Tracklist
Bulwark Bits: “You Are VILE” (Sam Stein and Tim Miller)
* I’d like to remind everyone of my world-leading arrangement of Tim’s rendition of O Canada
J.J. Fad - Supersonic
He’s the freshest DJ from coast to coast. Now, he’s the freshest DJ and can cut to boast. He may be a little light. He may be OK. ‘Cause we butt the soul, ‘cause when we be a little light, it sounds this way.
Now you party people know what Super Stonic Styneenns.
We didn’t come correct ‘cause J.J. Jazz just too poopy. Now baby, don’t you know that our rhymes are too bayonza?
Listen to her. Don’t be supersonic.
Boogie Down Productions - South Bronx
Doobie Brothers - What A Fool Believes
Pat Travers Band - Born Under A Bad Sign
Steve Miller Band - The Joker (Pompetus of Love)
JVC F.O.R.C.E. - Strong Island
Justified by Virtue of Creativity — For Obvious Reasons Concerning Entertainment (en.wikipedia.org)
Stone Roses - Fool’s Gold
Cardi G - Lovefool
Break Machine - Street Dance
* This was put together by the Village People’s producers. It charted higher in Europe than in the U.S.
Bobby Brown - My Prerogative
* Wrote this after getting flak for leaving New Edition
Teej - Slick Rick (Original Mix)
Teej - It’s A Feeling (Original Mix)
Korni Grupa - Moja generacija
* Yugoslavia’s 1974 Eurovision entry, the year ABBA won with “Waterloo.” Korni Grupa were a prog-rock band. Strange for Eurovision but good tune.
The Furby Fans - Furbies Op De Radio
* The most incredible thing. I uploaded the whole album.
Temptations - Ball of Confusion
* This song freaked me the f**k out when I was little.
The Furby Fans - I Love My Furby (English Version)
Samira Bensaid - Bitaqat Khub
* Morocco’s only Eurovision entry (1980). The country never returned, making the song a Eurovision trivia staple.
Nilüfer Yanya - Like I Say Really Fast
Teej, J.D. REID - Too Much TV (Original mix)
Jeep Beat Collective - B-Boys Breakdance Forever (Extended remix)
* From the album “Technics Chainsaw Massacre”
Bambaata Zulu Nation Soul Sonic Force - Zulu Nation Throwdown
Jody Watley - Looking For A New Love
* She said “Hasta la vista, baby” four years before Terminator 2.
Biz Markie - Vapors
* Biz’s morality play about clout and fair-weather friends
Pete Rock - Collector’s Item (Instrumental)
BONUS
For Luke Beasley
Advances in autoaudiomnemonic recording of “imagined” sound
We describe a pipeline for real-time reconstruction of imagined sound and its projection into an external auditory scene—Audio-Mnemonic Respatialization (AMR). The approach integrates neural signal acquisition, representation learning, and spatial rendering. We treat “imagined sound” as a latent trajectory in an auditory-mnemonic state space estimated from neural correlates of auditory imagery and recall. The resulting system decodes symbolic and sub-symbolic descriptors (timbral, phonetic, rhythmic, and spatial) and renders them through a parametric synthesizer and scene simulator. We outline acquisition strategies (EEG/MEG/ECoG/ear-EEG), decoding architectures (encoding/decoding models, state-space inference), tokenization (neural codecs), rendering (neural vocoders + HRTF/BRIR convolution), calibration protocols, latency budgets, and evaluation metrics. We close with risk analysis and governance notes.
1. Motivation and Framing
Auditory imagery—internally generated percepts with measurable correlates in auditory cortex, STG/STS, motor planning areas, and hippocampal networks—exhibits structure homologous to perception and performance. Empirically, imagery recruits tonotopic maps, rhythm-tracking networks, and language circuits with attenuated amplitude and altered phase relationships. The present aim is not a general brain-to-waveform translation, but a practical tool that:
* estimates a compact description of a user’s intended sound in real time;
* renders that description to audio;
* situates the audio in a scene consistent with the remembered or intended spatial context (the “mnemonic” component).
2. System Overview
The AMR system comprises five subsystems:
* Acquisition: continuous neural and kinematic streams.
* Preprocessing: artifact rejection, alignment, and feature extraction.
* Decoding: mapping neural features to audio tokens and scene parameters.
* Synthesis: token-conditioned neural vocoder + parametric instruments.
* Respatialization: room/position modeling (HRTF/BRIR/propagation) with head tracking.
A control plane coordinates calibration, confidence estimation, and user feedback.
3. Acquisition
3.1 Modalities
* Noninvasive: high-density EEG (64–256 ch), ear-EEG for practicality, and optional MEG in lab settings.
