DAYROVE ACOUSTIC & MEMORY SYSTEMS LAB · REV 2026.11

Applied Research in Wrist-Scale Acoustics, On-Device Quantized ASR & Verifiable Personal Memory.

How do you capture hours of spoken life on a 38mm watch battery, run multilingual speech-to-text strictly inside an iPhone’s Neural Engine without cloud uploads, and guarantee zero hallucinated citations when querying months of personal memory? This is our open engineering notebook.

BENCHMARK // DR-ASR-INT8

On-Device Neural Engine ASR Trade-off Matrix

Switch spoken language profile to compare speed factor, active memory footprint, and local checkpoint granularity across our three bundled/downloadable engines.

TARGET SILICON: APPLE A17 PRO / A18 NEURAL ENGINE · 45s ATOMIC BATCHES
PROTOCOL // WATCHCONNECTIVITY-2PC

Zero-Loss Two-Phase Watch Sync & Fault Injection

WatchOS deletes temporary transfer files aggressively and Bluetooth links drop mid-walk. Dayrove enforces a strict database-grade invariant: Watch audio is never eligible for local cleanup until the paired iPhone copies the payload out of ephemeral storage, verifies audio playability + SHA-256 hash, commits the main-actor index, and returns a signed receipt ACK.

STAGE 01 · WATCH 5m WAL Segment + Disk Guard (>50 MiB)
STAGE 02 · QUEUE Persistent Outbox & Retry Backoff
STAGE 03 · IPHONE Immediate Safe Copy & Frame Validation
STAGE 04 · INDEX Transactional Merge + Tombstone Check
STAGE 05 · RECEIPT Cryptographic Save ACK → Unlock Watch
[2PC-OK] clip_id=seg_03 · sha256=verified · iphone_index=committed · watch_ack=CONFIRMED
MONOGRAPHS // ARCHITECTURE

Core Engineering Notes

DR-NOTE-01 · ACOUSTIC CONTINUITY

Stop-Priority Generation Tokens & Call Interruption Safety

Preventing "zombie auto-resumes" on watchOS requires generation-stamped recording tickets. When a user taps Stop, all pending microphone warmups, segment-boundary handovers, and late audio-session interruption callbacks are atomically invalidated.

Duration(session) = Σ actual_frames(seg_i) // Gaps never fabricated
DR-NOTE-02 · VERSIONED DERIVATION

Immutable Source Audio & Sidecar Span Re-Transcription

Original audio is immutable truth. Transcripts live in versioned sidecar files (≤16 MiB bounded shards). Long-pressing a misheard paragraph re-runs Whisper/Qwen on just that audio time-range and presents a side-by-side diff while preserving parent versions.

Transcript_v(n+1) = Patch(Transcript_v(n), Span[t0, t1]) → AudioRef[clipID]
DR-NOTE-03 · VERIFIABLE MEMORY GRAPH

Deterministic Counting + Citation-Verified RAG

LLMs alone miscount exact occurrences. Dayrove routes counting & keyword queries through deterministic local full-text scan first, then synthesizes context with mandatory `[Date · ClipID · Timestamp]` anchors verified against local audio existence before rendering.

Verify(Citation) ∈ LocalAudioIndex ? PlayableJump(t - 2s) : DegradeGracefully
DR-NOTE-04 · BIOMETRIC INTEGRITY

Why We Report Raw SDNN HRV Instead of Invented Stress Scores

Wrist HRV (SDNN in ms) varies with posture, movement, and sparse sampling. Rather than inventing opaque 0–100 "stress numbers", Dayrove displays exact sample counts, time windows, and raw millisecond readings aligned honestly alongside your voice timeline.

HRV_Context = { mean_sdnn_ms, sample_count, window_utc } // Read-Only HealthKit