Voice Realtime

Real-time conversational AI for macOS with push-to-talk interface and multiple personas.

Features

• Push-to-Talk: Hold Cmd+Shift+T to speak, release to get AI response
• Push-to-Dictate: Hold Cmd+Shift+D to speak, release to type at cursor
• Read Selection: Select text, press Cmd+Shift+S to hear it read aloud
• Multiple Personas: Switch between Assistant, Tutor, Creative, and Casual modes
• Hybrid LLM: RedPill GPU TEE models with cryptographic attestation
• Fast STT: Lightning Whisper MLX - optimized for Apple Silicon
• Natural TTS: Moshi neural speech synthesis
• Continuous Conversation: Maintains context across turns

Requirements

• macOS with Apple Silicon (M1/M2/M3/M4)
• Python 3.12+
• Hammerspoon (for hotkeys)
• RedPill API key (get one at redpill.ai)

Getting Started

1. Clone and Setup

git clone https://github.com/HashWarlock/nobody.git
cd nobody
./setup.sh

The setup script will:

• Create a Python virtual environment at ~/voice-env
• Install all dependencies
• Symlink Hammerspoon config (auto-updates when you pull changes)
• Download required ML models (~1GB)
• Create a symlink at ~/voice-realtime

2. Configure API Key

Create a .env file with your RedPill API key:

cp .env.example .env
# Edit .env and add your REDPILL_API_KEY

Or manually:

echo "REDPILL_API_KEY=your-key-here" > .env

3. Grant Permissions

• System Settings > Privacy & Security > Accessibility
  • Add Hammerspoon and grant access
• System Settings > Privacy & Security > Microphone
  • Grant access to Terminal (or your terminal app)
• System Settings > Privacy & Security > Automation (for dictation)
  • Allow Hammerspoon to control "System Events"

4. Start Using

1. Open Hammerspoon (you should see "Voice Realtime ready!" alert)
2. Hold Cmd+Shift+T and speak
3. Release to hear the AI response

Hotkeys

Hotkey	Action
Cmd+Shift+T	Push-to-talk (hold to speak, release to get AI response)
Cmd+Shift+D	Push-to-dictate (hold to speak, release to type at cursor)
Cmd+Shift+S	Read selection (highlight text, press to hear it spoken)
Cmd+Shift+X	Stop/Cancel
Cmd+Shift+M	Model picker (searchable menu to switch LLM)
Cmd+Shift+1	Switch to Assistant persona
Cmd+Shift+2	Switch to Tutor persona
Cmd+Shift+3	Switch to Creative persona
Cmd+Shift+4	Switch to Casual persona
Cmd+Shift+R	Reload Hammerspoon config

Personas

Configure in personas.yaml:

Persona	Provider	Description
Assistant	RedPill	Fast, concise answers
Tutor	RedPill	Patient explanations with examples
Creative	RedPill	Brainstorming and ideation
Casual	Ollama	Friendly local conversation

Voice Pipeline

Component	Technology	Description
STT	Lightning Whisper MLX	10x faster than whisper.cpp
TTS	Moshi MLX	Neural speech synthesis
LLM	RedPill API	GPU TEE models with cryptographic attestation

STT Models

The default STT model is distil-medium.en. Available models:

Model	Size	Speed	Accuracy
tiny	39M	Fastest	Basic
small	244M	Fast	Good
distil-medium.en	769M	Fast	Great (English)
large-v3	1.5B	Slower	Best

Change in main.py:

transcriber = WhisperTranscriber(model="distil-medium.en")

Configuration

Environment Variables

Create a .env file (see .env.example):

# Required for cloud personas
REDPILL_API_KEY=your-key-here

# Optional: Override Ollama host for local personas
OLLAMA_HOST=http://localhost:11434

Changing LLM Models

Edit personas.yaml to change which model each persona uses:

personas:
  assistant:
    llm:
      provider: "redpill"
      model: "deepseek/deepseek-v3.2"  # Default, strong reasoning

GPU TEE Models (18 Total)

All models run in hardware-secured GPU TEE environments with cryptographic attestation.

Phala Network (9 models)

Model ID	Name	Features
deepseek/deepseek-v3.2	DeepSeek v3.2	Reasoning
phala/uncensored-24b	Uncensored 24B	Uncensored
phala/glm-4.7-flash	GLM 4.7 Flash	Multilingual, fast
phala/qwen3-vl-30b-a3b-instruct	Qwen3 VL 30B	Vision
phala/qwen2.5-vl-72b-instruct	Qwen 2.5 VL 72B	Vision
phala/qwen-2.5-7b-instruct	Qwen 2.5 7B	General, fast
phala/gemma-3-27b-it	Gemma 3 27B	General
phala/gpt-oss-120b	GPT OSS 120B	General
phala/gpt-oss-20b	GPT OSS 20B	General, fast

Tinfoil (5 models)

Model ID	Name	Features
moonshotai/kimi-k2.5	Kimi K2.5	Default, reasoning
moonshotai/kimi-k2-thinking	Kimi K2 Thinking	Reasoning
deepseek/deepseek-r1-0528	DeepSeek R1	Reasoning
qwen/qwen3-coder-480b-a35b-instruct	Qwen3 Coder 480B	Code
meta-llama/llama-3.3-70b-instruct	Llama 3.3 70B	General

Chutes (1 model)

Model ID	Name	Features
minimax/minimax-m2.1	MiniMax M2.1	General

Near-AI (4 models)

Model ID	Name	Features
deepseek/deepseek-chat-v3.1	DeepSeek Chat v3.1	General
qwen/qwen3-30b-a3b-instruct-2507	Qwen3 30B	General
z-ai/glm-4.7	GLM 4.7	Multilingual
z-ai/glm-4.7-flash	GLM 4.7 Flash	Multilingual, fast

Which Model Should I Use?

Use Case	Recommended Model	Why
General chat	moonshotai/kimi-k2.5	Default, strong reasoning
Complex reasoning	deepseek/deepseek-r1-0528	Reasoning-optimized with R1 architecture
Long context	moonshotai/kimi-k2-thinking	Deep reasoning, extended thinking
Coding	qwen/qwen3-coder-480b-a35b-instruct	Code-specialized, 262k context
Vision tasks	phala/qwen2.5-vl-72b-instruct	Best vision model
Fast + balanced	meta-llama/llama-3.3-70b-instruct	Llama 3.3, good all-around
Uncensored	phala/uncensored-24b	No content restrictions
Multilingual	z-ai/glm-4.7	Strong multilingual support

Troubleshooting

"launch path not accessible" error

The Python path or script path is wrong. Run ./setup.sh again or check that ~/voice-env/bin/python exists.

No response from AI

Check that REDPILL_API_KEY is set in .env file.

Transcription is wrong/gibberish

Try a larger Whisper model (e.g., small or distil-medium.en).

Hammerspoon not responding

Reload config with Cmd+Shift+R or restart Hammerspoon.

Dictation not typing

1. Check System Settings > Privacy & Security > Automation - Hammerspoon needs permission to control "System Events"
2. Make sure a text field is focused when you release the hotkey
3. Try Cmd+V manually after dictating - if text pastes, it's a permission issue

License

MIT