Fast In-Browser Speech Recognition via WebGPU and SIMD WebAssembly

A developer has released parakeet.wgsl, a dependency-free, local implementation of NVIDIA's Parakeet TDT 0.6B V2 speech recognition model running entirely in the browser. It leverages raw WebGPU compute shaders and SIMD WebAssembly to transcribe one hour of audio in just 20 seconds on modern hardware. This project demonstrates the feasibility of high-performance, private, and offline-capable Automatic Speech Recognition (ASR) directly on client devices without relying on cloud APIs. By compiling to WebGPU, it also paves the way for cross-platform GPU-accelerated transcription in native offline applications. The implementation is fully custom and dependency-free, using WebGPU Shading Language (WGSL) for neural network inference and SIMD WebAssembly for the audio processing frontend. It supports English transcription with punctuation, capitalization, and word-level timestamps.