Pace wake-word classifier

Pace wake-word classifier

PaceWakeWordClassifier.mlpackage is the bundled local keyword model for Always-On Companion Mode. It accepts a bounded two-second PCM window and runs before any Speech framework API is reachable.

Runtime contract

• Input: audio_samples, Float32 multi-array shaped [1, 32000].
• Audio: mono, 16 kHz, normalized to [-1, 1].
• Outputs: classLabel and classLabel_probs with exact labels background and hey_pace.
• Wake threshold: 0.986.
• Runtime policy: sample every 4,000 new samples and require two consecutive accepted windows. Pre-wake PCM remains in a bounded two-second ring plus one coalescing audio-ingress chunk; stop clears ingress immediately.

The model contains an audio-to-mel frontend, the frozen Google speech-embedding backbone, and a Pace-trained 45,169-parameter temporal CNN head in one Core ML package. No ONNX runtime or Python dependency ships in the app. Bundled weight SHA-256: 21bcc05e64da54f6bb26fcc06673cf951d0a24fce68d0efd4b7aa6614e42834c.

Training and evaluation

Training used locally synthesized Kokoro-82M v1.0 speech. Speaker families were disjoint by split:

The corpus included clean, quiet, room-echo, telephone-band, and pink-noise variants. Negative clips included close confusables such as “hey space,” “hey face,” “hey Grace,” “hey Ace,” “okay Pace,” and “Pace,” unrelated speech, pure noise, and silence. Threshold 0.986 was selected only from calibration results under a maximum 1% clip-level false-accept constraint.

Compiled Core ML results:

Across all 1,276 calibration and evaluation windows, the maximum absolute difference between Core ML and the PyTorch reference was 2.8610e-6. The mel and speech-embedding conversions independently matched their ONNX references within 1.4305e-6 and 2.8610e-5, respectively.

These are synthetic clip-level results, not false accepts per hour. They do not measure real microphones, rooms, accents, distance, hardware cost, or the production two-consecutive-window recall penalty. In particular, 55.94% held-out synthetic recall means misses remain common. The owner explicitly accepted that dogfood risk for this milestone; the hardware runbook remains the release gate.

Provenance and licenses

• Pace temporal head, conversion, and package metadata: MIT.
• livekit-wakeword feature-extraction resources and conversion reference: Apache-2.0.
• Google speech_embedding/1 backbone used by those resources: Apache-2.0.
• Kokoro-82M v1.0 weights used only to generate the local corpus: Apache-2.0.
• Acoustic transformations: FFmpeg filters over locally generated material.

No bundled openWakeWord wake-word head, CC BY-NC-SA model, commercial-service model, or third-party negative corpus was used.

The bundled PaceWakeWordClassifier-APACHE-2.0.txt license covers the Apache backbone and feature-extraction resources incorporated into the model.

Reproduction recipe

1. Generate “Hey Pace” positives and the documented hard negatives with the 28 speaker families above, keeping 16/4/8 voices disjoint.
2. Produce the five acoustic variants plus pure noise and silence at 16 kHz mono; never mix variants of one voice across splits.
3. Extract 16 consecutive 96-dimensional embeddings from each bounded two-second window with the Apache Google speech-embedding frontend.
4. Train the 45,169-parameter temporal CNN head with seed 927, AdamW (lr=0.001, weight decay 0.0003), positive-class weighting, and calibration-loss early stopping.
5. Choose the highest-recall calibration threshold whose false-accept rate is at most 1%, then evaluate that locked threshold once on the eight-voice holdout.
6. Convert the fixed 32,000-sample mel frontend, embedding backbone, and custom head into one Core ML ML Program; verify full-corpus output parity before replacing the bundled package.