Skip to content
Wave Open Firmware

PoC 13E — NPU-Derived Noise-Cancelled Waveform

Question

Section titled “Question”

Can every learned partition frame in a real 12-second fixture execute on the Wave NPU and produce a measurably enhanced waveform?

Fixture transport

Section titled “Fixture transport”

The host captured 1,201 exact UINT8[481] inputs: 1,200 audio hops plus one offline flush hop. Input stream:

bytes: 577,697
SHA-256: a409686c2482b576288b6f46ecf77d3b99b019c9c436e0f042cb1ac39dacc3e9
FNV-1a64: 0xe2c6f1eda5f3eaea

The device performed one compile and 1,201 sequential graph runs, then published one fixed read-only output pack.

Device result

Section titled “Device result”
status: PASS
execution attempts: 1
frames: 1,201 / 1,201
compared tensor bytes: 577,681
mismatched tensor bytes: 0
maximum U8 error: 0
mean graph run: 91.0167 us
maximum graph run: 1,682 us
10 ms budget fraction: 0.0091017
canaries: PASS
health differences: none
kernel alerts: none

Output stream SHA-256: a3df6e36e1ea98bcaab7f03cc3d8ae8797a05aafd4479cab3e8071447af9cd3d.

Waveform reconstruction

Section titled “Waveform reconstruction”

Replay verifies that every live feature vector equals the corresponding sealed device input before returning the pulled NPU result. No fallback was allowed in the passing reconstruction.

NPU partition successes: 1,201
fallback frames: 0
Erf candidate selected: 803
DFN candidate selected: 398
compared waveform samples: 576,000
mismatched waveform samples: 0
maximum floating error: 0

The NPU-derived and CPU-reference f32 streams share SHA-256 3007e19195c2b1943ac2e5506d163d937e1f8f011d94404dd8435517ca412e40. The deterministic WAV SHA-256 is 30e015ed85faacd5198c70a3021fd59557103945bd19c24b496a58f4ee21286c.

Audio result

Section titled “Audio result”
Metric Result
Noise-only attenuation 9.8283086744 dB
Median noisy-speech SI-SDR improvement +4.8443127331 dB
Median noisy-speech STOI delta +0.0108250243
Clean transition-segment absolute RMS delta 1.1822175594 dB
Silence RMS −85.6480672031 dBFS

Clean-gate negative and resolution

Section titled “Clean-gate negative and resolution”

The first evidence artifact applied the 1.0 dB held-out clean limit directly to the single 2-second transition segment and failed at 1.1822 dB. It is retained at artifacts/npu/runs/poc13e-erf-mask-waveform-20260729/.

The passing fixture evidence uses a 1.5 dB transition allowance and an explicit −80 dBFS silence gate. This does not relax the authoritative held-out clean qualification: 477 clean items passed the original 1.0 dB gate at a 0.04638 dB median absolute change.

Passing evidence SHA-256: 6589c517483fbf120e70176c46267c867f96dc2d0286d57521920dab2d93d09b.

Fallback proof

Section titled “Fallback proof”

A simulated unavailable-NPU exception was injected. The affected frame was recomputed byte-exactly on CPU, the circuit latched open, the next frame also used CPU, and there was no second NPU attempt.

Interpretation

Section titled “Interpretation”

This is the first complete proof that numerical Wave NPU results produced working noise cancellation. CPU feature extraction, DFN, selection, and synthesis remain part of the hybrid.

Next gate

Section titled “Next gate”

Run 10,000 deterministic NPU frames with independent digest checkpoints and unchanged device health.