/voice-chat

Voice-controlled Claude Code. Speak your commands/questions, Claude processes them verbatim (as if you typed them), and responds with Inworld Olivia voice.

Usage

/voice-chat               # Start voice-controlled mode
/voice-chat stop          # Exit voice mode

Requirements (from .claude/voice-chat-requirements.md)

R1: Voice Input → Claude Code Verbatim

Your spoken words are transcribed and passed directly to Claude Code — exactly as if you typed them.

R2: Response in Inworld Olivia Voice

Claude Code's response is spoken using Inworld TTS with Olivia voice.

Prerequisites

1. voice-bridge server running with INWORLD_API_KEY set:

   cd ~/Documents/GitHub/voice-bridge
   VOICE_BACKEND=inworld INWORLD_API_KEY=<your-key> npm start
   

2. Microphone permissions granted

Instructions

Starting Voice Mode

When /voice-chat is invoked:

1. Announce start:

   mcp__voice-mcp__voice_speak({
     text: "Voice mode active. Speak your command.",
     voice_id: "Olivia",
     speed: 1.0
   })
   

2. Display: 🎤 **Listening...**

3. Listen for voice:

   mcp__voice-mcp__voice_listen({
     timeout_seconds: 30,
     language: "en-GB"
   })
   

4. Display the transcription: 👤 **You said:** "<transcribed text>"

5. CRITICAL: Process the transcribed text as if the user typed it directly.

   • Do NOT interpret it as a "voice chat question"
   • Execute it as a normal Claude Code command/request
   • Use all available tools (Read, Edit, Grep, Bash, etc.) as needed
   • Generate your normal response

6. Display speaking indicator: 🔊 **Speaking...**

7. Speak the response (or summary if long):

   mcp__voice-mcp__voice_speak({
     text: "<response or summary>",
     voice_id: "Olivia",
     speed: 1.0,
     wait: true
   })
   

8. Smart Output Router - Decide what to speak:

   Content Type	Action	Voice Says
   Short (< 100 words)	Speak fully	Full response
   Medium (100-300 words)	Display + announce	"Here's what I found"
   Long (> 300 words)	Display only	"Please read this"
   Contains code	Display + announce	"Here's the code"
   Contains table	Display + announce	"Here's a table"
   Error	Speak	Brief error description

9. Loop back to step 2 for next command

Exit Conditions

• User says: "stop", "exit", "bye", "goodbye", "end voice mode"
• User types: /voice-chat stop
• Timeout with no input

Stopping Voice Mode

mcp__voice-mcp__voice_speak({
  text: "Voice mode ended.",
  voice_id: "Olivia",
  speed: 1.0
})

Display: 👋 **Voice mode ended.** Returning to text input.

Example Flow

User: /voice-chat

🎤 **Listening...**
Olivia: "Voice mode active. Speak your command."

[User speaks: "What projects am I working on?"]

👤 **You said:** "What projects am I working on?"

[Claude searches vault, finds projects]

🔊 **Speaking...**

🤖 **Claude:** You have 3 active projects: Alpha, Beta, and Gamma.

Olivia: "You have 3 active projects: Alpha, Beta, and Gamma."

🎤 **Listening...**

[User speaks: "Create a meeting note for today"]

👤 **You said:** "Create a meeting note for today"

[Claude creates meeting note using normal workflow]

🔊 **Speaking...**

🤖 **Claude:** Created Meeting - 2026-01-29 [Title].md

Olivia: "Meeting note created."

🎤 **Listening...**

[User speaks: "Stop"]

Olivia: "Voice mode ended."

👋 **Voice mode ended.** Returning to text input.

Key Principle

The transcribed text IS the user's command. Process it exactly as you would if they typed it. The only difference is:

• Input comes from voice transcription
• Output is spoken via Inworld Olivia TTS

Troubleshooting

• Male voice playing: Check INWORLD_API_KEY is set and voice-bridge restarted
• No voice at all: Check voice-bridge is running on port 4000
• TTS error: Verify API key is valid base64-encoded Inworld key