Modelpublic

AfkaraLP/mister-president-wakeword

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes6downloads

Model Card

mister president — Wake Word Model

Custom wake word model for the phrase "mister president", trained with OpenWakeWord and Pocket TTS.

Model Summary

Property	Value
Wake word	`mister president`
Architecture	WakeWordNet
Input	16 frames x 96 features (raw audio embeddings)
Output	Probability [0, 1]
Calibrated threshold	0.75
Dimensions	16 frames x 96 features

Benchmark Results

Evaluated on held-out project data and (where available) OpenWakeWord test data.

Metric	Value
Threshold	0.75
Recall (val)	0.957
FP Rate (val)	0.058

Data Sources

—Synthetic training data generated via Pocket TTS
—Adversarial negative samples (TTS-generated similar phrases)
—Noise and silence negative examples

Limitations

—Trained on synthetic data from a single speaker pipeline (Pocket TTS).
—May have reduced accuracy on speakers with different accents or speech patterns.
—False activations can occur with speech that phonetically resembles "mister president".
—The model expects raw 16kHz mono audio, not pre-processed features.

Files

  mister_president.onnx
  mister_president.onnx.data
  mister_president_results.json
  mister_president_benchmark.json
  README.md

Install

bash

pip install openwakeword

Usage

Option 1: Streaming (recommended for real-time use)

python

import numpy as np
from huggingface_hub import hf_hub_download
from openwakeword import Model

# Download model from this repo
model_path = hf_hub_download(
    repo_id="<your-username>/mister-president",
    filename="mister_president.onnx",
)

# Load into openwakeword (supports multiple models)
oww = Model(wakeword_model_paths=[model_path])
model_name = list(oww.models.keys())[0]

# Stream audio in 1280-sample (80ms) chunks
sr, audio = scipy.io.wavfile.read("your_audio.wav")
if sr != 16000:
    import scipy.signal
    audio = scipy.signal.resample(
        audio, int(len(audio) * 16000 / sr)
    ).astype(np.int16)
audio = audio.astype(np.int16)

threshold = 0.75
for i in range(0, len(audio), 1280):
    chunk = audio[i : i + 1280]
    if len(chunk) < 1280:
        chunk = np.pad(chunk, (0, 1280 - len(chunk)))
    prediction = oww.predict(chunk)
    score = prediction.get(model_name, 0.0)
    if score >= threshold:
        print(f"Wake word detected! score={score:.3f}")

Option 2: Single clip inference (batch)

python

import numpy as np
import scipy.io.wavfile
import onnxruntime as ort
from huggingface_hub import hf_hub_download
from openwakeword.utils import AudioFeatures

# Download and load the ONNX model
model_path = hf_hub_download(
    repo_id="<your-username>/mister-president",
    filename="mister_president.onnx",
)
sess = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])

# Load audio (16kHz mono)
sr, audio = scipy.io.wavfile.read("your_audio.wav")
if len(audio.shape) > 1:
    audio = audio[:, 0]
if sr != 16000:
    from scipy.signal import resample
    audio = resample(audio, int(len(audio) * 16000 / sr)).astype(np.int16)
audio = audio[:32000]  # Pad or truncate to 2 seconds

# Compute embeddings and run inference
fe = AudioFeatures()
feats = fe.embed_clips(audio[None, :], batch_size=1).astype(np.float32)
score = sess.run(None, {sess.get_inputs()[0].name: feats})[0][0][0]
print(f'Wake word score: {score:.3f}')
print(f'Detected: {score >= 0.75}')