VoxrtAsr for Android
An Android library for streaming, on-device speech recognition, built on a custom Rust inference runtime and a 32M-parameter NeMo FastConformer model.
VoxrtAsr runs streaming speech recognition on VoxRT, a from-scratch Rust inference runtime built for phone-class hardware instead of ONNX Runtime, PyTorch Mobile or LiteRT. It uses a 32M-parameter NeMo FastConformer model with cache-aware streaming in roughly 1.1 second chunks, takes 16kHz mono audio, and outputs punctuated and capitalized text.
It ships arm64-v8a builds with NEON acceleration and an x86_64 emulator build, under Apache-2.0 for the Kotlin wrapper with a proprietary but redistributable compiled runtime and a CC-BY-4.0 upstream NeMo model. Sibling libraries on the same runtime and NEON kernels cover voice-activity detection and always-on wake-word detection.
Measured release-build real-time factors range from about 0.24 on a 2018 flagship SoC to 0.35 on a 2020 midrange SoC during live microphone transcription, leaving headroom for other on-device work.