Voice Chat Skill (Relaxed Settings)

Start a voice conversation with relaxed silence detection settings.

Purpose

Quickly start voice mode with patient defaults that work well for natural conversation, avoiding premature cutoffs from aggressive silence detection. Use this instead of /voice when you want more forgiving listening.

Usage

Invoke with /voice-chat or /voice-chat [message]

Examples

code

/voice-chat                          # Start listening for voice input
/voice-chat Hello, what's on your mind?  # Speak a message and listen

Default Settings

Setting	Value	Reason
`voice`	"af_sky"	Preferred Kokoro voice
`vad_aggressiveness`	0	Least aggressive - won't cut off speech prematurely
`listen_duration_min`	4	Wait at least 4 seconds before considering silence
`listen_duration_max`	120	Allow up to 2 minutes of speaking
`wait_for_response`	true	Listen for user response after speaking
`metrics_level`	"minimal"	Clean output - just response text, saves tokens

Workflow

Starting Voice Mode

•If no message provided: Start listening immediately with a friendly prompt
•If message provided: Speak the message, then listen for response

Continuing Conversation

•Continue using the mcp__voice-mode__converse tool with the same settings
•Each response loops back to listening mode
•Conversation continues until user says "stop" or "end voice mode"

Ending Voice Mode

User can say:

•"Stop voice mode"
•"End voice"
•"Switch to text"
•"That's all for now"

When ending, acknowledge and return to text-based interaction.

Implementation

When this skill is invoked, use these tool parameters:

python

mcp__voice-mode__converse(
    message="[user message or greeting]",
    wait_for_response=True,
    voice="af_sky",
    vad_aggressiveness=0,
    listen_duration_min=4,
    listen_duration_max=120,
    metrics_level="minimal",
)

Multi-Sentence Responses (Preferred)

For longer responses, break them into multiple sequential tool calls. This provides better readability and a more natural conversation flow:

python

# For multi-sentence responses, queue them up like this:
mcp__voice-mode__converse(message="First sentence.", wait_for_response=False, ...)
mcp__voice-mode__converse(message="Second sentence.", wait_for_response=False, ...)
mcp__voice-mode__converse(message="Final sentence. Now listening.", wait_for_response=True, ...)

Guidelines:

•Set wait_for_response=False for all messages except the last one
•Only the final message should have wait_for_response=True
•Keep each message to 1-2 sentences for natural pacing
•Queue all messages in a single tool call block so they play sequentially

Adjusting Settings

Users can request adjustments:

•"Be more aggressive with silence detection" → vad_aggressiveness=2
•"Listen longer before cutting off" → listen_duration_min=6
•"Quick responses only" → listen_duration_max=30

Troubleshooting

Voice not being detected

•Check microphone permissions
•Try mcp__voice-mode__service("whisper", "status") to verify STT service

Speech being cut off early

•Lower vad_aggressiveness to 0
•Increase listen_duration_min

Service not running

python

mcp__voice-mode__service("whisper", "start")
mcp__voice-mode__service("kokoro", "start")

Key MCP Resources

•voicemode://docs/quickstart - Basic usage
•voicemode://docs/parameters - Full parameter reference
•voicemode://docs/troubleshooting - Common issues

Notes

•Voice mode works best in quiet environments
•Responses are transcribed via Whisper STT
•TTS uses Kokoro or OpenAI depending on configuration

TTS Pronunciation Quirks

Hyphens are often interpreted as "minus" by TTS:

•"T-004" → "T minus 004" (say "T zero zero four" instead)
•"8-10" → "8 minus 10" (say "8 to 10" instead)

Be mindful of phrasing when the text will be spoken aloud.