Mimir

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

Bioacoustic Contestants 2026-05-24

15 min read

Bioacoustic Contestants 2026-05-24

Purpose

This pass moves the harness from “one current song, many decoder knobs” toward actual contestants: different generated birdcall word shapes fight against the same indexed cepstral decoder family and the same cepstral warp/blur damage.

The original score was intentionally brutal:

performance = sqrt(clamp(realtime / 50x, 0..1) * convergence)
anchor      = sqrt(timing_accuracy * frequency_accuracy)
score       = performance * anchor
  • convergence is clock confidence times anchor coverage.
  • timing_accuracy falls with clock-fit residual in samples.
  • frequency_accuracy is current word identity precision; exact word identity stands in for the emitted frequency contour until the decoder exposes per-formant residuals.
  • realtime is managed harness speed, not the final native SIMD/DSP shape.

Contestants

  • current-birdcall: current four-syllable formant word.
  • redpoll-trill: fast repeated chips with strong temporal onsets.
  • robin-warble: softer irregular phrase, wider curved formant motion.
  • thrush-ladder: repeated interval ladder with a shifted reply.
  • thornbill-zigzag: high-band fast zig-zag syllables.
  • nightingale-cascade: dense pretty cascade; not yet in the fast receipt.
  • aquasynth-formant-weaver: AquaSynth-shaped formant/FM word with payload ornaments.
  • canary-packet-trill: six-syllable packet call with hard timing chips and payload-colored formant notes.
  • finch-burst-packet: shorter five-syllable boundary candidate. Useful as a speed/blur failure witness, not the current recommendation.

Language Score Update

The harness now reports the score the project actually cares about:

language_score = realtime_multiplier
               * timing_accuracy
               * frequency_accuracy
               * payload_bitrate_bps

This is raw enough to hurt. It rewards a receiver that is fast, places anchors accurately in time, preserves frequency/shape identity, and recovers data-bearing words. payload_bitrate_bps is computed from the profile’s payload bits per event, event spacing, and payload accuracy.

Important cut: payload accuracy is now classified from the observed word shape. For each anchored event, the harness compares the observed feature against that event’s payload alphabet and records the classified payload symbol. This is slower and meaner than treating event identity as payload recovery, but it is the first score that deserves to be called bitrate evidence.

The score also stopped using word precision as a fake frequency score. Each decoded observation now carries a cepstral shape accuracy from the winning template distance. That is still a compact proxy, but it is pointed at the frequency/formant surface instead of identity bookkeeping.

Clock fitting now uses a global propagation-delay hypothesis. Candidate anchors vote for one delay, and only anchors coherent with that delay are allowed to own the clock fit. Observations outside the expected fully captured window remain diagnostics; they no longer poison the timing kernel. This fixed a real harness bug where truncated end-of-window calls were being counted as expected anchors.

The first fast panel intentionally tested only the first four contestants, two decoders, and clean/blur degradations so the loop could produce receipts inside one turn.

Fast Receipt

Run:

dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --bioacoustic-contestants --seconds 0.5 --max-songs 4 --max-decoders 2 --max-degradations 2 --output artifacts/bioacoustic-contestants

Receipt:

artifacts/bioacoustic-contestants/contestants-20260524-180639/contestant-summary.json

Best result:

song=robin-warble
decoder=compact-fast-index
degradation=blur-light
score=0.178
performance=0.178
anchor=1.000
timing=1.000
frequency=1.000
convergence=0.485
realtime=3.3x
correct=2/3

Read: robin-warble produces fewer anchors in the short window, but the anchors that survive are clean. The current score rewards that more than many sloppy anchors. This may be correct for clock acquisition, but the next pass should separate “lock fast” from “lock beautifully once found” so the harness cannot overvalue sparse perfection.

Packet Contest Receipts

Spectral Receipts

The current canary packet has rendered spectral witnesses under artifacts/spectra/canary-packet-192847/:

  • source-spectrogram-linear.png: linear-frequency view.
  • source-spectrogram-log.png: log-frequency view, closer to the perceptual shape we keep talking about.
  • source-spectral-peaks.csv: per-frame peak bins for quick inspection.

