Mimir

Realtime cameras, microphones, chirplets, and field evidence organized into one coherent stream machine.

Bioacoustic Training Harness

3 min read

Bioacoustic Training Harness

Purpose

The bioacoustic receiver is tuned by benchmark receipts, not by console optimism. Mimir.BufferSmoke --bioacoustic-train runs a small hypothesis panel over the current motif language and records both queryable CultCache evidence and audio artifacts.

This is authoring and calibration machinery. It is allowed to allocate, score, serialize, and leave receipts. It is not the runtime hot loop.

Command

dotnet run --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --bioacoustic-train --sample-rate 48000 --seconds 0.5 --output artifacts/bioacoustic-training

Each run writes:

  • bioacoustic-training.cc: typed CultCache results with provenance, decoder settings, degradation name, scores, timing, and artifact hashes.
  • run-summary.json: short best-result summary.
  • source-pre-warp.wav: the emitted source song.
  • <hypothesis>/<degradation>/pre-warp.wav: source audio for that scored case.
  • <hypothesis>/<degradation>/post-warp.wav: mel-cepstral warp/blur round-trip.
  • <hypothesis>/<degradation>/reconstructed-from-detections.wav: song rebuilt from detected word features.
  • <hypothesis>/<degradation>/summary.json: per-case receipt mirror.

Current Hypotheses

  • baseline-mfcc-index: current indexed MFCC/log-mel smoke shape.
  • compact-fast-index: smaller feature and hash surface to test speed against identity loss.
  • robust-wide-index: wider cepstral/augmentation surface to test degraded-path survival.
  • highband-room-index: upper-band-biased receiver to avoid low room junk.

The harness intentionally scores identity, timing, and speed separately:

total = identity * 0.62 + timing * 0.18 + speed * 0.20

Aggregate score is a triage knife. Identity tells whether the word language survived. Timing tells whether detected words landed on the canonical clock. Speed tells whether the receiver shape can plausibly become a streaming machine.

Latest Local Receipt

Run:

artifacts/bioacoustic-training/bioacoustic-20260524-150729/

This run adds global clock hypothesis scoring. Timing is no longer raw per-word absolute error; it is the confidence of a coherent clock/path fit over detected words. Single-anchor fits are deliberately weak, and sample-rate slope is only allowed when enough anchors exist.

Best aggregate result:

compact-fast-index / clean-roundtrip
total=0.752 identity=0.908 timing=0.905 speed=0.132 realtime=6.6x

Useful observations:

  • compact-fast-index is still the clean-throughput winner when enough anchors survive.
  • baseline-mfcc-index is still the steadier degraded-path receiver.
  • highband-room-index shows real value under warp, but needs more anchor density before it can own timing.
  • compact-fast-index remains faster, but its warped-path identity and timing confidence are weaker.
  • robust-wide-index still costs too much for the current panel, but it gives useful pressure under heavier degradation.
  • The global clock hypothesis is the right scoring authority, but short windows and weak degradations can still starve it. Future training runs should score time-to-lock separately from identity. Anchor coverage is now persisted in the CultCache clock snapshot so optimization can target lock density directly.

Previous receipt:

Run:

artifacts/bioacoustic-training/bioacoustic-20260524-103438/

Best aggregate result:

compact-fast-index / clean-roundtrip
total=0.700 identity=0.997 timing=0.357 speed=0.085 realtime=4.2x

Useful observations:

  • compact-fast-index is faster on clean/light blur, but collapses under warp.
  • baseline-mfcc-index keeps better identity under light warp than compact.
  • highband-room-index wins some blur/warp-blur cases, which supports room-path band selection as a real tuning axis.
  • robust-wide-index is not currently worth its extra CPU except as pressure for heavier blur/warp cases.
  • Timing mean absolute error is still poor under warped cepstral reconstruction. That means the next cut is global delay/clock/path hypothesis over detected words, not just “more cepstra.”

Architecture Pressure

The benchmark now uses CultLib at the BufferSmoke edge, but Mimir itself should be a CultMesh app. Phones, microcontrollers, Raven, and Starfire need typed state surfaces for schedule/codebook snapshots, calibration receipts, learned path response, and work offload. CultCache receipts are the durable proof layer; the next runtime cut should promote the live state contracts into mesh-owned documents rather than leaving them as harness-only artifacts.

Backlinks