Her Voice 🎙️

Give your agent a voice. Audio responses powered by Kokoro TTS — a compact, naturally expressive model running entirely on-device.

✨ Features

Highly optimized response time thanks to on-the-fly audio streaming technology. 100% free, no API keys required. Inspired by Samantha and Sky.

• - ⚡ On-the-fly Streaming — Audio plays as it generates, very low latency
• 👄 The Voice of an angel — Cutting-edge local text-to-speech model Kokoro TTS
• 🧠 TTS Daemon — Keep the model warm in RAM for instant responses (can be disabled to save RAM)
• 🖥️ Persist Mode — Drag & drop audio, paste text, use as a voice station
• 🔧 Fully Configurable — Voice, speed, visualizer, notification sounds
• 🍎 MLX + PyTorch — Native Metal acceleration on Apple Silicon, PyTorch fallback everywhere else
• 🎨 Real-time Visualizer — Floating 60fps LED bars that react to speech (macOS only)

First-Run Setup

CODEBLOCK0

Note: SKILL_DIR is the root directory of this skill — the agent resolves it automatically when running commands.

The setup wizard will:

1. 1. Detect platform and select TTS engine (MLX on Apple Silicon, PyTorch elsewhere)
2. Find or install the appropriate TTS backend (mlx-audio or kokoro)
3. Install espeak-ng (Homebrew on macOS, apt on Linux)
4. Patch espeak loader if needed (macOS compatibility)
5. Compile the native visualizer binary (macOS only)
6. Download the Kokoro model
7. Create config at INLINECODE2

Check status anytime:
CODEBLOCK1

Post-Setup: Names & Pronunciation

After setup, configure the agent and user names:
CODEBLOCK2

TTS pronunciation tip: If the user's name is non-English, figure out a phonetic English spelling that Kokoro will pronounce correctly. Store it in user_name_tts and use that spelling whenever speaking the name aloud. The real name stays in user_name for display purposes.

Speaking Text

CODEBLOCK3

Options

Flag	Description
INLINECODE5	Skip the visualizer for this call
INLINECODE6

Keep visualizer open after playback ends | | --save PATH | Save audio to WAV file instead of playing | | --voice NAME | Override the configured voice | | --speed N | Override the configured speed multiplier | | --mode MODE | Override visualizer mode (v2 or classic) |

Agent Workflow

When the user wants voice responses:

1. 1. Check voice mode — is voice enabled or did the user ask for it?
2. Play notification sound (instant feedback while TTS generates):

   afplay /System/Library/Sounds/Blow.aiff &
   

1. 3. Speak the response:

   python3 SKILL_DIR/scripts/speak.py "Response text here"
   

1. 4. Always provide text alongside voice — accessibility matters.

Notification Sound

The notification sound plays instantly (~0.1s) while TTS generates (~0.3-3s). This gives the user immediate feedback that the agent is responding.

Configure in ~/.her-voice/config.json:
CODEBLOCK6

Available macOS sounds: Blow, Bottle, Frog, Funk, Glass, Hero, Morse, Ping, Pop, Purr, Sosumi, Submarine, Tink. Located in /System/Library/Sounds/.

TTS Daemon

The daemon keeps the Kokoro model warm in RAM, eliminating ~1.1s of startup overhead per call.

The daemon auto-resolves the mlx-audio venv — no need to find the venv Python manually.

CODEBLOCK7

INLINECODE28 auto-detects the daemon: uses it if available, falls back to direct model loading.

The daemon is optional. Without it, speech still works — just ~1s slower per call as the model loads each time. Skip the daemon to save ~2.3GB RAM.

Note: The daemon writes its PID file and socket after the model is fully loaded and ready to accept connections. They live in ~/.her-voice/ with restricted permissions (owner-only access). The daemon won't survive a reboot — start it again after restart if needed.

Visualizer

A floating overlay with three animated LED bars that react to speech in real-time. 60fps, native macOS (Cocoa + AVFoundation). macOS only — on other platforms, audio plays without the visualizer.