* Semi-invasive/invasive (research): ECoG/SEEG yield higher SNR and bandwidth but are not assumed for consumer deployment.
* Ancillary: IMU for head pose; eye tracking for attentional state; optional EMG (orofacial) to exploit covert articulation during imagined speech/song.
3.2 Sampling and Synchronization
* EEG/ear-EEG at 1–2 kHz, anti-aliasing per montage.
* Hardware timestamping and PTP-like clocking to bound inter-stream jitter <2 ms.
* Sliding windows of 200–500 ms with 50% overlap balance latency and stability.
4. Preprocessing
* Artifact handling: band-limited regression for EOG/EMG; ICA/AMICA where feasible; spectral gating for mains interference.
* Feature stacks:
* Rhythmic: delta–beta phase, inter-areal phase-locking value (PLV), and amplitude envelope correlations with perceptual rhythm templates.
* Tonotopic: spatial filters aligned to estimated frequency gradients in auditory cortex (data-driven beamformers).
* Linguistic/prosodic: Riemannian geometry features of covariance matrices on SPD manifolds; CCA with articulatory priors.
* Self-normalization: subject/session drift handled via adaptive whitening and online affine recalibration using a short moving baseline.
5. Decoding Architecture
5.1 Learning Strategy
A two-stage model:
* Stage A: Encoding models (perception mode). While the user listens to a stimulus battery, we learn mappings from audio features → neural responses (ridge/ARD regression or shallow convs).
* Stage B: Inversion/decoding (imagery mode). We invert the learned models under a sparsity/low-rank prior to recover token probabilities and continuous controls from neural data. A controlled state-space model (SSM) with a Kalman-like update or neural SSM (S4/SSM-Transformer) yields streaming stability.
5.2 Target Representation
We avoid direct waveform regression. Instead we decode:
* Neural-codec tokens (e.g., residual vector-quantized units): compact, perceptually aligned, streamable.
* Symbolic descriptors: pitch class/contour, tempo, meter, onset patterns.
* Phonetic units for imagined speech/song (phones or discrete self-supervised units).
* Scene parameters: azimuth, elevation, distance, source width; room size, RT60, early/late energy ratio; absorption profile.
* Confidence per unit and per channel.
Posterior over tokens is tempered to manage hallucination; a small beam (B=4–8) preserves diversity without destabilizing latency.
6. Synthesis
6.1 Neural Vocoder
Tokens → waveform via a lightweight, streaming neural vocoder (multi-band, multi-period discriminator).
* Frame size: 20–40 ms, stride 10 ms.
* Runtime on laptop-class GPU/Neural-NPU ≤ 10 ms/frame; CPU fallback ~25 ms/frame at 22.05 kHz.
6.2 Parametric Layer
Symbolic descriptors drive parallel generators:
* Monophonic pitch with differentiable oscillators;
* Drum events via sample-select + micro-timing jitter;
* Articulatory TTS for phonetic units when a voiced output is desired.Blending weights are derived from confidence; the mixture regularizes the generative model when the token posterior is diffuse.
7. Respatialization (Mnemonic Component)
7.1 Source Localization from Imagery
Parietal and posterior temporal features carry coarse spatial intent. We regress:
* Azimuth/elevation (spherical),
* Perceived distance (log-scaled),
* Apparent width (coherence proxy).Outputs are smoothed through the SSM to avoid discontinuities.
7.2 Scene Construction
Three tiers, selected by compute budget:
* HRTF-only: individualized HRTF (or K-means cluster HRTF) with binaural panning and late-reverb approximation.
* Parametric BRIR: image-source model with frequency-dependent absorption, yielding early reflections and tail which are conditioned on the decoded mnemonic room features (room size, RT60).
* Wave-based (lab): fast modal solvers for small rooms or neural acoustic fields pre-fit to candidate spaces.
Head-pose modulates binaural rendering in real time. Optionally, decoded “memory tags” (e.g., “cathedral”, “tile bathroom”) select priors over room parameters.
8. Calibration Protocol
* Perception phase (8–12 min). User listens to a stratified stimulus set covering speech, environmental sounds, instrument families, rhythmic grids, and room archetypes. We fit Stage-A encoders.
* Guided imagery (4–6 min). Cued recall of a subset forms paired (neural, token) supervision to warm-start Stage-B decoders.
* Online adaptation. During normal use, confidence-weighted teacher forcing from the synthesizer’s own tokens lets the decoder adapt without catastrophic drift (EMA with low learning rate).
* Periodic mini-recalibration when SNR degrades (detected via Riemannian distance of covariance from baseline).
9. Latency Budget
* Acquisition + preprocessing: 8–15 ms (ear-EEG front-end + short window).
