nanowakeword-wake-up

NanoWakeWord Model — "wake up"

A wake word detection model trained with nanowakeword for the phrase "wake up".

Model Files

Training Data

• Positive: TTS-generated "wake up" (2500 clips, 904 Piper speakers) + real recordings from pollen-robotics/wake-up-real + synthetic clips from TigreGotico/synthetic-wakeword-wake_up
• Negatives: 5000 adversarial TTS + 3000 phoneme-adversarial + speech commands from pollen-robotics/speech-commands-v0.02 + arcosoph/AE29H_float32 + arcosoph/RACON_11h_v1
• Background noise: Real robot mic recordings + arcosoph/datasets_zip (SonicWeave-v2)

Training Recipe

• Architecture: DNN (layer_size=64, n_blocks=4, embedding_dim=128, dropout=0.3)
• Loss: Bias-weighted asymmetric BCE (positive bias=0.65, negative weight=2.0)
• Optimizer: AdamW (lr=0.0008, weight_decay=0.01, onecycle schedule)
• Steps: 40,000 train + 8,000 distillation
• Augmentation: 10 rounds (noise, gain, pitch, SNR)
• Checkpoint averaging: Top-5 (EMA alpha=0.05)

Usage

Python (nanowakeword package)

from nanowakeword.model import Model

# Load the full model
model = Model.from_pretrained("pollen-robotics/nanowakeword-wake-up", filename="wake_up_v1.onnx")

# Or load the lite model
model = Model.from_pretrained("pollen-robotics/nanowakeword-wake-up", filename="wake_up_v1_lite.onnx")

ONNX Runtime (direct)

import onnxruntime as ort
import numpy as np

session = ort.InferenceSession("wake_up_v1.onnx")
# Input: (1, 32000) float32 audio at 16kHz
# Output: (1, 1) probability

Real-time detection (nanowakeword listener)

from nanowakeword.listener import Listener

listener = Listener(
    model_paths=["pollen-robotics/nanowakeword-wake-up"],
    wake_word_names=["wake_up_v1"],
    chunk_size_ms=100,
)
listener.start()

Deployment

This model is designed for Raspberry Pi 4 deployment on the Reachy Mini robot. Use the lite model (wake_up_v1_lite.onnx) for lower CPU usage, with the full model as a secondary verifier if needed.

<!-- ml-intern-provenance -->

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern