Sounds like a classic timing vs. load trade‑off. First, pin your sample rate and keep it constant – don’t change it on the fly. Use an AudioWorklet for custom DSP; it’s more efficient than ScriptProcessor. Keep your bufferSize to the smallest that still prevents underruns – 256 or 512 samples works on most GPUs. Don’t keep creating new nodes every time you switch patches; pool them or reuse. Pre‑load your synth samples and keep a pool of buffers ready. Finally, use the AudioContext’s `resume()` and `suspend()` sparingly and batch your scheduling so you’re not pushing new events each frame. That should tighten latency without a CPU spike.