AI Platform · 2026
Nemo Speech Rs
Tech stack
Rust, OpenBLAS
Role
Design & Development
01. About the project
nemo-speech-rs is a high performance Rust library engineered specifically for CPU optimized inference of NVIDIA NeMo Automatic Speech Recognition (ASR) models. I built the entire inference engine from scratch, including a custom Tensor framework utilizing OpenBLAS and GEMM for accelerated CPU operations.
02. What I built
Engineered a complete ASR inference engine in Rust from scratch, starting from audio signal preprocessing and normalization, then building the custom Tensor abstraction layer using OpenBLAS for CPU GEMM acceleration. I implemented the full FastConformer ASR architecture including the convolutional encoder, multi-head self-attention (MHSA), and add TDT, CTC, RNN-T decoders support.
03. The result
Delivered a highly optimized, lightweight compiled library with a binary footprint of less than 10MB. This is drastically smaller compared to the official NVIDIA NeMo speech toolkit, which requires the massive PyTorch library and can take hundreds of megabytes.