parakeet.wgsl

Fast, accurate transcription in the browser, powered by WebGPU and WebAssembly.

This demo runs NVIDIA Parakeet TDT 0.6B v2 (English-only) entirely on-device. Audio never leaves the browser.

The model is downloaded when you transcribe for the first time, then cached for later visits.

Check out parakeet.wgsl on GitHub and npm for use in your own projects.

Pick an audio file (or ).

Checking browser support…

Runtime details

Checking WebGPU support…
Downloaded model

Checking downloaded model…

Acknowledgements & licenses

parakeet.wgsl is distributed under the MIT License.

NVIDIA Parakeet TDT 0.6B v2

NVIDIA licenses the original model under Creative Commons Attribution 4.0. Attribution and license details

FluidAudio

The long-form transcript policy, merge behavior, and recovery behavior are each derived from FluidAudio under Apache License 2.0. License text

ffmpeg.wasm and FFmpeg

Audio decoding uses the MIT-licensed ffmpeg.wasm fork and an audio-only FFmpeg build licensed under LGPL-2.1-or-later. ffmpeg.wasm license · FFmpeg license. The corresponding source, build, and relinking instructions are available in the pinned fork.

Supporting components

The distribution also incorporates Emscripten runtime and system-library portions and a build-time coremltools k-means helper. Emscripten notices · musl copyright and license · coremltools helper license

Kennedy sample audio

The optional sample is a public-domain White House Audio Collection recording cataloged by the John F. Kennedy Presidential Library and Museum.

Complete third-party notices, source references, and relinking instructions