Baseline CDSL Seed Design

The model sums measured complex pressure by angle and side after explicit digital biquad filters, gain, polarity, and delay. SPL calibration, physical driver offsets, cabinet diffraction, and distortion under equalized drive remain assumptions. This baseline is fixed by request and was not chosen by the optimizer; the full candidate search is shown only as context.

Baseline is the fixed seed: L26RO4Y below 200 Hz, L22MG 200-800 Hz, GRS PT6816 800-2500 Hz, and ND25FW4 above 2500 Hz. Comparison factors use broad psychoacoustic frequency weighting and are separate from hard filter-count constraints. Weighted factor wins: chosen 45%, baseline 51%, insufficient/context 4%.

Hard Filters

Biquad counts and sparse-EQ limits are acceptance criteria only; they are not scored as acoustic advantages once they pass.

Open side-by-side comparison page

The search is not just the preliminary stack. It tries L10NEO, both ScanSpeak 10F variants, GRS, MU10, and ND25 across multiple LR2/LR4 crossover grids. Lower score is better; the score favors cosine/dipole-like front and rear polars, strong side nulls, adjacent-driver pattern match near crossovers, measured-frequency validity, larger psychoacoustically weighted 2-10 kHz SPL43.6 - SPL45.8 separation, known in-band THD traces, and the 15-biquad/channel cap as a hard filter. This baseline is fixed by request and generated with the same gain, delay, mixed-order LR, and sparse flat-EQ process as the chosen design.

Download ranked candidate search JSON (top 0 of 54504)

These rows are measured-driver diagnostics from Juan's baffleless HDF5 data before synthetic crossover summation. They explain the upper-band tradeoff: GRS is closest to a clean dipole shape, L10NEO and the 10F ScanSpeaks provide more configured x-c angle separation, and the dual-ScanSpeak split is kept as an experimental finalist because it adds a crossover between two near-identical radiators.

Distortion and SPL evidence comes from measurements/juan/UMs desnudos/contexto.txt and its associated screenshots:

• Juan's screenshot notes rank 8424 distortion/SPL best, 8414 close behind, and MU10 worst among the 8424/8414/MU10 upper-mid comparison.
• Raw REW .mdat THD extraction adds L10NEO to the comparison and does not support describing L10NEO as worse-distortion than the ScanSpeak pair.
• The same notes flag 8424 rear-side directivity and high-angle order as weaker than 8414, especially above 2 kHz.
• GRS is treated as the measured dipole/directivity-order reference, but it does not provide the largest configured x-c angle separation above 2 kHz.
• L10NEO remains a high-separation alternate; it is not selected in the balanced primary due to the composite directivity/crossover score, not because of worse raw THD evidence.
• Driver contribution peaks are not automatically flattened when they are helping the complex 0-degree sum; the search now separately penalizes narrow 8-12 kHz normalized-contour ridges and steep side/rear polar transitions.

Download measured-driver directivity audit CSV

THD and measurement level are extracted directly from the front 0-degree REW .mdat files, not from screenshots. The table shows each driver's fundamental SPL during the distortion sweep plus REW's stored THD trace. This avoids comparing a low-level distortion sweep against a high-level one.

• Raw REW .mdat THD was extracted from each driver's front 0-degree measurement; harmData[0] is treated as REW's THD trace and harmData[1] as the fundamental SPL trace.
• In the 2-7 kHz THD band, L10NEO's median fundamental SPL is 86.6 dB (+3.0 dB vs SS10F8414G10 single, +1.9 dB vs SS10F8424G00 SN086, +3.7 dB vs MU10RB-SL A), so it was not measured at a lower level than the ScanSpeak references.
• L10NEO 2-7 kHz median THD is 0.157% (-56.1 dB), which is not worse than the available 8414/8424 raw REW THD traces.
• These are not perfectly level-matched sweeps; distortion rankings should be treated as measured-at-level evidence, not normalized maximum-SPL evidence.

L10NEO vs ScanSpeak

• Raw REW evidence does not favor ScanSpeak on THD: L10NEO is 0.157% median THD at 86.6 dB SPL in 2-7 kHz, while the best available ScanSpeak trace is SS10F8424G00 SN086 at 0.177% and 84.7 dB SPL.
• The baseline is a fixed seed, not an optimizer-selected rejection of L10NEO; use it only as the requested reference stack.
• Conclusion: L10NEO must not be rejected on distortion/SPL evidence; any non-L10 selection has to be justified by directivity/crossover integration, not by the available raw THD traces.

Download raw REW THD / SPL audit CSV

Effective THD is estimated from the filtered front 0-degree driver contributions. Only drivers with usable REW THD traces contribute distortion data, and harmonic pressures are combined incoherently. The coverage column shows what fraction of the weighted fundamental contribution has THD evidence in each band.

Download effective system THD CSV

Cascaded Filter Topology

Architecture: cascaded mixed-order Linkwitz-Riley split tree with shared high-pass carryover (LR4 / LR4 / LR4). The higher branch of each split carries the previous high-pass stages into the next split, so later drivers include the first-stage and intermediate high-pass filters before their own branch filters.

