Back to the example index

Preset and conversion audition

Compare the expanded original/repository recipe bank and clean-room DX7 conversion fixtures. Start with low device volume. The master host gain is 0.12 (−18.4 dB); energy-matched mode adds separate per-source attenuation, while dry mode omits that trim. No patch is rewritten. Matching RMS is not perceptual proof.

A/B sources and single-note settings Source A Source B Register Velocity Material Host comparison DSP quality profile Phrase seed (unsigned 32-bit integer)

Single notes use a 0.8-second gate. The seeded phrase repeats six isolated notes with reproducible articulation and velocities; MIDI notes are bounded to 48–84. A/B use common padded slots so release tails never overlap and RMS windows agree. Stop cancels buffers immediately. Dispose closes the owned AudioContext; Play starts a fresh session. This listening surface plays locally rendered dry DSP buffers, not a real-time worklet stress test.

PCM16 WAVs are synthesized locally at 48 kHz with the exact selected material and playback host gains, including release tails and slot padding. Record the seed, A/B sources, register, velocity, material and gain mode to repeat comparisons. No recordings or patches are uploaded. Changing a setting invalidates the download.

Sound report

Choose sources, then measure their low/middle/high register and soft/medium/hard velocity grid.

Reports show raw peak, gate RMS (single notes), common whole-window RMS (phrases), and full-render RMS before playback gain. Applied host trims are shown separately. Unweighted RMS is not LUFS or a perceptual quality verdict. Repeat matched and dry comparisons and note attack/body/decay/brightness preferences yourself; no listening results are fabricated. Chords need additional headroom.

Conversion limits

The DX7 fixtures are original parameter recipes, not third-party patch banks. OPM.js is four-operator synthesis, not a DX7 emulator. A/B compares converted recipes and OPM voices; no six-operator audio reference is rendered. Carrier removal and approximate envelopes/levels/velocity/feedback/LFO transfer functions can be audible; v5 retains fixed-frequency operators in Hz.

Record a human listening finding

Automated reports never become listening results. After auditioning A and B, record only what you heard. Entries stay in this page's memory until you export them; no audio, patch or finding is uploaded. Use an anonymous listener label.

Listener and playback route Anonymous listener label Device / browser / OS Output route, headphones, device volume I actually listened to both A and B at these selected settings.

Preference per criterion Attack Body Brightness Decay Release Control transitions (ratio, feedback, ADSR) Gain comfort Notes (up to 1024 characters) Subjective gain notes for this matched or dry comparison (no assumed volume units)

0 local findings in memory.

Loading voices.