Wake Word

Wake word detection enables a hands-free “Hey Siri” style experience. The service listens continuously for a keyword phrase, captures the user’s utterance that follows, and sends it to the AI automatically.

Starting Wake Word

await aiService.StartWakeWord("Hey Copilot");

Once started, the service:

1. Continuously listens for the wake word phrase
2. When detected, captures the user’s utterance via speech-to-text
3. Sends the utterance to the AI via TalkTo()
4. If the turn carries questions, skips wake word detection and listens directly for the user’s reply (conversation continuation)
5. Otherwise, loops back to listening for the wake word

Stopping Wake Word

aiService.StopWakeWord();

This cancels the background listening loop and returns the service to an idle state. The WakeWord property is cleared and IsWakeWordEnabled returns to false.

Checking State

if (aiService.IsWakeWordEnabled)
{
    Console.WriteLine($"Listening for: {aiService.WakeWord}");
}

Conversation Continuation

When the AI asks a question, the wake word loop skips keyword detection and immediately listens for the user’s reply. This creates a natural back-and-forth without requiring the wake word after every response.

“Is that a question” is a typed signal, not a guess at the wording: the AI answers with a structured turn whose Questions collection drives the decision. When the model can’t produce one, the service falls back to checking whether the reply ends in ?.

The conversation continues as long as the AI keeps asking. Once a turn carries no questions, the loop returns to wake word detection.

The follow-up window

An open microphone with nobody answering is a problem — the next unrelated thing said in the room becomes the answer. FollowUpTimeout bounds the wait:

aiService.FollowUpTimeout = TimeSpan.FromSeconds(30);   // default is 20 seconds
aiService.FollowUpTimeout = null;                       // wait indefinitely

When it expires, PendingQuestions is cleared and the loop goes back to requiring the wake word.

If voice interruption is enabled and the user says a quiet word during TTS playback, the conversation breaks immediately and returns to wake word detection. See Voice Interruption below.

Voice Interruption

During text-to-speech playback, the service can listen for voice interruptions. This is controlled by the QuietWords property:

// Default quiet words: cancel, quiet, shut up, stop, nevermind, never mind, hush
// Customize:
aiService.QuietWords = ["stop", "cancel", "quiet"];

// Disable voice interruption entirely:
aiService.QuietWords = null;

Two interruption behaviors:

• Quiet words — If the user says only a quiet word (e.g., “stop”), TTS is silenced and the conversation loop breaks, returning to wake word detection
• Other speech — If the user says anything else during TTS (e.g., “actually, tell me about…”), TTS is silenced and the new utterance is sent to the AI as the next message, continuing the conversation

Note

Quiet words must be the user’s entire utterance to trigger. Saying “cancel this appointment” will not trigger the quiet word — it will be treated as a new message to the AI.

Constraints

• Only one wake word at a time — calling StartWakeWord() while already active throws InvalidOperationException
• Exclusive with ListenAndTalk — ListenAndTalk() throws InvalidOperationException if wake word is active (they both use the microphone)
• Waits for in-progress work — StartWakeWord() waits for any active TalkTo() call to complete before starting the listener

Example: Settings Toggle

async Task ToggleWakeWord()
{
    if (aiService.IsWakeWordEnabled)
    {
        aiService.StopWakeWord();
    }
    else
    {
        await aiService.StartWakeWord("Hey Copilot");
    }
}

How It Works

Wake word detection is powered by Shiny.Speech’s ISpeechToTextService. The service calls ListenForKeyword() to detect the wake phrase, then ListenUntilSilence() to capture the user’s command. Both are platform-native implementations — no cloud API is needed for the wake word detection itself.

Note

The quality of wake word detection depends on the platform’s speech recognition engine. For best results, use short, distinctive phrases (2-3 words) that are unlikely to occur in background conversation.

Android

Android’s native speech recognition stops and restarts between segments, causing brief pauses during continuous wake word listening. This can result in missed wake words during the restart gap. For a smoother experience on Android, use Azure Speech Services which provides continuous streaming recognition without interruptions. See Platform Notes for setup details.