<!-- agent-source-envelope -->
# Artifact: Parakeet WGSL Browser ASR | PTLLM

> Parakeet v3 scored 0/82 word errors and ran 3.51x faster than native WhisperKit on the same eight LibriSpeech clips, while missing a deliberately aggressive…

Canonical: https://posttrainllm.com/artifacts/parakeet-wgsl-browser-asr
<!-- /agent-source-envelope -->

Reference-scored browser ASR

Report artifact

2026-09-02

This closes the accuracy gap left by the original long-form smoke. On the frozen 34.365-second fixture, the official browser engine beat the native incumbent on WER, matched all four proper nouns, produced no repetition errors, and made no external warm requests. Its 33.84x median real-time factor is strong but below the preregistered 50x bar, so the formal decision remains reject rather than moving the goalposts.

WebGPU, browser, ASR, Apple Silicon, local first

## Headline Numbers

Evidence class applies to each value: measured = retained evaluation or
runtime result; historical = recorded result not freshly reproduced;
derived = calculated or decided from records; observed = current artifact
state; not-measured = an explicit evidence gap.

**Browser WER:** 0.0%

0/82 errors · eight test-clean clips

Evidence: measured

**Native WER:** 8.54%

7/82 errors · identical clips and references

Evidence: measured

**Paired latency:** 3.51x

914.7ms browser · 3211.8ms native

Evidence: measured

**Short-clip RTFx:** 33.84x

median · frozen gate required ≥50x

Evidence: measured

## Competitive Context

| System | Metric | Score | Size / Class | Comparable? | Readout |
| --- | --- | --- | --- | --- | --- |
| Parakeet TDT 0.6B v3 · browser | WER · total decode | 0.0% · 914.7ms | 684.4MB first download | Direct | Apple Metal WebGPU; zero warm external requests. |
| WhisperKit Mac-native baseline | WER · total decode | 8.54% · 3211.8ms | 1.5GB local Core ML model | Direct | Whisper large-v3 turbo; CPU and Neural Engine. |

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

## Frozen promotion gates

| Gate | Result | Decision |
| --- | --- | --- |
| WER delta ≤2 points | −8.54 points | pass |
| Median warm RTFx ≥50 | 33.84x | fail |
| No repetition regression | 0 vs 0 | pass |
| Warm external requests | 0 | pass |

## Release Blockers

### 50x short-clip gate missed

The 2–7 second clips expose an approximately 80–100ms browser dispatch floor; median throughput reached 33.84x despite beating native latency by 3.51x.

**Unblock:** Do not change this decision. A separately frozen batching or long-form latency experiment may test whether fixed overhead amortizes.

### Large optional first download

The pinned v3 browser model downloaded 684,385,250 bytes.

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

## Evidence

- [Paired result](https://github.com/PostTrainLLM/posttrainllm/blob/main/evals/verified-wins/parakeet-asr-result-v1.json)

- [Experiment protocol and interpretation](https://posttrainllm.com/docs/techniques/parakeet-wgsl-browser-smoke)

- [Frozen manifest](https://github.com/PostTrainLLM/posttrainllm/blob/main/evals/verified-wins/parakeet-asr-v1.json)

- [Complete experiment archive](https://posttrainllm.com/experiments)

## Next Release Action

Closed as a quality and native-latency win with a formal speed-threshold miss. Use it to teach paired evaluation, fixed-cost amortization, and why frozen gates matter.