The log spectrogram shows the actual packet anatomy: hard vertical onset gaps, two lower timing chips, then payload-colored formant ladders. The early peak CSV shows the first anchor band climbing around 2.67-3.12 kHz between 0.040s and 0.070s, which is exactly the kind of explicit spectral witness the harness should have been producing from the beginning.

Current Honest Payload Receipt

Run:

dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --bioacoustic-contestants --seconds 0.75 --song canary-packet-trill --decoder packet-razor-index --max-songs 1 --max-decoders 1 --max-degradations 5 --output artifacts/bioacoustic-contestants

Receipt:

artifacts/bioacoustic-contestants/contestants-20260524-192847/contestant-summary.json

Best result with independent payload classification in the decoder sweep:

song=canary-packet-trill
decoder=packet-razor-index
degradation=blur-light
language_score=88.380
realtime=5.6x
timing=0.860
frequency=0.957
payload_bitrate=19.0 bps
payload_accuracy=1.000
correct=6/6

Damage panel:

packet-razor clean-roundtrip   score=88.129 payload_bitrate=19.0 bps payload=1.000 timing=0.859 frequency=0.931
packet-razor blur-light        score=88.380 payload_bitrate=19.0 bps payload=1.000 timing=0.860 frequency=0.957
packet-razor warp-light        score=65.172 payload_bitrate=17.4 bps payload=0.913 timing=0.872 frequency=0.776
packet-razor warp-light-blur   score=66.679 payload_bitrate=13.5 bps payload=0.707 timing=1.000 frequency=0.922
packet-razor warp-heavy-blur   score=81.515 payload_bitrate=13.5 bps payload=0.707 timing=1.000 frequency=0.922

This is the current honest floor: a 2-bit payload alphabet carried by the canary packet. Payload classification is now direct waveform/template correlation over the anchored local payload alphabet, not MFCC identity reuse. That raised the honest score and made payload survive all current warp/blur cases. Tightening canary-packet-trill spacing from 0.125s to 0.105s improved bitrate and the degradation floor. A 0.100s spacing was tested and rejected because overlap started eating anchor accuracy. A 3-bit variant did not earn its keep; clean was similar, but degraded payload recovery collapsed. The earlier 8-bit and 4-bit readings were schedule-entangled and are no longer treated as bitrate evidence.

The leaderboard is now split:

  • clean result: canary-packet-trill + packet-razor-index, 88.129
  • best overall: canary-packet-trill + packet-razor-index under blur-light, 88.380
  • remaining failure: warp-light frequency accuracy is still low at 0.776, but timing and payload now stay alive.
  • rejected 3-bit dual-axis packet: best 12.849, clean 8.084, heavy warp nearly dead. A separate band/rhythm bit did not survive well enough to keep in the built-in contestant panel.

Log-Mel Flux / Log-Mel Leak Receipt

The arena now includes two separate damage domains:

  • cepstral warp/blur: the older MFCC-shaped leaky pipeline.
  • log-mel warp/blur: a perceptual-surface leak that bends and blurs the log-mel field before reconstruction through the original STFT phase.

The decoder panel also tested two flux-proposal variants:

  • packet-razor-flux-index: same packet razor identity surface, but candidate starts come from positive log-mel spectral-flux peaks.
  • packet-sprint-flux-index: a wider sprint variant with the same flux proposal owner.

Runs:

dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --bioacoustic-contestants --seconds 0.75 --song canary-packet-trill --decoder packet-razor-index --max-songs 1 --max-decoders 99 --max-degradations 9 --output artifacts/bioacoustic-contestants
dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --bioacoustic-contestants --seconds 0.75 --song canary-packet-trill --decoder packet-razor-flux-index --max-songs 1 --max-decoders 99 --max-degradations 9 --output artifacts/bioacoustic-contestants
dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --bioacoustic-contestants --seconds 0.75 --song canary-packet-trill --decoder packet-sprint-flux-index --max-songs 1 --max-decoders 99 --max-degradations 9 --output artifacts/bioacoustic-contestants

Receipts:

  • artifacts/bioacoustic-contestants/contestants-20260524-212432/contestant-summary.json
  • artifacts/bioacoustic-contestants/contestants-20260524-212451/contestant-summary.json
  • artifacts/bioacoustic-contestants/contestants-20260524-212518/contestant-summary.json

