> Canonical page: https://posttrainllm.com/inspiration/parakeet-wgsl

A note of thanks · Browser runtimes

# parakeet.wgsl

Public kernels, a package format, and a reproducible browser run made parakeet.wgsl a concrete runtime lesson. We verified execution locally without treating it as an accuracy win or integration decision.

01 · The idea

## What stayed with us

Raw WebGPU plus SIMD-WASM can run Parakeet TDT browser inference with public kernels, a package format, cache behavior, a converter, and a benchmark harness.

02 · The local translation

## What we did with it

The pinned package transcribed a 17 minute 10 second public fixture in 7.2 to 7.8 seconds on Apple WebGPU. Warm output was deterministic and no external request was observed.

03 · The boundary

## Where the comparison stops

This validates reproducible browser execution, not transcription accuracy against a shared reference or readiness for default product integration. The 386.5 MiB package remains material.

Source trail

## Follow the work

- Original project [parakeet.wgsl source ↗](https://github.com/narcotic-sh/parakeet.wgsl)
- Our evidence · validated proof [parakeet.wgsl ↗](https://posttrainllm.com/studies/parakeet-wgsl)

Independent appreciation. The named projects have not endorsed or affiliated with PostTrainLLM.