Soniqo KWS v1
Experimental local wake-word classifier for Hey Soniqo, Hi Soniqo, and Hello Soniqo. Intended pronunciations include «со́нико» and «сони́ко». Bare Soniqo and “Okay Soniqo” are not intended activations.
A small temporal convolutional network trained from scratch, with fixed log-mel preprocessing inside the graph. No shared pretrained acoustic embedding is required. VoxCPM2 generated synthetic training speech; VoxCPM2 is not included in this model. Training code, audio, and datasets are private and are not included.
Files and inference
model.mlpackage/: FP32 Core ML export, about 1 MiB.model.onnx: FP32 ONNX export, about 1 MiB, opset 17.inference_config.json: experimental streaming settings.release_manifest.json,evaluation_summary.json,checksums.json: identity, aggregate evaluation, and SHA-256 checksums.
Both graphs accept audio, float32 shape [1, 51200]: 3.2 seconds of mono
16 kHz PCM in [-1, 1]. They return score, float32 shape [1]. No external
mel transform or normalization is needed. Resample microphone audio to 16 kHz.
The caller maintains a chronological 3.2-second window, scores every 100 ms,
and requires two consecutive scores >= 0.5. Rearm after a score below threshold;
allow at most one event per second. Reset audio and detection state between
sessions. Threshold 0.5 is an experimental setting, not device calibration.
Training and measured limits
V1 used 3,850 automatically screened synthetic clips from 180 source voices, with 500 development clips from 20 other source voices. In isolated synthetic streaming probes at threshold 0.5 it detected 242/250 positives and triggered on 6/250 hard-negative clips. The Indian-reference subgroup detected 19/26.
A later source-disjoint synthetic accent audit detected 70/71 screened clips. Only 71 of 162 generated clips passed the automated ASR gate; rejected clips and small per-accent sample sizes limit interpretation. Source accents are self-reported; synthesized accent preservation was not independently verified.
Core ML and ONNX passed numerical parity checks against PyTorch. These are
synthetic development diagnostics, not validated real-world accuracy. Human
clip review, real microphone recall, noisy environments, continuous false
activations/hour, and physical iPhone latency/power have not been established.
The release manifest intentionally remains research_only, with
productionEligible: false. Exact phrase playback can activate the model;
it does not authenticate a speaker or establish command intent.
License
Copyright (c) 2026 Soniqo. All rights reserved. Public availability does not grant a broad reuse or redistribution license. See LICENSE.
- Downloads last month
- -