Best result in this panel:

song=canary-packet-trill
decoder=packet-razor-index
degradation=logmel-warp-light-blur
language_score=92.522
realtime=5.7x
timing=1.000
frequency=0.939
payload_bitrate=17.4 bps
payload_accuracy=0.913
correct=5/6

Flux-proposal results:

packet-razor-flux logmel-warp-light      score=43.577 realtime=2.7x payload_bitrate=17.4 bps timing=1.000 frequency=0.925 correct=5/6
packet-razor-flux clean-roundtrip        score=33.466 realtime=2.2x payload_bitrate=19.0 bps timing=0.862 frequency=0.930 correct=6/6
packet-sprint-flux logmel-warp-light-blur score=26.761 realtime=2.0x payload_bitrate=14.7 bps timing=1.000 frequency=0.926 correct=5/6

Read: spectral flux is a good biological witness and a plausible proposal owner, but the naive managed implementation does not yet win the arena. It spends too much time computing log-mel flux separately from the indexed fingerprint pass. The shape should not be discarded; it should be fused into the existing feature extraction so each STFT frame emits both log-mel/MFCC identity and positive-flux onset evidence. Current best transmission through the leaky pipeline remains canary-packet-trill + packet-razor-index, now at 92.522 on log-mel warp/light blur.

Streaming Packet Razor Receipt

The next cut stopped treating the packet receiver like an offline search problem. packet-razor-streaming-faust is schedule-owned and packet-local:

  • no dense MFCC/LSH proposal sweep;
  • no second FFT pass for flux;
  • no wide identity table lookup;
  • only scheduled packet windows are inspected;
  • each window tests the four local payload hypotheses;
  • timing is refined with a tiny parabolic local correction;
  • the Faust lowering target is faust/mimir_packet_razor_frontend.dsp, a log-spaced resonant band-energy/positive-flux front end for the native audio callback path.

Run:

dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --bioacoustic-contestants --seconds 0.75 --song canary-packet-trill --decoder packet-razor-streaming-faust --max-songs 1 --max-decoders 99 --max-degradations 9 --output artifacts/bioacoustic-contestants

Receipt:

artifacts/bioacoustic-contestants/contestants-20260524-214530/contestant-summary.json

Best result:

song=canary-packet-trill
decoder=packet-razor-streaming-faust
degradation=logmel-blur-light
language_score=1121.012
realtime=62.6x
timing=0.995
frequency=0.944
payload_bitrate=19.0 bps
payload_accuracy=1.000
correct=6/6

Nine-leak panel:

clean-roundtrip          score=798.225  realtime=44.7x payload=1.000 timing=0.993 frequency=0.943 correct=6/6
blur-light               score=720.550  realtime=56.8x payload=1.000 timing=0.995 frequency=0.669 correct=6/6
warp-light               score=527.312  realtime=55.7x payload=1.000 timing=0.862 frequency=0.577 correct=6/6
warp-light-blur          score=749.635  realtime=60.0x payload=1.000 timing=0.996 frequency=0.658 correct=6/6
warp-heavy-blur          score=672.210  realtime=57.4x payload=1.000 timing=0.996 frequency=0.618 correct=6/6
logmel-blur-light        score=1121.012 realtime=62.6x payload=1.000 timing=0.995 frequency=0.944 correct=6/6
logmel-warp-light        score=1066.350 realtime=68.2x payload=1.000 timing=0.995 frequency=0.826 correct=6/6
logmel-warp-light-blur   score=872.989  realtime=55.0x payload=1.000 timing=0.996 frequency=0.837 correct=6/6
logmel-warp-heavy-blur   score=992.207  realtime=68.2x payload=1.000 timing=0.997 frequency=0.767 correct=6/6

This is the current honest arena leader by more than an order of magnitude over the earlier 92.522 best. The architectural lesson is stronger than the number: once the receiver is streaming and schedule-aware, the problem becomes local payload discrimination over a tiny candidate set. Wide MFCC search is an acquisition/reference tool, not the hot loop. The next non-cheating production cut is an acquisition state that uses Faust/front-end band flux to find the first clock hypothesis, then switches to scheduled packet windows exactly like this receipt.

