Voice Loop

A minimal on-device voice agent loop. Runs entirely on Mac M4 / Apple Silicon.

Now with even snappier audio responses — TTS starts on the first sentence while the LLM is still generating the rest, so you hear the reply almost immediately.

Need a custom voice model or production voice agent? See Trelis Voice AI Services.

Features

• Smart turn detection — Silero VAD + pipecat's Smart Turn v3, so the agent waits when you pause mid-sentence
• Voice interruption — speak over the agent; WebRTC AEC3 cancels echo from speakers so your voice cuts through
• Editable persona — SOUL.md controls the agent's style, live-reloaded each turn
• Optional long-term memory — enable with --memory; the agent learns durable facts about you in MEMORY.md and consolidates every 5 turns
• Fully local — no API keys, no cloud. Everything runs on-device

Stack

• Moonshine (CPU) for speech-to-text transcription
• Gemma 4 E4B (MLX/Metal) for response generation
• Kokoro (CPU) for TTS (streaming)
• Silero VAD + Smart Turn v3 for turn detection
• WebRTC AEC3 (via LiveKit APM) for voice interruption

Setup

brew install portaudio espeak-ng uv
git clone https://github.com/TrelisResearch/voice-loop.git
cd voice-loop
uv sync

First run downloads Gemma 4 E4B (~3GB), Moonshine (~250MB), Kokoro (~300MB).

Usage

# Recommended defaults (TTS + smart turn + voice interrupt all on)
uv run voice_loop_mac.py

# + chime on utterance + soft ticks while generating
uv run voice_loop_mac.py --chime

# + persistent memory (reads/writes MEMORY.md)
uv run voice_loop_mac.py --memory

# Text-only mode (no TTS)
uv run voice_loop_mac.py --no-tts

# Disable voice interruption (keypress only)
uv run voice_loop_mac.py --no-aec

# Different voice (see below)
uv run voice_loop_mac.py --voice bf_emma

# Use the smaller E2B model (faster, slightly lower quality)
uv run voice_loop_mac.py --model mlx-community/gemma-4-E2B-it-4bit

# Custom silence timeout
uv run voice_loop_mac.py --silence-ms 500

# Debug: record mic stream to a WAV
uv run voice_loop_mac.py --record

Recommended Kokoro voices

Only the higher-quality voices are listed here:

Voice	Accent	Gender	Notes
af_heart	US	Female	Top pick — Grade A (default)
af_bella	US	Female	Grade A-, HH training
bf_emma	UK	Female	Grade B-, HH training
am_fenrir	US	Male	Grade C+, H training
am_puck	US	Male	Grade C+, H training
am_michael	US	Male	Grade C+, H training
bm_fable	UK	Male	Grade C, MM training
bm_george	UK	Male	Grade C, MM training

Architecture

   Mic (16kHz) ──► Silero VAD ──► Smart Turn ──► Moonshine ──► Gemma 4 E4B ──► Kokoro ──► Speakers
                                                                    ▲                         │
                                                        SOUL.md + MEMORY.md                   │
                                                                                              ▼
   Mic during TTS ──► WebRTC AEC3 (LiveKit APM) ──► Silero VAD ──► voice interrupt ◄──────────┘

How it works

1. Mic capture via sounddevice (16kHz mono)
2. Silero VAD detects speech vs silence
3. Smart Turn confirms end-of-turn on silence (default on)
4. Moonshine transcribes your audio to text (CPU)
5. Gemma 4 E4B responds using SOUL.md (+ MEMORY.md if --memory) as system prompt
6. Kokoro synthesizes sentence pairs in a background thread while the previous pair plays — gapless, low-latency audio
7. WebRTC AEC3 cleans mic during TTS playback → Silero VAD on cleaned audio → voice interrupt

Press any key during TTS to interrupt.

Persona & Memory

• SOUL.md — persona / style (always loaded, live-reloaded each turn)
• MEMORY.md — long-term facts. Only read/written when --memory is passed. When enabled, the agent extracts new durable facts after each turn and consolidates every 5 turns.

Both files are re-read at the start of every turn, so edits take effect immediately.

Memory usage

~3.5 GB total. Fits easily in 16GB.

Credits

Built with:

• Moonshine — STT
• Kokoro — TTS
• Silero VAD — voice activity detection
• Smart Turn v3 — end-of-turn detection
• LiveKit APM — WebRTC AEC3
• mlx-vlm — MLX multimodal inference
• Gemma 4 — LLM

Raspberry Pi

See docs/raspberry-pi.md for a full feasibility analysis. Short version: Pi 4 is too slow; Pi 5 + LiteRT-LM is the best current option but still borderline for natural-feeling voice.

License

Apache 2.0.