# Volocal

A fully local voice AI app for iOS that runs speech-to-text, a language model and text-to-speech entirely on an iPhone's Neural Engine, GPU and CPU, with real-time conversation and no cloud calls after setup.

Volocal streams speech through a local speech-to-text model, a local large language model and a local text-to-speech model to hold a real-time voice conversation entirely on an iPhone, with no cloud service, no API keys and no internet access needed after the initial model download. It supports interrupting the assistant mid-sentence and uses hardware echo cancellation so the microphone does not pick up the app's own spoken output.

The three models are deliberately spread across different compute units so they do not compete for the same hardware: speech-to-text runs on the Neural Engine through CoreML, the language model, a quantized Qwen3.5-2B, runs on the GPU through llama.cpp and Metal, and text-to-speech runs on CPU and GPU through CoreML. The project explains that earlier choices, running text-to-speech through MLX on the GPU or an ONNX-based streaming speech model, caused dropouts and hangs from contention with the language model, which an all-CoreML combination resolved along with reducing memory use.

It is written in Swift and SwiftUI, downloads about 2.3 GB of models on first launch with per-model progress, and is explicitly described as a work in progress.

Website: https://apps.apple.com/app/volocal/id6761493288