thalovant/zoguilo-wakeword

The wake word "Zoguilo", said zo-gui-lo /zo.ɡi.lo/ ("zo-ghee-lo"): an invented three-syllable word, soft and easy for a child to say, with no web or trademark hits, for French, English, Spanish, Portuguese, Italian, German and Dutch speakers. Each language's spelling was trained: fr Zoguilo, en Zo-ghee-lo, es Soguilo, pt Zoguilô, it Zoghilo, de Sogilo, nl Zogielo, pl Zogilo, ro Zoghilo.

It is a 0.9 MB ONNX head trained with wakeforge on the streamed features of TigreGotico/wakehubert-tiny (int8), loaded in OpenVoiceOS by ovos-ww-plugin-wakeforge in place of openWakeWord. This is v2 (main, git tag v2); earlier versions are kept under their tags (Versions).

Use

OpenVoiceOS

pip install --pre "ovos-ww-plugin-wakeforge>=0.18.0a1"

mycroft.conf:

{
  "listener": {"wake_word": "zoguilo"},
  "hotwords": {
    "zoguilo": {
      "module": "ovos-ww-plugin-wakeforge",
      "model": "https://huggingface.co/thalovant/zoguilo-wakeword/resolve/v2/zoguilo.onnx",
      "listen": true,
      "active": true
    },
    "hey_mycroft": {"active": false}
  }
}

The plugin downloads the model once (to ~/.local/share/wakeforge/), fetches the featurizer the model names at a pinned revision, and uses the model's own threshold. If your openWakeWord entry has another name than hey_mycroft, put "active": false on that one.

• x86 PCs without AVX-512 (AVX2: most desktops and laptops): the released plugin scores up to 0.07 off there until OpenVoiceOS/ovos-ww-plugin-wakeforge#55 is merged; until then install pip install "git+https://github.com/goldyfruit/ovos-ww-plugin-wakeforge@perf/x64-int8-precision". ARM boards, Apple silicon and AVX-512 CPUs score exactly with the released plugin.
• Less CPU: "head_skip_below": 0.01 (in #56, from git+https://github.com/goldyfruit/ovos-ww-plugin-wakeforge@perf/head-skip) cuts the CPU by a third (Raspberry Pi 4: 106 to 72 ms per second of audio); about one detection in four can come 80 ms later.

Thalovant satellite (thalovant-voice 0.10.0 or later)

wake:
  word: /path/to/zoguilo.onnx

Threshold

The default, 0.864, was set on 97 h of development speech, never the test set (0.49 debounced false wakes per hour on it, 0.55 on its YouTube part). Override it with "threshold": lower answers more readily, higher wakes less by mistake. Measured on the held-out audio below:

threshold	false wakes/h: YouTube / debates / music	wake words found, 18 conditions (mean)
0.6	0.64 / 0.44 / 0.20	84.6 %
0.8	0.32 / 0.33 / 0.05	81.0 %
0.864 (default)	0.22 / 0.33 / 0.05	78.9 %
0.948	0.10 / 0.22 / 0.05	73.9 %

How good it is

Every number is on audio the model never trained on: speakers, voices, videos, rooms, noises and music tracks held out of training. Raw reports: eval.json, stress_recall.json, stress_false_wakes.json, latency.json, plugin_validation.json.

False wakes, streamed through ovos-plugin-manager as OVOS runs it (198 h in all, 55 false wakes, 0.28/h):

audio	hours	false wakes	per hour
read speech (MLS, LibriSpeech test) + AudioSet	70	26	0.37
YouTube videos (YODAS2, 7 languages)	72	16	0.22
European Parliament debates (VoxPopuli, 8 languages)	36	12	0.33
music, with and without vocals (MUSAN)	20	1	0.05

Wake words found, each held-out clip in each condition (edge-tts voices excluded from training and clips voice-converted onto unseen real speakers, 522 clips that a phoneme model confirms say the word; fr, en, es, pt, it, de, nl, pl, ro). English voices: 82.1 % on average over the 18 conditions, other languages: 77.8 %.

condition	found
after silence	88 %
speech right before it, no pause	86 %
a command right after it	89 %
in the middle of a sentence	85 %
far side of a real room (-36 dBFS, room noise)	86 %
noise, 10 dB SNR	89 %
noise, 5 dB SNR	78 %
noise, 0 dB SNR	67 %
music with vocals, 10 dB	85 %
music with vocals, 5 dB	73 %
another person talking over it, 5 dB	53 %
phone line (8 kHz G.711)	82 %
Opus at 12 kbit/s	89 %
said fast (x1.25)	84 %
said slowly (x0.8)	91 %
quiet speaker (-48 dBFS)	59 %
clipped microphone	63 %
pitch +4 semitones (child-like)	74 %

Latency: the wake word is reported a median 86 ms after the end of the word, and 90 % of the time within 164 ms, through the plugin at 1280-sample chunks.

The training pipeline's own held-out evaluation at the default threshold: 70 h of read speech by unseen speakers, 0.33 false wakes/h; wake words found after silence / after speech: edge-tts voices 94 % / 93 %, OmniVoice speakers 96 % / 94 %, voice-converted real speakers 95 % / 92 %.

