OPM.js v1.13.0

Package version 1.13.0, a GitHub release (not on npm). No voice-format (still v7) or worklet protocol changes. The internal render-Worker start message gains a dither field. API addition: opt-in WAV dither (below). Rendering only changes by rounding: on a mixed 35-preset render with edits and releases, under 0.1% of Float32 output samples differ from v1.12.0, by at most 3e-8.

WAV export dither

• encodeWav, createWavEncoder and renderSequenceInWorker accept dither: true (default false).
  • It adds deterministic TPDF dither before rounding to PCM16/PCM24: the difference of two xorshift32 uniforms, ±1 LSB triangular, with a fixed seed per encoder.
  • Quiet passages, fades and reverb tails keep sub-LSB detail as a steady noise floor instead of truncation distortion, and digital silence becomes that noise.
  • In tests, a 0.4 LSB sine encodes to silence without dither and to an unbiased 0.4 LSB average with it.
  • Full scale is clamped, never wrapped.
  • Streamed chunks continue one sequence, so they are byte-identical to full-buffer and Worker output, and a rejected chunk does not advance it.
  • It is rejected for Float32. Undithered output is unchanged.

Faster rendering

• Operator phases, increments and phase modulation are kept in half-turns, so the new deterministic sinPi (sin(π·u), same degree-11 polynomial with π folded into its coefficients, |error| ≤ 1.4e-11) reduces phases with two additions and no multiply by π. fastSin(x) is sinPi(x/π). Both sinPi and fastExp2 evaluate their polynomials in Estrin form, roughly halving the dependent multiply-add chain on feedback, LFO and pitch paths.
• The sine operator kernels use module-local constants and a local copy of the polynomial: optimized code otherwise re-checks imported bindings on every operator evaluation.
• Envelope segments carry their geometric gain from frame to frame (re-anchored to the exact dB rule every 64 frames) and skip per-operator reclassification while a frame is known to stay inside the same attack, decay or release segment.
• A voice whose operator gains are all zero for a frame (for example a sustain-zero one-shot that was never released) only advances its phases, noise clocks and feedback history instead of evaluating oscillators; its output is still exactly zero and a later ADSR edit resumes bit-identically.
• Voice LFO settings and per-operator targets are snapshotted at note-on into fixed numeric fields, so mixed banks with differently shaped lfo objects no longer make per-frame reads polymorphic.
• The decimator takes one output frame of oversampled inputs per call; the generic waveform path skips operators with zero static level (their phase still advances and is still checked). Each note caches its all-sine frame function and pitch envelope, and the per-frame PM step and AM folding are unrolled.
• Voices that combine sine with other periodic waveforms, but no noise, now take per-algorithm frames when their operator levels are static, like all-sine voices. Each operator has its own waveform call site, so the JIT inlines only the shapes a voice uses. All-sine voices keep their own frames. Measured against the previous step: 1.2–1.5× for 8 voices with two non-sine operators, across all seven shapes. Output is bit-identical to the generic path; a test now checks this for all eight algorithms, using a silent noise operator to force the generic loop. The old identity test compared the fast path against itself.
• Held and centered voices now render without per-frame heap allocation. V8 boxes a double that crosses a call it did not inline, and a merge with such a result boxes every sample that flows through it, even while that branch never runs. Before this change, eight held voices at 48 kHz produced about 7.5 MB/s of short-lived numbers, a 16-voice bank cycling through every preset about 17 MB/s, and eight voices with a square operator about 92 MB/s. All of these meant frequent young-generation GCs on the audio thread.
  • The decimator stores each output in its state, and render reads it from there.
  • renderVoice returns a status flag.
  • The DC blocker, steal fades and pitch-envelope frames no longer pass or merge per-frame doubles across calls.
  • The generic waveform path scales sine and non-sine oscillators in separate branches, and the waveform dispatcher types every case as a number. Non-sine waveform voices measured up to about 10% faster.
  • Exact envelope evaluation (64-frame re-anchoring, segment entry and boundary substeps) goes through storeGain, which takes no doubles and returns none. 32 voices retriggering every block previously allocated about 3.4 MB/s there.
  • Waveform voices can still allocate when the JIT's inlining budget leaves a shape or the LFO's fastExp2 out of line: eight voices with one square operator and vibrato measured about 48 MB/s in their frames.
  • Output is bit-identical.
• Measured on Apple M5 / Node 26 against v1.12.0 (best of seven interleaved CPU-time runs, 48 kHz, 1 s per case):
  • standard 1 voice 1.63×; 8 held voices 1.62×; 8 steady voices 1.69×; attack/decay 1.85×; 32 voices 1.54×.
  • high 8 voices 1.56×; eco 8 voices 1.52×.
  • Pitch envelope 1.52×; per-operator LFO targets 1.78×.
  • Square operator 2.08×, and other non-sine shapes 1.6–2.0×.
  • A 16-voice mixed all-sine bank 2.04×.
  • A 16-voice run cycling through all 35 bundled presets with releases measured about 1.6×, and retriggered noise hi-hats about 1.65×.
  • These are single-machine Node results, not browser AudioWorklet or device measurements.