First Physical ASIO Packet Receipt

The first physical run used the current Focusrite ASIO path:

dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --render-contestant-f32 --song canary-packet-trill --output artifacts/asio/canary-packet-192k-f32.raw --sample-rate 192000 --seconds 4
.\native\probes\asio_audio_cadence\build\Release\asio_audio_cadence.exe --set-sample-rate 192000 --play-f32-mono artifacts\asio\canary-packet-192k-f32.raw --record-f32-interleaved artifacts\asio\scarlett-canary-packet-192k-f32.raw --play-gain 1.0
dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --analyze-contestant-asio-f32 --input artifacts\asio\scarlett-canary-packet-192k-f32.raw --sample-rate 192000 --channels 4 --seconds 4 --song canary-packet-trill --schedule-offset-samples 1623

The first ASIO attempt accidentally ran at the driver’s existing 44100 Hz state despite the requested sample rate. The corrected run used --set-sample-rate 192000, captured 765,952 frames at 192 kHz across four inputs, and found the loopback packet by correlation at about +1623 samples (8.453 ms). With that acquisition offset supplied, the streaming packet receiver decoded the loopback channels perfectly:

channel=2 events=37/37 payload=1.000 timing=1.000 confidence=1.000 delaySamples=1623.000000 delayUs=8453.125 maeSamples=0.000000
channel=3 events=37/37 payload=1.000 timing=1.000 confidence=1.000 delaySamples=1623.000000 delayUs=8453.125 maeSamples=0.000000

The acoustic inputs did not yet prove the room path:

channel=0 events=0/37 payload=0.000 timing=0.000 confidence=0.000
channel=1 events=1/37 payload=0.000 timing=1.000 confidence=0.545 delaySamples=1587.500000 delayUs=8268.229

Interpretation: the packet machine survives the real Focusrite digital loopback path at 192 kHz. The room/mic path needs either more acoustic level, better mic channel state, or a front-end acquisition pass that can lock before scheduled packet-local decoding takes over. The important architectural result still stands: acquisition and tracking are separate states. Once the clock offset is known, the streaming packet hot loop is tiny and exact.

Second Physical ASIO Packet Receipt

After placing both microphones, the same 192 kHz packet run was repeated:

dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --render-contestant-f32 --song canary-packet-trill --output artifacts/asio/canary-packet-192k-f32.raw --sample-rate 192000 --seconds 4
.\native\probes\asio_audio_cadence\build\Release\asio_audio_cadence.exe --set-sample-rate 192000 --play-f32-mono artifacts\asio\canary-packet-192k-f32.raw --record-f32-interleaved artifacts\asio\scarlett-canary-packet-192k-rerun-f32.raw --play-gain 1.0

Capture:

sampleRate=192000 frames=765952 channels=4
input[0]="Input 1" input[1]="Input 2" input[2]="Loopback 1" input[3]="Loopback 2"
source rms=0.00694 peak=0.02913
ch0 rms=0.02612 peak=0.32488
ch1 rms=0.01530 peak=0.21927
ch2/ch3 rms=0.00692 peak=0.02913
correlation lags: ch0=2519, ch1=2122, ch2=1623, ch3=1623 samples

With the loopback acquisition offset (1623 samples), loopback remained perfect and the two mic channels showed partial lock:

ch0 events=14/37 payload=0.162 timing=0.745 delayUs=8500.880
ch1 events=23/37 payload=0.135 timing=0.706 delayUs=8445.016
ch2 events=37/37 payload=1.000 timing=1.000 delayUs=8453.125
ch3 events=37/37 payload=1.000 timing=1.000 delayUs=8453.125

With per-channel acquisition offsets, the co-streamer shotgun mic on channel 1 locked:

ch1 offset=2122 events=37/37 payload=1.000 timing=0.828 delayUs=11166.397 maeSamples=9.952719

The cardioid/near mic on channel 0 improved but remained unreliable:

ch0 offset=2519 events=26/37 payload=0.595 timing=0.726 delayUs=13075.222 maeSamples=18.120421

The channel correlations were negative for both mic inputs, so the streaming packet scorer now treats polarity as a nuisance variable in the packet-local path. That is a real-world acoustic invariant, not a broad decoder patch: mic, preamp, wiring, and room reflections can flip the sign of a narrow waveform template without changing the packet identity.

