Show HN: Real-time AI (audio/video in, voice out) on an M3 Pro with Gemma E2B

On-device, real-time multimodal AI. Have natural voice and vision conversations with an AI that runs entirely on your machine.

Parlor uses Gemma 4 E2B for understanding speech and vision, and Kokoro for text-to-speech. You talk, show your camera, and it talks back, all locally.

Research preview. This is an early experiment. Expect rough edges and bugs.

I'm self-hosting a totally free voice AI on my home server to help people learn speaking English. It has hundreds of monthly active users, and I've been thinking about how to keep it free while making it sustainable.

The obvious answer: run everything on-device, eliminating any server cost. Six months ago I needed an RTX 5090 to run just the voice models in real-time.

Google just released a super capable small model that I can run on my M3 Pro in real-time, with vision too! Sure you can't do agentic coding with this, but it is a game-changer for people learning a new language. Imagine a few years from now that people can run this locally on their phones. They can point their camera at objects and talk about them. And this model is multi-lingual, so people can always fallback to their native language if they want. This is essentially what OpenAI demoed a few years ago.

Voice Activity Detection in the browser (Silero VAD). Hands-free, no push-to-talk. Barge-in. Interrupt the AI mid-sentence by speaking. Sentence-level TTS streaming. Audio starts playing before the full response is generated.

• Python 3.12+
• macOS with Apple Silicon, or Linux with a supported GPU
• ~3 GB free RAM for the model

Models are downloaded automatically on first run (~2.6 GB for Gemma 4 E2B, plus TTS models).

| Stage | Time | |---|---| | Speech + vision understanding | ~1.8-2.2s | | Response generation (~25 tokens) | ~0.3s | | Text-to-speech (1-3 sentences) | ~0.3-0.7s | | Total end-to-end | ~2.5-3.0s |

Decode speed: ~83 tokens/sec on GPU (Apple M3 Pro).