PoC 13A — Learned Erf-Mask Host Qualification

Question

Can a deterministic learned model built around quantized Erf improve noisy speech while preserving clean speech, before any new Wave device call?

Falsifiers

• training does not improve validation loss;
• model artifacts do not reproduce;
• synthetic median SI-SDR improvement is below 3 dB;
• natural median SI-SDR improvement is below 1 dB;
• natural STOI delta is below −0.02;
• noise attenuation is below 6 dB;
• 477-item clean-speech median absolute RMS delta exceeds 1 dB;
• any invalid, non-finite, or clipped evaluation item;
• any device call.

Model

The host computes a causal 960-point STFT, causal noise estimate, and 481 learned slopes plus 481 biases. Quantized logits use U8 scale 0.03125 and zero point 128. The independent Erf oracle returns U8 scale 0.0078125 and zero point 128.

The final denoiser evaluates a learned Erf candidate and a DeepFilterNet CPU candidate every frame. It selects DFN when retained frame power is at least 0.7943282347 of noisy power; otherwise it selects the learned candidate.

Negative iterations retained

• The first full learned-Erf waveform lost about 5.75 dB on clean speech.
• A weak residual Erf postfilter did not satisfy out-of-distribution synthetic quality.
• The causal DFN-or-Erf selector was the first candidate to satisfy every held-out gate.

Result

status: PASS
items: 1,230
invalid: 0
synthetic median SI-SDR improvement: +3.2597667132 dB
natural median SI-SDR improvement:   +5.2338097012 dB
natural median STOI delta:           +0.0135532935
noise median attenuation:            10.5837640380 dB
clean median absolute RMS delta:      0.0463758547 dB
partition calls:                      740,742
Erf-selected frames:                  450,140
DFN-selected frames:                  290,602
device calls:                         0

Evidence

Artifact	SHA-256
Model manifest	bd7ffc58d253855e3ddc546def711f91812da8cb9ecd88dc658b56aada40ce5e
Weights	ab722366943c9e090f4f97fa5916bb3bf0da0d42bd04a924e1f30b5b256c030b
ONNX	effebd8e351d857fcc284558b7c551869eb191ba0eb4f572bbde6e45f7168539
Evaluation	32d7b224253e2a05f230555db7cfdf4bb6094c81d832d5793851a582a376ee2d
Host qualification	ad286a54ca1c3311cb1a44818c777de2dd62dca1b8d155b715e076f03710693d

Verification

cd waved/tools/onnx-hybrid
uv run pytest -q
uv run ruff check .

Final regression result after PoC 13F: 187 tests passed; Ruff passed.

Interpretation

The learned host architecture and quantization boundary are qualified. This does not prove compilation, NPU execution, NPU residency, or general ONNX support.

Next gate

Compile the one fixed Erf UINT8[481] -> UINT8[481] graph twice without executing it.