Obsolete Schedule-Entangled Receipt

The best current receipt is:

dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --bioacoustic-contestants --seconds 0.75 --song canary-packet-trill --decoder packet-razor-index --max-songs 1 --max-decoders 1 --max-degradations 5 --output artifacts/bioacoustic-contestants

Receipt:

artifacts/bioacoustic-contestants/contestants-20260524-185328/contestant-summary.json

Best result before independent payload classification:

song=canary-packet-trill
decoder=packet-razor-index
degradation=warp-light-blur
language_score=396.683
realtime=6.7x
timing=1.000
frequency=0.922
payload_bitrate=64.0 bps
payload_accuracy=1.000
correct=5/5

This receipt is still useful for timing/frequency/speed pressure, but not as a bitrate claim. It assumed payload recovery from event identity instead of decoding the payload alphabet from the observed word.

Worst result in that five-degradation panel:

degradation=warp-light
language_score=199.244
realtime=5.1x
timing=1.000
frequency=0.724
payload_bitrate=54.1 bps
correct=5/5

The packet-razor-index profile is the current speed ceiling: 512-point FFT, 16 mel bins, 6 cepstral coefficients, one projection table, 8-bit hashes, and a tight proposal budget. It only became viable after adding the missing warp-light template augmentation; without that, plain warp-light collapsed to 0.720.

Boundary tests:

  • packet-sprint-index is steadier than razor but slower. Best canary result: 255.908, realtime 4.5x, timing 1.000, frequency 0.976.
  • packet-needle-index buys more speed and loses some frequency fidelity. Best canary result: 258.298, realtime 5.2x, frequency 0.845.
  • finch-burst-packet proves the current lower bound is too short: its best razor result was 280.146, but blur-light collapsed to 4.358; with sprint it recovered blur but topped out at 218.368.

Broad panel sanity check:

dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --bioacoustic-contestants --seconds 0.75 --max-songs 8 --max-decoders 5 --max-degradations 2 --output artifacts/bioacoustic-contestants

Receipt:

artifacts/bioacoustic-contestants/contestants-20260524-185708/contestant-summary.json

This clean/blur-only panel still picked canary-packet-trill + packet-razor-index as best, with language_score=264.923, realtime=6.0x, timing=1.000, frequency=0.944, and payload_bitrate=46.7 bps on clean roundtrip. The score is lower than the targeted five-degradation best because the broad panel did not include the warp-blur cases where canary/razor hit 396.683.

Lessons

  • The old decoder recognized words but emitted coarse proposal offsets. Timing residuals were thousands of samples even when identity was correct.
  • Adding waveform refinement around the winning word fixed some anchor cases but was too slow when done as a one-sample brute scan.
  • The current refiner is a two-stage phrase search: coarse scan, then fine snap near the best basin. It is still managed and expensive, but it turns feature identity into a real timing anchor for some contestants.
  • robin-warble currently gives the best anchor accuracy under light blur.
  • current-birdcall and redpoll-trill still identify words but often anchor inside the motif instead of at the canonical motif start.
  • The next scoring improvement should expose per-formant/frequency residuals directly instead of using word identity as a proxy.
  • The 96 ms six-syllable canary packet is the best current tradeoff. Pushing the motif down near 78 ms increases nominal bitrate but loses too much degradation resistance.
  • The global delay hypothesis is not optional. A single smeared anchor can move a least-squares fit by hundreds of samples if it is allowed to claim clock authority directly.
  • AquaSynth’s useful pressure is the separation of authoring intent from realtime DSP: these packet songs should become precompiled/formally described emission surfaces, with runtime controls exposed as stable parameters rather than regenerated inside the audio callback.

Real Bird References

Xeno-canto API v3 now requires an API key, so this pass pulled Wikimedia Commons mirrors of Xeno-canto recordings instead. Local reference files are under artifacts/birdsong-reference/; they are research inputs and should not be committed as repo source.

Downloaded references:

  • Wattled Guan, XC250442, Niels Krabbe, CC BY-SA 3.0, Wikimedia Commons / Xeno-canto mirror.
  • Lesser Redpoll, XC482789, Pascal Christe, CC BY-SA 4.0, Wikimedia Commons / Xeno-canto mirror.
  • Brown Thornbill, XC446870, Wikimedia Commons / Xeno-canto mirror fetch was attempted but the downloaded file was only 1990 bytes and should be treated as invalid until re-fetched.

