The engine
Mamba-3, per-sample,
Mamba-3, per-sample,
in pure C++17.
A multi-head MIMO state-space model running per-sample in the time domain — no STFT in the model path, no external inference runtime, no GPU dependency at runtime.
Architecture
Specification
Six stacked SSM blocks with RoPE rotary embedding, trapezoidal discretization, FiLM conditioning, and MIMO rank-4 heads.
| Parameter | Value |
|---|---|
| Architecture | Mamba-3 SSM, multi-head MIMO R=4 |
| d_model / d_state | 128 / 32 |
| Layers | 6 (stack-allocated SSMBlocks) |
| Parameters | ~1.66M trainable |
| Processing | Per-sample, time-domain |
| Weights | Flat binary model.bin, mmap, version-checked (V2) |
| Dependencies | C++17 stdlib only |
| Latency | 0 samples reported |
Verification
The test suite is the datasheet.
| Claim | Evidence |
|---|---|
| CLAP conformance | clap-validator in CI |
| Cross-platform build | GitHub Actions matrix — Linux · macOS · Windows |
| Numerical correctness | Bit-exact Python ↔ C++ parity (CTest) |
| RT-safety | Zero heap allocation on audio thread; deferred atomic state reset; lock-free SPSC queue |
| Bypass behavior | Click-free 50 ms equal-gain crossfade |
| Stability | 60 s @ 44.1 kHz long-run, no NaN/Inf |
Roadmap
R=8 spectral engine
The next generation adds an STFT frontend (n_fft=2048), complex-pole SSM, CREPE f0 conditioning, and cross-track attention at ~4.26M parameters. Training pipeline validated on datacenter GPUs; the C++ engine rewrite is in progress.
Latency disclosure: the spectral engine introduces
1024-sample algorithmic latency (n_fft/2). Current shipping engine: 0
samples. Both figures will be published with reproducible measurements.