Technical Articles

Train an ONNX Wake Word Model Online in 2026

Enter a custom phrase, generate a portable model, and run wake-word detection locally on multiple platforms.

Train an ONNX Keyword Spotting Model Online in 2026

Put several fixed voice commands in one compact KWS model and connect them directly to application actions.

Why openWakeWord Recall Varies by Wake Word

How a frozen speech-embedding backbone creates phrase-dependent separability, per-word tuning and “word picking.”

openWakeWord Not Detecting? Fix Low Recall Step by Step

Check model loading, 16 kHz audio, framing, scores, thresholds, VAD and training coverage before retraining.

Picovoice Free Trial Ended? Three Offline Alternatives

Train openWakeWord yourself, generate a portable ONNX/TFLite model, or stay with a commercial SDK.

Replace Picovoice on Python with an Offline ONNX Model

Run a custom wake word on Linux, Windows, or macOS without a runtime AccessKey.

Offline Keyword Spotting on Android

Detect fixed voice commands with ONNX Runtime and map them to app actions without cloud ASR.

Keyword Spotting in the Browser

Capture the microphone, run ONNX locally, and control the page without an inference backend.

Offline Voice Commands on Linux

Connect a multi-keyword model to a kiosk, robot, desktop tool, or local service with Python.

Custom Wake Word on ESP32-S3 with INT8 TFLite Micro

Run the tested ESP-IDF example, inspect measured latency, and replace the matched model and output head.

Custom Wake Word Training in 2026

Type a word, get a small offline ONNX model in minutes. No recordings, no GPU. English, German, French, Japanese, Chinese.

Picovoice Alternative — Custom Wake Words, No Recurring Fees

Picovoice's free tier stopped June 30, 2026. Migrate to a one-time-payment, offline ONNX model with no per-device fees.

openWakeWord Alternative — Stable Recall, No Tuning

openWakeWord's custom models swing from 25% to 90% recall. Voicute's from-scratch TCN holds 90%+ — in English, German, French, Japanese, and Chinese.

Stable 90%+ Wake Word Recall, Explained

Why frozen embeddings swing and from-scratch training stays stable — the architecture behind consistent recall.

The Best Wake Word Tools in 2026, Compared

openWakeWord vs Picovoice vs DIY vs Voicute — languages, multi-keyword, and the three questions to ask first.

Multi-Keyword Wake Word Detection: One Model, 10 Commands

Why one-word-per-model doesn't scale, and how a shared backbone detects 2–10 keywords in ~100 KB.

Multiple Wake Words, One Model

Every family member gets their own wake word, one firmware covers all SKUs, switching words is a config edit.

A Custom German Wake Word, No Recordings Needed

Why English tools fail on German, and how TTS synthesis + real-audio fine-tuning gets recall above 90%.

Custom Wake Words for Home Assistant Voice

The openWakeWord add-on path for English, and the sidecar path for non-English and multi-keyword models.