FFmpeg is now installed through winget as Gyan.FFmpeg 8.1.1. The current shell used the installed binary directly from:

C:\Users\Meta\AppData\Local\Microsoft\WinGet\Packages\Gyan.FFmpeg_Microsoft.Winget.Source_8wekyb3d8bbwe\ffmpeg-8.1.1-full_build\bin\ffmpeg.exe

Decoded WAVs remain ignored scratch artifacts. Committed receipts live under artifacts/spectra/real-birds-ffmpeg/:

  • lesser-redpoll/Acanthis_cabaret_-_Lesser_Redpoll_XC482789-log-mel.png
  • lesser-redpoll/Acanthis_cabaret_-_Lesser_Redpoll_XC482789-log-mel-flux.png
  • lesser-redpoll/Acanthis_cabaret_-_Lesser_Redpoll_XC482789-cepstrum.png
  • lesser-redpoll/Acanthis_cabaret_-_Lesser_Redpoll_XC482789-mel-cepstral-features.csv
  • wattled-guan/Aburria_aburri_-_Wattled_Guan_XC250442-log-mel.png
  • wattled-guan/Aburria_aburri_-_Wattled_Guan_XC250442-log-mel-flux.png
  • wattled-guan/Aburria_aburri_-_Wattled_Guan_XC250442-cepstrum.png
  • wattled-guan/Aburria_aburri_-_Wattled_Guan_XC250442-mel-cepstral-features.csv

The Brown Thornbill file remains invalid. FFmpeg reports:

Failed to find two consecutive MPEG audio frames
Invalid data found when processing input

The mel-cepstral receipt tool chooses the loudest active window when no start time is supplied, then writes log-mel, cepstrum, and per-frame feature CSV. The feature CSV now includes positive log-mel spectral flux: frame-to-frame new-energy motion in mel space. This is a better timing-anchor surface than the plain cepstrum because sustained bands recede and attacks/glides light up.

The first real-bird readings:

Lesser Redpoll active window: start=2.000s, duration=4.000s
centroid median=2440 Hz, p90=3041 Hz, max=3981 Hz
bandwidth median=2427 Hz, p90=2683 Hz
peak-energy frame near 1.732s, centroid=3224 Hz, bandwidth=1768 Hz
log-mel flux median=0.0870, p90=0.1166, p99=0.2079, max=0.3437
top flux frames around 1.716-1.720s and 1.478-1.482s
 
Wattled Guan active window: start=137.000s, duration=4.000s
centroid median=2217 Hz, p90=4229 Hz, max=4445 Hz
bandwidth median=1072 Hz, p90=1854 Hz
peak-energy frame near 1.722s, centroid=2021 Hz, bandwidth=766 Hz
log-mel flux median=0.0749, p90=0.1034, p99=0.1818, max=0.2370
top flux frames around 0.034-0.036s and the 3.248-3.744s high-band shifts

Read against the current canary-packet-trill, the lesson is blunt:

  • The Redpoll reference is not a neat de Bruijn-ish token stream. It has dense chatter, short vertical attacks, noisy broadband texture, and repeated contour fragments that stay recognizable even when individual peaks blur.
  • The Guan reference is the opposite useful witness: a slow, broad, curved low-mid phrase with stable harmonic/cepstral structure. It is poor as a high bitrate packet model but excellent for response/group-delay probing.
  • Our canary packet currently anchors time well because its syllables are deliberately sharp, but it is still too ladder-like. Real birds are carrying identity through redundant contour, roughness, onset statistics, and cepstral envelope motion at the same time. That is the next scoring surface.
  • The log-mel flux receipts are the clearest current clue. They expose repeated curved strokes in Redpoll and slower phrase-boundary/harmonic-motion anchors in Guan. The next decoder should treat flux ridges as anchor proposals, then classify the surrounding log-mel/cepstral word shape instead of asking the cepstrum to do all the work alone.

Next Cut

  1. Extract more real mel-cepstral contours: syllable onset density, frequency glide slope, formant spacing, rhythm variance, and motif repetition.
  2. Add per-formant frequency residual to the contestant score.
  3. Promote the two-stage refiner into a streaming state model: predict next expected word time from the current clock and only search a narrow basin.
  4. Teach the generator to combine Redpoll-style onset/chatter density with Guan-style smooth cepstral envelope probes.
  5. Run the full contestant panel and keep mutating the top two shapes until score improves rather than spreading attention across ornamental variants.