Input
  +-- LR4 HP 70 Hz (2 biquads, global boundary)
    +-- LR4 split @ 200 Hz
        +-- LR4 LP 200 Hz (2 biquads) -> L26RO4Y
        +-- LR4 HP 200 Hz (2 biquads) -> next split
      +-- LR4 split @ 800 Hz
          +-- LR4 LP 800 Hz (2 biquads) -> L22MG (nude)
          +-- LR4 HP 800 Hz (2 biquads) -> next split
        +-- LR4 split @ 2500 Hz
            +-- LR4 LP 2500 Hz (2 biquads) -> GRS PT6816
            +-- LR4 HP 2500 Hz (2 biquads) -> ND25FW4 (nude 18mm)

• Shared-tree count assumes the DSP can route a high-pass bus into the next split stage.
• Standalone-channel count is what is exported per driver when each output channel must contain all inherited upstream filters.
• LR2 splits use one Q=0.5 biquad per edge and invert the next/downstream branch; LR4 splits use two Q=0.7071 biquads per edge without a required polarity inversion.
• Flat-EQ is capped after crossover filters so every exported driver channel stays at or below 15 total biquads.
• Gain, delay, and polarity are not counted as biquads.

Crossovers are mixed LR2/LR4 cascades: LR2 uses one Q=0.5 biquad per edge and inverts the downstream branch, while LR4 uses two Q=0.7071 biquads per edge. The flat-EQ filters are a sparse summed-response least-squares fit assigned per driver. They use 12 active correction filters out of 15 candidates; the unconstrained full fit would keep about 14 filters above the pruning threshold. The hard export limit is 15 biquads per driver, including cascaded crossover filters before flat-EQ. This is still a simulated equalization pass, not a final measured-room/prototype EQ.

• Filter JSON
• CamillaDSP-style YAML
• Hypex-style DSP listing
• Design manifest

Target flatness is evaluated on the synthetic front 0-degree sum. The raw gated response still contains narrow features that should not be over-corrected from this preliminary phase model. The more meaningful target here is the smoothed trend before prototype measurement. Primary flatness target met under the 15-biquad-per-driver cap: no.

Optional High-Frequency PEQ

PEQ filters can boost or cut. After the front-sum flatness fit, the generator tries a small set of +/- PEQ candidates around 7.5-12.5 kHz and accepts them only when the 8-12 kHz polar-transition penalty improves without materially degrading front-sum flatness. Penalty before/after: 1.914 -> 1.914.

• L26RO4Y source is a 25 cm deep / 32 cm diameter cylindrical-baffle measurement, not a nude baffleless capture.
• Processed phase was reconstructed from separately loaded measurements after per-measurement impulse peak alignment; synthetic complex summation is therefore preliminary.
• Rear polarity convention is inherited from the measurement files and existing 0-180 mapping; it has not been independently validated with raw impulse polarity.
• Absolute SPL and maximum-SPL claims remain limited by the available local evidence; raw THD is extracted for available 0-degree REW files but not normalized to matched drive level.
• DI and beamwidth figures match the existing repo convention and use front-horizontal data only, not a full 3D power response.
• Vertical-plane beaming and pseudo-baffle diffraction from the final stacked mounting geometry are not modeled.
• The 10 kHz/top-octave behavior is treated as suspect validation territory; search and comparison weights emphasize 2-10 kHz and do not let a narrow top-octave feature decide the design.

The x-c metric is now configurable and uses SPL43.6 - SPL45.8. For a cosine dipole, these angles imply 0.33 dB ideal separation, so the optimizer no longer uses the old fixed 30/60-angle threshold. The speaker aim is modeled as tilted 23.5 deg toward the listener, not perpendicular to the ear line.

The suspended-driver data is most reliable for choosing drivers and horizontal crossover behavior. It is not a complete model of the final physical stack. Vertically aligned drivers can create vertical lobing, and a narrow support spine or neighboring driver motors can behave like a thin diffracting baffle.

• The synthetic HDF5 is a horizontal 0-180 degree sum of separately suspended driver measurements; it does not model vertical-plane lobing from center-to-center spacing.
• Non-GRS cone/dome drivers are assumed approximately axisymmetric when reasoning about vertical behavior, but the GRS planar is not axisymmetric and needs measured or modeled vertical data.
• A first-order vertical lobing model is feasible if driver coordinates and acoustic-center offsets are supplied, but it would only cover source interference, not pseudo-baffle diffraction.
• The pseudo-baffle/diffraction from a vertical mounting spine, neighboring motor structures, wiring, and driver frames is not identifiable from the current individual-driver HDF5 data; it needs CAD/BEM or assembly measurements.
• Recommendation: use this synthetic model for driver/filter shortlisting, then measure the actual stacked fixture horizontally and vertically before treating the crossover as final.

Upper-mid side-feature diagnostic: in the 1.6-3.2 kHz region, front 90-degree response reaches -47.1 dB relative to front 0 degrees at 2695 Hz. This is a side null, not a 90-degree SPL peak. Filling it would reduce the dipole null and the intended CDSL separation.

Metric definition: SPL43.6 - SPL45.8. A cosine dipole gives 0.33 dB with the current angles; higher values above 2 kHz indicate more separation between the ipsilateral and contralateral angles for CDSL/crosstalk-cancellation use.

Download metrics CSV

• CDSL design explorer
• Filter transfer dashboard
• Directivity dashboard
• CDSL vs LX521 dashboard
• Polar explorer

Generated: 2026-05-21T21:53:32+00:00

Inputs: output/data/polar_data_juan_baffleless.h5, output/data/polar_data_lx521_system.h5

Synthetic HDF5: output/data/polar_data_juan_cdsl_baseline_synthetic.h5

Smoothing: None; gate: 0.5 ms / 3.0 ms.

This is a synthetic, equalized pressure sum. It is not a replacement for a measured prototype with final driver spacing and baffle geometry.