# VoxrtAsr for iOS

A Swift package for streaming, on-device speech recognition on iOS, built on a custom Rust inference runtime (VoxRT) rather than ONNX Runtime or PyTorch Mobile, running a 32-million-parameter FastConformer model with roughly 90ms per-chunk latency on an iPhone 13 Pro Max.

VoxrtAsr is a Swift package that adds streaming, on-device automatic speech recognition to an iOS app, distributed through Swift Package Manager. It runs a NeMo FastConformer model (32 million parameters) on VoxRT, a from-scratch Rust inference runtime built specifically for streaming voice workloads on phone-class hardware rather than a general-purpose runtime such as ONNX Runtime, PyTorch Mobile or LiteRT; two sibling packages built on the same runtime and NEON kernel set cover voice-activity detection and always-on wake-phrase detection.

Both CTC and RNN-T decoding modes are supported, with RNN-T reporting a 3.27% word error rate on LibriSpeech test-clean against 4.90% for the CTC mode, at roughly 15% lower per-chunk compute cost. Measured on an iPhone 13 Pro Max, the real-time factor is 0.08 to 0.10, meaning about 90% of one CPU core stays free during live transcription, with around 150 MB of runtime memory.

The Swift wrapper is Apache-2.0 licensed; the compiled inference runtime itself is proprietary but redistributable through this package, and the underlying NeMo model weights are CC-BY-4.0.

Website: https://github.com/VoxRT/voxrt-asr-ios