Recovering a Working Processing Contract for an Undocumented Streaming Speech-Enhancement Model and eploying It as a Full-Band Real-Time Audio Plugin

Speech-enhancement networks are often passed on as bare inference graphs, without the sample rate, filter bank, framing, tensor layout or mask semantics they were trained with. We derived a working processing contract for an undocumented streaming ONNX model of unknown training data. Inspection of the graph’s nodes and constants determined its tensor layout ([subband, bin, re/im]), PCEN front end, dual-path frequency/time GRU U-Net and mask algebra: direct, reverberant and noise masks, built by a law-of-cosines construction, that sum to 1 + 0j. Controlled experiments selected the remaining conventions: operation at 16kHz on two 8kHz pseudo-QMF subbands, a 256-point square-root-Hann STFT, unnormalised spectra and a hop of 64 subband samples. The hop is not identified: a hop of 128 subband samples also denoised, and moving to 64 (8ms, the hop of the related TRU-Net) raised PESQ on VoiceBank+DEMAND from 2.428 to 2.567 at the same latency and 1.8 times the CPU time. We built minoise, a real-time plugin, tested at host rates of 16, 44.1 and 48kHz, that runs the model through a 7.25kHz complementary linear-phase crossover, integer-sample latency alignment, a high-band gain tracker and a two-dial voice/ambience mix whose default output is bit-identical to the delayed input. On the VoiceBank+DEMAND test set minoise reached a wideband PESQ of 2.567 and an SI-SDR of 17.32dB at a reported latency of 37.9ms and a single-thread real-time factor of 0.084 on mono input. DeepFilterNet3 scored higher on every metric (3.162 and 18.96dB) at a latency of 40ms. Compared with RNNoise (2.450 and 13.26dB, 30ms), minoise was higher in PESQ (paired mean difference 0.117, 95% CI [0.093, 0.141]), CSIG, CBAK, COVL, SSNR, STOI and SI-SDR, and lower on DNSMOS BAK, OVRL and P.808, which for these short clips are computed on looped audio. Layout and mask-application conventions that the graph rules out lowered PESQ by 0.083 to 0.727; swapping real and imaginary parts cost only 0.083 PESQ while making the output orthogonal to the reference. Metric code, system outputs and per-utterance scores are released.