stimulus-speak-then

Voice-controlled Stimulus actions using local wake word detection.

Say "Hey Jarvis" to wake, then speak commands that trigger Stimulus actions.

Installation

npm install stimulus-speak-then
# or
yarn add stimulus-speak-then

Setup

1. Install the package

npm install stimulus-speak-then
# or
yarn add stimulus-speak-then

Models are automatically copied to public/models/ on install.

If automatic install fails, manually copy the models:

cp node_modules/stimulus-speak-then/models/*.onnx public/models/

2. Configure ONNX Runtime WASM

The package uses ONNX Runtime for wake word detection, which requires WebAssembly files. You must configure where these files are loaded from before importing the controller.

Option A: Use CDN (recommended for most setups)

// app/javascript/controllers/index.js
import { application } from "./application"
import * as ort from "onnxruntime-web"

// Configure WASM paths BEFORE importing the controller
ort.env.wasm.wasmPaths = "https://cdn.jsdelivr.net/npm/onnxruntime-web/dist/"

// Use dynamic import to ensure WASM config is set first
const { SpeakThenController } = await import("stimulus-speak-then")
application.register("speak-then", SpeakThenController)

Option B: Serve WASM files locally

Copy WASM files to your public directory:

cp node_modules/onnxruntime-web/dist/ort-wasm*.wasm public/

Then configure:

import { application } from "./application"
import * as ort from "onnxruntime-web"

ort.env.wasm.wasmPaths = "/"

const { SpeakThenController } = await import("stimulus-speak-then")
application.register("speak-then", SpeakThenController)

3. Register the controller (alternative if WASM is pre-configured)

If your bundler already handles WASM files or you've configured ort.env.wasm.wasmPaths elsewhere:

// app/javascript/controllers/index.js
import { application } from "./application"
import { SpeakThenController } from "stimulus-speak-then"

application.register("speak-then", SpeakThenController)

Usage

Add speak:command actions to elements. Say "Hey Jarvis" to wake, then speak the command.

<div data-controller="speak-then">
  <span data-speak-then-target="indicator"></span>
  
  <button data-action="speak:next->player#next">Next</button>
  <button data-action="speak:previous->player#previous">Previous</button>
  <button data-action="speak:pause->player#pause">Pause</button>
</div>

Multi-word commands

Use underscores for multi-word phrases:

<button data-action="speak:play_music->player#play">Play Music</button>
<button data-action="speak:turn_off_lights->home#lightsOff">Lights Off</button>

Configuration

<div data-controller="speak-then"
     data-speak-then-confidence-value="0.5"
     data-speak-then-sleep-value="5000"
     data-speak-then-models-path-value="/models"
     data-speak-then-wake-model-value="hey_jarvis_v0.1.onnx"
     data-speak-then-lang-value="en-US"
     data-speak-then-debug-value="true">
</div>

Value	Default	Description
confidence	0.5	Wake word detection threshold (0-1)
sleep	5000	Ms of silence before sleeping
models-path	/models	Path to ONNX model files
wake-model	hey_jarvis_v0.1.onnx	Wake word model filename
lang	en-US	Language for speech recognition (BCP 47 code)
debug	false	Enable debug logging to console

Combining with click actions

<button data-action="click->player#next speak:next->player#next">
  Next
</button>

Listening for events

<div data-controller="speak-then my-controller"
     data-action="speak-then:wake->my-controller#onWake speak-then:sleep->my-controller#onSleep speak-then:error->my-controller#onError">
</div>

Event	Description
speak-then:wake	Fired when wake word is detected
speak-then:sleep	Fired when returning to sleep after timeout
speak-then:error	Fired on errors (detail contains { error })

CSS classes

The controller adds speak-then-awake class to the element when listening:

.speak-then-awake {
  border-color: green;
}

.speak-then-awake [data-speak-then-target="indicator"] {
  display: block;
}

Using without the controller

import { SpeakThen } from "stimulus-speak-then"

const speakThen = new SpeakThen(element, {
  basePath: "/models",
  wakeModel: "hey_jarvis_v0.1.onnx",
  confidence: 0.5,
  sleepAfter: 5000,
  lang: "en-US",
  onWake: () => console.log("Listening..."),
  onSleep: () => console.log("Sleeping..."),
  onError: (error) => console.error("Error:", error)
})

await speakThen.start()

Custom Wake Words

You can train your own wake word using openWakeWord.

Training a Custom Model

1. Install openWakeWord:

   pip install openwakeword

2. Generate synthetic training data:

   python -m openwakeword.train_custom_model \
     --phrase "hey computer" \
     --output_dir ./my_wake_word \
     --n_samples 5000

3. Train the model:

   python -m openwakeword.train_custom_model \
     --phrase "hey computer" \
     --output_dir ./my_wake_word \
     --train

4. Export to ONNX: The training process outputs an ONNX file (e.g., hey_computer.onnx).

5. Copy to your models directory:

   cp ./my_wake_word/hey_computer.onnx public/models/

6. Configure the controller:

   <div data-controller="speak-then"
        data-speak-then-wake-model-value="hey_computer.onnx">
   </div>

Using Pre-trained Models

openWakeWord provides several pre-trained models. Download from the openWakeWord releases and copy to your models directory:

• hey_jarvis_v0.1.onnx (included)
• alexa_v0.1.onnx
• hey_mycroft_v0.1.onnx
• ok_google_v0.1.onnx

Note: The melspectrogram.onnx and embedding_model.onnx files are shared across all wake word models and must remain in your models directory.

How it works

1. Wake word detection runs locally using ONNX models (~3MB) via WebAssembly
2. Command recognition uses the Web Speech API after wake (requires Chrome)
3. Commands dispatch custom events (speak:command) that Stimulus routes to actions

Browser support

• Wake word: Any modern browser (runs in WebAssembly)
• Commands: Chrome/Edge (Web Speech API)

Requirements

• HTTPS (or localhost) for microphone access
• Chrome/Edge for command recognition after wake

Troubleshooting

WASM files not found (404 errors)

If you see errors like:

Failed to load resource: the server responded with a status of 404 (Not Found)
wasm streaming compile failed
both async and sync fetching of the wasm failed

This means ONNX Runtime can't find its WebAssembly files. Make sure you've configured ort.env.wasm.wasmPaths before importing the controller. See Configure ONNX Runtime WASM.

Models not found

If wake word detection fails silently, verify models were copied:

ls public/models/
# Should show: embedding_model.onnx  hey_jarvis_v0.1.onnx  melspectrogram.onnx

If missing, copy manually:

mkdir -p public/models
cp node_modules/stimulus-speak-then/models/*.onnx public/models/

Vite/Rails specific issues

With Vite, you must use dynamic imports to ensure WASM configuration happens first:

// This won't work - static imports are hoisted
import { SpeakThenController } from "stimulus-speak-then"  // Runs first!
ort.env.wasm.wasmPaths = "..."  // Too late

// This works - dynamic import respects execution order
ort.env.wasm.wasmPaths = "..."  // Runs first
const { SpeakThenController } = await import("stimulus-speak-then")  // Runs second

License

MIT