Compact ONNX wake-word model for detecting “Okay Hermes” in short local audio windows.
The current root-level v2 export is the WavLM-teacher → merged RepCNN-student model trained with the robust-v1 augmentation policy from wakeword-forge. Compared with the previous public artifact, v2 explicitly covers low-gain microphones, far-field listening, onset jitter, noise/reverb assets, and spectrogram masking.
config.json — canonical Hub config; points to wakeword_v2.onnx and wakeword_v2.json
README.md — model card and usage notes
LICENSE — Apache-2.0 license
What this model does
Takes a 3-second mono audio window.
Returns a direct probability that the wake phrase “Okay Hermes” is present.
Runs locally with ONNX Runtime.
Does not transcribe speech.
Does not identify speakers.
Does not send audio anywhere by itself.
Input
Name:waveform
Shape:(batch, time); use exactly 48,000 samples per row for a 3-second window at 16 kHz
Type:float32
Audio: mono PCM, 16 kHz, 3 seconds
Audio should be resampled to 16 kHz mono and padded or cropped to exactly 48,000 samples per window.
Output
Name:score
Shape:(batch,)
Type:float32
Meaning: wake-word probability for each input window
The output is already a probability. Do not apply another sigmoid unless you have changed the exported graph.
Thresholds
Recommended deployment threshold:
0.6973556280136108
The trained EER threshold is lower:
0.5128782391548157
The deployment threshold is intentionally higher because the v2 model gives a non-trigger zero-audio score around 0.624035120010376. Start with the recommended threshold, then tune upward for false wakes or downward for missed activations in your own room and microphone setup.
Basic trigger rule:
trigger = score >= 0.6973556280136108
For always-on use, apply smoothing or debouncing across overlapping audio windows. A practical starting point is 3-second windows every 0.5 seconds with 1–2 consecutive windows above threshold.
Python example
python
1import json
2import numpy as np
3import onnxruntime as ort
4from huggingface_hub import hf_hub_download
56repo_id ="Neopabo/okay-hermes-repcnn-onnx"78# Fetch config first so normal Hub usage touches a counted metadata file.9config_path = hf_hub_download(repo_id=repo_id, filename="config.json")10withopen(config_path,"r", encoding="utf-8")as f:11 config = json.load(f)1213model_path = hf_hub_download(repo_id=repo_id, filename=config["onnx_file"])# wakeword_v2.onnx1415opts = ort.SessionOptions()16opts.intra_op_num_threads =117opts.inter_op_num_threads =118opts.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
1920session = ort.InferenceSession(model_path, sess_options=opts, providers=["CPUExecutionProvider"])2122# Replace this with real mono 16 kHz audio, exactly 3 seconds long.23waveform = np.zeros((48000,), dtype=np.float32)2425score = session.run(26[config["output_name"]],27{config["input_name"]: waveform[None,:]},28)[0][0]2930trigger =float(score)>=float(config["recommended_threshold"])31print(float(score), trigger)
Technical summary
Wake phrase: Okay Hermes
Version: v2
Backend: wavlm-repcnn
Inference model: merged / reparameterized RepCNN
Teacher model: microsoft/wavlm-base
Audio frontend: included in the ONNX graph; callers provide waveform audio directly
Output: direct probability named score
ONNX input shape: dynamic time axis named time; use 48,000 samples for the documented 3-second window
Wake-word detection only; not general speech recognition.
Not speaker verification or identity recognition.
False accepts and false rejects are possible.
Thresholds may need adjustment for each deployment environment.
No training audio or source datasets are included in this repository.
Privacy
The model runs locally and only processes audio supplied by the surrounding application. Privacy depends on how that application captures, buffers, logs, stores, or transmits audio.
