Trigger Word for End-of-Speech Detection ("Walkie-Talkie Mode")

[evbrandy]

Feature Request: Trigger Word for End-of-Speech Detection ("Walkie-Talkie Mode")

Problem

When using VoiceMode in noisy environments (walking on streets, cafes, outdoors), the silence-based end-of-speech detection (WebRTC VAD) struggles to determine when the user has finished speaking. Background noise prevents reliable silence detection, leading to:

• Premature cutoffs (VAD thinks noise is the end of speech)
• Long waits (VAD never detects silence, waits for listen_duration_max)

Proposed Solution

Add support for an optional trigger word (like "over" in walkie-talkie/radio communication) that explicitly signals the end of speech. When the user says "over" at the end of their message, VoiceMode stops listening immediately.

Configuration

# In voicemode.env or environment variable
VOICEMODE_TRIGGER_WORD=over
# Or for non-English: VOICEMODE_TRIGGER_WORD=fertig

Behavior

1. User speaks: "Please help me refactor the authentication module, over"
2. VoiceMode detects trigger word → stops recording immediately
3. Trigger word is stripped from transcription → Claude receives: "Please help me refactor the authentication module"

Implementation Options

Option A: Post-transcription Detection (Simple)

After STT completes, check if transcription ends with trigger word and strip it.

Location: tools/converse.py around line 1655

# After: response_text = stt_result.get("text")
if TRIGGER_WORD and response_text:
    # Check for trigger word at end (case-insensitive)
    text_lower = response_text.lower().rstrip()
    trigger_lower = TRIGGER_WORD.lower()
    if text_lower.endswith(trigger_lower):
        # Strip trigger word
        response_text = response_text[:-(len(TRIGGER_WORD))].rstrip()
        # Also strip common trailing punctuation/comma before trigger
        response_text = response_text.rstrip('.,;:')

Pros: Simple, minimal code changes
Cons: Doesn't solve the waiting problem - still waits for silence/timeout before transcribing

Option B: Real-time Keyword Detection (Better UX)

Detect trigger word during recording to stop immediately.

Approaches:

1. Vosk - Lightweight offline speech recognition, can do streaming
2. Periodic STT - Send audio chunks every few seconds to Whisper, check for keyword
3. Porcupine/similar - Wake word detection (would need custom model training)

Location: tools/converse.py in record_audio_with_silence_detection()

Pros: Immediate response when user says trigger word
Cons: More complex, additional dependencies, potential latency

Recommended Approach

Start with Option A for quick implementation, then consider Option B as an enhancement if users need faster response in noisy environments.

Use Cases

1. Outdoor/mobile use - Walking, public transport, street noise
2. Open office - Background chatter interferes with silence detection
3. Hands-free operation - Clear signal without needing to press a button
4. International users - Can configure trigger word in their language

Workaround (Current)

Users can currently work around this by:

disable_silence_detection=True, listen_duration_max=30

But this requires waiting the full duration or awkward silence.

Additional Considerations

• Trigger word should be configurable (different languages, preferences)
• Should be optional (disabled by default for backward compatibility)
• Could support multiple trigger words: VOICEMODE_TRIGGER_WORDS=over,done,finished
• Consider adding a parameter to converse(): trigger_word: Optional[str] = None

---

Environment:

• VoiceMode version: 7.4.2
• Use case: concept elaboration via voice while mobile

[kindcreator]

@evbrandy
Have you read this thread? #58

(Direct link -> #58)

[evbrandy]

@kindcreator Thanks for the hint, the title was about "wake word" while I propose the "end-of-speech detection", quite the contrary. But it seems the other issue also discusses the end-of-speech situation.