Tymbal: Neural Synthesizer on a Single Chip

Tymbal is a polyphonic real-time synthesizer on a single NUCLEO-N657X0-Q board (STM32N657: Cortex-M55 at 800 MHz + Neural-ART NPU), controlled via MIDI keyboard, with I2S-DAC output. The audio is computed in 4 ms blocks; at 800 MHz that's 3.2 million cycles per hop for everything: voices, harmonic synthesis, neural network, filter bank, detune, limiter, with zero underruns. The architecture uses a PQMF four-subband domain (12 kHz each) to reduce cost and give the network a compact input. The harmonic skeleton is rendered directly into subbands using complex filter responses, with an anti-chiptune layer (auto-unison, spectral bloom, micro-drift). A causal TCN with 12 layers, dilations 1..32, C=88 channels runs on the NPU as a residual refiner, quantized to int8 via PTQ. The key insight was that a network trained naively on residual prediction loses to a linear FIR of 520 coefficients; the fix was to freeze the linear part as a 'zero layer' (pred = FIR(x) + net(x)), so the network learns only the nonlinear residual. Final eval: FIR alone gives +11.25 dB suppression, FIR+net fp32 +17.16, int8 +16.89, with network contributing +5.64 dB. Extensive bring-up findings include ECC random content after reset, disabled AXISRAM banks, NOR-flash path issues, NPU clock divider incorrect (400 instead of 800 MHz), and a false assert. Budget optimization started at 4.18M cycles/hop, finally achieving 2.47M worst-case chord, 0 underrun. Techniques included a measurement loop (QEMU vs board cycles), NPU epoch profiling, pumping (non-blocking NPU calls interleaved with rendering), moving hot tables to DTCM, and MVE-based memory copy for state ring.