent. First fix: use a stable clock source and keep it isolated from charger/noisy domains; confirm PLL lock behavior is stable. Anchor parts: RTC crystal FC-135 and a stable MEMS oscillator (e.g., SiT1602).
Mapped to: H2-9
10 Buttons cause pops in audio — mechanical coupling or analog bias disturbance?
First check: event-align pops to button edges and compare near-field mechanical taps vs actual switch actuation; then scope mic bias and AFE input to see if a bias step coincides. If bias droops/steps at the pop, it is electrical coupling/return-path; if only physical taps create it, it is mechanical/acoustic coupling. First fix: debounce and slow bias transitions, add RC damping on mic bias, and separate button return from the analog reference path; add ESD protection on exposed lines (e.g., TPD4E05U06).
11 USB plug-in causes reboot or recording drop — inrush/UVLO or ESD?
First check: measure VBUS inrush and system rail dip at plug-in; then read UVLO/brownout/reset flags. If rails dip below UVLO, it is inrush/power-path switching; if pops/resets depend on touch point/cable position, it is ESD return-path. First fix: add a robust power mux/soft-start (e.g., TPS2121), raise UVLO margin, and place USB-C TVS with a short return (e.g., TPD4E02B04). Validate by repeated plug cycles while recording.
12 Standby battery drains fast — armed VAD current, mic-bias leakage, or periodic SD wake?
First check: measure sleep current by state (deep sleep vs “armed” VAD standby vs USB attached); then log periodic wake sources (SD polling, UI timers, radio beacons). If current jumps only in armed mode, mic bias/AFE is leaking; if it spikes periodically, background storage/housekeeping is waking the system. First fix: gate mic bias unless needed, batch housekeeping, and use low-Iq rails for always-on domains (e.g., TLV755P), while keeping charger behavior predictable (e.g., BQ25895).