Modes

• - v2 (default) — Three-tier pure red, center raw amplitude, sides with lag
• classic — Original smooth gradient look

Controls

Key	Action
ESC	Quit
Space

Pause/Resume (file mode) |

| ← → | Seek ±5s (file mode) | | ⌘V | Paste text to speak (persist mode) |

Persist Mode

Keep the visualizer on screen between playbacks. Use as a standalone voice station: CODEBLOCK8

In persist mode:

• - Drag & drop audio files (.wav, .mp3, .aiff, .m4a) onto the visualizer to play them
• ⌘V pastes clipboard text → streams directly from TTS daemon with full visualizer animation
• Idle breathing — subtle center bar pulse when waiting for input

Standalone Usage

CODEBLOCK9

Disable Visualizer

CODEBLOCK10

Configuration

Config file: INLINECODE30

CODEBLOCK11

Key Settings

Key	Default	Description
INLINECODE31	INLINECODE32	Agent's name (e.g. "Jackie")
INLINECODE33

"" | User's real name | | user_name_tts | "" | Phonetic spelling for TTS (e.g. "Mah-toosh" for Matúš) | | voice | af_heart | Base voice name | | voice_blend | {af_heart: 0.6, af_sky: 0.4} | Voice blend weights | | speed | 1.05 | Speech speed multiplier | | language | en | Language code | | tts_engine | auto | TTS engine: auto, mlx, or pytorch | | model | mlx-community/Kokoro-82M-bf16 | Model identifier (MLX) | | visualizer.enabled | true | Show visualizer overlay | | visualizer.mode | v2 | Animation mode (v2/classic) | | visualizer.remember_position | true | Save window position between sessions | | notification_sound.enabled | true | Play sound before speaking | | notification_sound.sound | Blow | macOS system sound name | | daemon.auto_start | true | Advisory flag only — the daemon never self-starts. When true, the agent should start it on first voice use (saves ~1s/call, costs ~2.3GB RAM) | | daemon.socket_path | ~/.her-voice/tts.sock | Unix socket path |

Voice Selection

Voice Blending

Mix multiple voices for a unique sound. Configure voice_blend in config:
CODEBLOCK12

The blended voice is stored as a .safetensors file in the model's voices directory (e.g., af_heart_60_af_sky_40.safetensors). Create it by running TTS once — speak.py looks for the pre-blended file automatically.

Error Handling

Error	Cause	Fix
"mlx-audio not found"	Venv missing or broken	Run INLINECODE71
"espeak-ng not found"

Phonemizer missing | brew install espeak-ng | | Compilation failed | Xcode tools missing | xcode-select --install | | "Model not found" | First run, no download | Run setup.py or speak once | | Daemon "not running" | Crashed or rebooted | Start daemon again | | No sound output | macOS audio permissions | Check System Settings → Sound → Output | | Visualizer not showing | Binary not compiled | Run setup.py | | "kokoro not found" | PyTorch venv missing | Run setup.py | | PyTorch CUDA error | GPU driver mismatch | pip install torch --force-reinstall in kokoro venv | | "soundfile not found" | Missing dependency | pip install soundfile in kokoro venv |

Requirements

• - macOS + Apple Silicon recommended for best experience (MLX engine + visualizer + notification sounds)
• Linux/Intel Mac supported via PyTorch Kokoro engine (no visualizer)
• Windows is not supported
• Xcode Command Line Tools for visualizer on macOS (xcode-select --install)
• INLINECODE80 for phonemization (brew install espeak-ng on macOS, apt install espeak-ng on Linux)
• ~500MB disk (model + venv)
• ~2.3GB RAM when daemon is running

Uninstall

Remove all Her Voice data (config, venvs, compiled binary, daemon state):
CODEBLOCK13

How It Works

1. 1. Kokoro 82M — A compact neural TTS model with two backends: MLX (Apple's framework for native Metal GPU acceleration on Apple Silicon) and PyTorch (works everywhere). The engine is auto-detected based on platform, or can be forced via the tts_engine config option (auto, mlx, or pytorch)
2. Streaming — Audio generates and plays simultaneously. First sound in ~0.3s (with daemon) vs ~3s batch
3. Visualizer — Native macOS app (Swift/Cocoa) reads raw PCM from stdin, plays via AVAudioEngine with real-time amplitude metering
4. Daemon — Unix socket server holding the model in RAM. Eliminates Python import + model load overhead on every call