Browser runtime proof Report artifact 2026-09-01

Parakeet WGSL Browser ASR

This is a strong browser-systems result and a deliberately incomplete model-quality claim. The path is local after its verified cache is present, warm transcripts are repeatable, and runtime is dramatically faster than real time. Accuracy was not scored against a reference, visible transcript errors remain, and the first download is 386.5 MiB.

Headline Numbers

First inference

7.225s 17m 10s audio · 142.6x real-time

Warm inference

7.68–7.80s 132–134x real-time

Verified cache

386.5 MiB 29 browser assets

Competitive Context

System Metric Score Size / Class Comparable? Readout
Parakeet WGSL warm browser run real-time factor 132–134x 1,030.49-second public MP3 Direct Three repeated transcripts were byte-identical.
WhisperKit Mac-native baseline WER and domain terms not measured future fresh comparison Not comparable No reference transcript was scored in this bounded smoke.

Direct rows share this artifact's eval setup. Directional rows are useful market context but should not be read as leaderboard claims.

What the smoke established

ClaimStateBoundary
Browser executionvalidatedChrome Apple/Metal FP16 subgroup path
Warm repeatabilityvalidatedidentical transcript hash
Local after cachesupportedmain-tab network observation
Accuracyunmeasuredvisible errors; no WER reference

Release Blockers

Accuracy not qualified

The smoke used no shared reference transcript or WER scorer and visible proper-name, phrase, and duplication errors remain.

Unblock: A WhisperKit comparison is a new owner-led experiment, not remaining historical work.

Large optional first download

The verified cache totals 405,252,493 bytes.

Unblock: Any product use must keep the download explicit, optional, and measured on Chrome and Safari.

Evidence

Next Release Action

Closed as a validated browser-runtime proof, not a selected ASR layer. Use it as the browser-audio lab case study.