* Decoder update (SSM + small Transformer head): 10–18 ms.
* Vocoder frame: 8–12 ms.
* Spatialization + head-pose: 1–3 ms.End-to-end (steady-state): 40–65 ms, adequate for interactive monitoring.
10. Evaluation
* Content similarity
* FAD (Frechét Audio Distance) between target class exemplars and reconstructions.
* STOI/ESTOI for speech-like imagery; Kendall tau on pitch contours; beat-alignment F-score for rhythm.
* Spatial fidelity
* Localization error (degrees), distance RMSE, interaural parameter deviation (ILD/ITD).
* Subjective
* MUSHRA panels; triadic comparisons for spatial plausibility; intra-subject test-retest.
* Operational
* Dropout rate under motion, re-calibration time, sustained performance over 60–90 minutes.
11. Failure Modes and Mitigations
* Mode collapse (decoder outputs generic tokens).Mitigation: entropy flooring on token posterior; parametric layer blend; scheduled sampling.
* Confusion between heard and imagined (perception leakage).Mitigation: gating classifier distinguishing exogenous vs endogenous patterns; physical earplugs or bone-conduction monitoring.
* Spatial drift (decoded azimuth wanders).Mitigation: head-pose-anchored stabilization; SSM damping; re-anchor via brief “look-at” cue.
* User fatigue / electrode shift.Mitigation: adaptive re-weighting of channels via Riemannian minimum distance to class.
12. Interfacing and Formats
* Token stream: time-stamped, CBOR/MessagePack, <2 kB/s typical.
* Scene parameters: 50–200 Hz float tuples; optional OSC bridge.
* Audio: 22.05–48 kHz PCM; ambi-bus option (ACN/SN3D) for external renderers.
* Security: local inference by default; export requires explicit user consent and embedded provenance metadata (W3C C2PA-style).
13. Applications
* Compositional capture (silent composition, live ideas to tokens).
* Assistive communication (imagined speech to articulate TTS with individualized timbre).
* Forensic recall under controlled protocols (strict governance required).
* Audio scene notes: attach BRIR-like “memory snapshots” to photos/videos for later recall.
14. Ethical and Governance Notes
Imagery decoding is categorically sensitive. We enforce consent-locked operation, on-device processing, deletion by default, and externally verifiable provenance watermarks on all renders. No background or passive decoding. Institutional deployments require IRB-equivalent oversight and adversarial red-team testing for model inversion and membership inference risks.
15. Implementation Sketch (near-term)
* Hardware: 32–64 ch ear-EEG with dry electrodes; Bluetooth LE Audio is avoided in favor of wired/USB to keep jitter low; 6-DoF IMU in the headband.
* Runtime: small-footprint inference on consumer laptop or edge NPU; TorchScript or ONNX; real-time priorities for audio thread; ring buffers with back-pressure.
* Models:
* Stage-A encoders: sparse linear + shallow convs;
* Stage-B decoder: 2-layer SSM-Transformer head over Riemannian features;
* Vocoder: streaming multi-band;
* Spatializer: fast HRTF convolution with pre-tabulated HRIRs + parametric late reverb.
16. Limitations
Noninvasive SNR remains the primary constraint; individual variability in imagery quality is significant; linguistic reconstructions degrade with covert articulation suppression; cross-day generalization requires periodic alignment; causal decoding imposes a tight bias-variance tradeoff; re-creating highly specific timbres without a subject-specific prior remains challenging.
17. Conclusion
AMR redefines “recording” as decoding and rendering of intended acoustic content, with explicit modeling of the remembered spatial context. The design above is engineering-first: use encoders learned during listening to invert imagery in real time; map to a tokenized audio manifold; synthesize through a constrained vocoder; and respatialize via a fast scene model informed by decoded mnemonic cues. The overall system is attainable with present commodity compute, provided careful calibration, conservative priors, and strict governance.
Appendix A: Minimal Operator Workflow (for practitioners)
* Fit encoders with 10-minute listening battery.
* Run guided imagery calibration; verify token entropy and basic spatial control.
* Engage real-time mode; monitor confidence, PLV stability, and latency.
* If token entropy < target or drift > threshold, trigger 60-second mini-recalibration.
* Export token stream + BRIR parameters for archival; audio renders carry provenance.
Appendix B: Candidate Stimulus Sets (abbreviated)
* Speech: phoneme-balanced sentences; prosodic contours.
* Music: monophonic pitch staircases; drum patterns; instrument banks across ADSR regimes.
* Scenes: small room, large hall, outdoors, reflective corridor; azimuth sweeps at three distances.
This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit djic.substack.com/subscribe
