Downloads · 30 days
0
12384-0ashntoi/mr-president-openwakeword
mr-president-openwakeword is a audio classification model from 12384-0ashntoi. Use it for the audio classification task on the model card, and read the license before you ship it in a product. It is set up for openwakeword. The card lists the license as apache-2.0.
A custom openWakeWord wake-word detection model for the phrase "mister president".
Downloads · 30 days
0
Access
Public
Updated Jul 23, 2026
Repo size
4.3 MB
Likes
1
Public
Click a slice to open those files.
.data2.2 MB · 50%
From the Hugging Face model README
A custom openWakeWord wake-word detection model for the phrase "mister president".
Trained on synthetic speech from three complementary TTS families:
en_US-libritts_r-medium, 700 speakers) — bulk timbre, pace, and prosody coverageOn top of the modal-voice clips we add a whisperized held-out subset (STFT phase-randomisation of modal clips) so the model is evaluated against whispered "mister president". Training features include raw positive windows, stochastic openWakeWord augmentation, adversarial near-homophone negatives, ACAV100M background features, and hard-negative mining against the 10.7-hour openWakeWord validation corpus.
End-to-end openWakeWord embedding-window evaluation, 80 ms frames, 1580 held-out modal positives + 700 held-out whispered positives + 700 held-out Kokoro expressive positives + 1200 held-out adversarial negatives, plus the 10.7-hour general-speech FP corpus that ships with the base openwakeword package.
| Threshold | Recall<br>(modal) | Recall<br>(whisper) | Recall<br>(Kokoro voices) | Adversarial<br>FPR | FP / hour<br>(10.7 h corpus) |
|---|---|---|---|---|---|
| 0.5 | 99.8 % | 99.9 % | 100.0 % | 2.1 % | 0.00 |
| 0.7 | 99.8 % | 99.9 % | 100.0 % | 1.8 % | 0.00 |
| 0.9 | 99.8 % | 99.9 % | 100.0 % | 1.1 % | 0.00 ← recommended |
af_bella, af_heart, af_nicole, am_puck, and am_santa.mister presidents, sister president, mystery pheasant, mr president, missed her president, minister president, mister precedent, …) that fire. Intentionally hard: those phrases are near-homophones or phrase-context confusables for "mister president".import openwakeword
from openwakeword.model import Model
# Download the shared melspectrogram + embedding backbone (~4 MB, one-off)
openwakeword.utils.download_models()
model = Model(
wakeword_models=["mister_president.onnx"],
inference_framework="onnx",
)
# Feed 16 kHz mono 16-bit PCM audio in 80 ms frames (1280 samples)
prediction = model.predict(frame)
if prediction["mister_president"] > 0.9: # recommended threshold
print("wake!")
Keep mister_president.onnx.data beside mister_president.onnx; the exported ONNX model uses external tensor data.
| File | Purpose |
|---|---|
mister_president.onnx | The classifier head. Input [1, 16, 96] embeddings, single sigmoid output. |
mister_president.onnx.data | External tensor data required by the ONNX file. |
negative_phrases.txt | Adversarial phrases used to generate hard negatives. Format <count>\t<phrase>. |
mister_president_config.yaml | Training and inference configuration. |
metrics.json | Full evaluation metrics. |
metrics_by_threshold.json | Threshold sweep at 0.5, 0.7, and 0.9. |
hey, okay, hello, yo, dear) and suffixes (please, now, today, sir, again, right now). Examples: mister presidents, sister president, mystery pheasant, missed her president, mister precedent, mister resident, minister president, mister present, mister pleasant, mister peasant.Positive and adversarial clips run through the openWakeWord augmentation pipeline 3 rounds, with background mixing, room impulse responses, pitch shift, band-stop filtering, seven-band EQ, tanh distortion, gain changes, and coloured noise:
openwakeword).1.19e-7.0.9 — the highest-value operating point on this model:
Lowering to 0.5 does not improve measured recall on this held-out set and increases adversarial confusable fires from 1.1 % to 2.1 %.
mister president has many realistic near-homophones and phrase-context confusables. The adversarial set is intentionally hard and some confusables still fire even at 0.9.metrics.json also retains stochastic augmented split metrics from one destructive augmentation pass for transparency.Apache-2.0 for code and this model card. The model weights inherit the base openWakeWord backbone's CC-BY-NC-SA-4.0 because they were derived using Google's speech-embedding model under those terms.