Reduce listener recomputation with exact embedding reuse

[petebytes]

Summary

• keep WakeWordModel.predict() stateless and add isolated WakeWordStream cache state
• have WakeWordListener automatically use a fresh stream while retaining fallback compatibility for predict-only wrappers
• reuse a speech embedding only when its 76-frame mel input exactly matches the shifted prior input
• batch all changed embedding windows into one ONNX call and reset stream state after detection

Why

The Python listener advances an overlapping 2-second window every 80 ms. Before this change, one prediction ran one mel inference, sixteen speech-embedding inferences, and one classifier inference. With one classifier that is 18 ONNX calls per prediction, or 225 calls/second at the listener's 12.5 Hz cadence.

A naive rolling mel buffer is not score-equivalent because the mel model can produce window-dependent values around silence. This implementation still computes the full mel tensor and uses exact equality before reusing any embedding.

Verification

Full upstream suite:

uv run --frozen pytest tests/ -q
79 passed, 2 skipped

The changed Python files pass Ruff check and format check. The package builds as both wheel and sdist.

A 107-clip downstream regression corpus covering 4,032 overlapping windows produced identical peak scores and identical results at every threshold:

Threshold	Recall before/after	False accepts before/after
0.5	20/25	50/82
0.6	20/25	48/82
0.7	20/25	48/82
0.8	20/25	46/82
0.9	18/25	34/82

Measured work on that corpus:

mel invocations:        4,032
embedding invocations:  4,032
classifier invocations: 4,032
total:                  12,096 = 37.5 invocations/second
embedding windows:      17,527 / 64,512
max peak score delta:   0

On 187 overlapping quiet-audio windows with single-threaded ONNX Runtime:

stateless CPU: 2.9967s
stream CPU:    0.2488s
reduction:     12.05x
max score delta: 0

Remaining gate

This is intentionally a draft until it completes a live microphone workday to compare real-world detections and false accepts.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}