Against openWakeWord

The same harness, the same audio and the same voices, each model through its own OVOS plugin (openWakeWord 0.6.0, onnx):

	false wakes/h: read / YouTube / debates / music	found, English voices	found, other languages	latency median / p90
zoguilo v2	0.37 / 0.22 / 0.33 / 0.05	82.1 %	77.8 %	86 / 164 ms
openWakeWord hey_jarvis	0.14 / 0.25 / 0.00 / 0.15	85.5 %	25.1 %	78 / 255 ms
openWakeWord alexa	0.74 / 1.56 / 0.44 / 4.84	87.9 %	71.1 %	-3 / 116 ms

Against openWakeWord's alexa, Zoguilo v2 wakes by mistake less on every kind of audio (7 times less on YouTube; with music, 1 false wake in 20 hours against 97) and finds the word about as often overall: more often for non-English speakers, less often for English ones. Against hey_jarvis, it wakes by mistake less on YouTube and music and more on read speech and debates (0.37 against 0.14, 0.33 against none per hour), finds a few points fewer English speakers, and three times more speakers of other languages.

CPU and memory, one thread, ms of CPU per second of audio (10 = 1 % of one core), measured through ovos-plugin-manager on the devices themselves:

device	zoguilo	with head_skip_below: 0.01	openWakeWord hey_jarvis
Raspberry Pi 3 (Cortex-A53)	244.7	159.3	508.0
Raspberry Pi 4 (Cortex-A72)	106.5	72.0	248.8
Raspberry Pi 5 (Cortex-A76)	40.4	28.6	102.6
Jetson Orin Nano (Cortex-A78AE)	36.8	24.9	86.4
Ryzen 7 5800 desktop (AVX2)	10.0		22.0
Linux x86_64, AMD EPYC 7763 (AVX2) (GitHub runner)	14.8		29.5
Linux aarch64, Neoverse N2 (GitHub runner)	18.3		44.9
macOS 15, Apple M1 (GitHub runner)	14.0		37.2
Windows x64 (AMD) (GitHub runner)	16.9		35.4
Windows ARM64 (GitHub runner)	19.5		50.8

Process memory on the desktop with one wake word: 107 MB in all (+33 MB for the model and the featurizer), against 217 MB (+143 MB) for openWakeWord. Two wake words in OVOS share wakeforge's featurizer: 17.6 ms/s against 46.3 ms/s on the desktop.

Runs everywhere, scores the same

On every 64-bit platform tried, every 80 ms score of a 106 s golden stream matches the AVX-512 VNNI machine the model was calibrated on within 1.2e-5 (tolerance 0.002) and the detections are identical: Linux x86_64 (onnxruntime 1.20.1, 1.23.2, 1.31.1) and aarch64, macOS on Apple silicon, Windows x64 and ARM64, and the real Raspberry Pi 3, 4, 5 and Jetson Orin Nano above (with the AVX2 fix of #55 on x86 PCs). 32-bit ARM is not supported: onnxruntime publishes no armv7 wheel. Cross-platform CI.

Training

• Positives: synthetic speech in 9 languages (OmniVoice including child voices, edge-tts, Piper, Kokoro/StyleTTS2, PocketTTS, Supertonic, Magpie, Google) and Chatterbox voice conversion onto MLS and LibriSpeech speakers, every clip checked by a phoneme model.
• Negatives: MLS and LibriSpeech speech, 140 h of YouTube videos (YODAS2, CC BY 3.0; other videos than the test's), music (MUSAN, other artists than the test's), AudioSet; near-miss words and phrases in 7 languages, and the 2 s before every detection of the previous version on that YouTube and music audio.
• Real-room conditions in training (new in this version): speech right before and after the word, another talker over it, room responses up to 1.2 s with their tail (MIT IR survey, RWCP, simulated rooms), quieter audio.
• Two GRU heads (with and without the voice-converted clips) averaged and calibrated; epoch and threshold chosen on 97 h of development speech that includes 71 h of other YouTube videos.

Versions

• v2 (main, tag v2): this model. Against v1, on the same held-out audio: false wakes on YouTube 1.74/h to 0.22/h, debates 0.89/h to 0.33/h, music 0.85/h to 0.05/h; found in the middle of a sentence 83 % to 85 %. v1's lower threshold found more in the easier cases: after silence 97 % (v2 88 %), a quiet speaker 87 % (v2 59 %), a clipped microphone 92 % (v2 63 %). A threshold of 0.6 brings v2 to 85 % on average, still with fewer false wakes than v1 (table above).
• v1 (tag v1): trained on read speech only; threshold 0.253; kept for comparison.

Limits

• Every test voice is synthetic or voice-converted; nobody has recorded "Zoguilo" for the tests yet.
• Someone talking over the word at a similar level (a loud TV) is the hardest case: about half are found.
• A quiet or distant speaker (a child speaking softly) and a clipping microphone are found less often than in v1: lower the threshold (0.6) where that matters more than the odd false wake.

License

CC BY-NC 4.0: free for non-commercial use with attribution ("Zoguilo wake word by Thalovant"); commercial use requires permission from Thalovant. The WakeHuBERT featurizer it runs on is Apache-2.0 and licensed separately. Versions downloaded before 2026-11-05 were published under Apache-2.0.