High-performance inference of NVIDIA's Parakeet TDT 0.6B V2 English transcription model, in the browser.
A fully custom, dependancy-free implementation with raw WebGPU compute shaders and SIMD WebAssembly audio frontend.
1 hour of audio transcribed in 8.4 seconds (Apple M5 Max, Safari 26.6).
Scores 1.69% and 3.19% WER on LibriSpeech test-clean and test-other,
respectively, matching
NVIDIA's reported results
Try the live demo
npm install parakeet.wgslThe package is ESM-only and requires a WebGPU-capable browser in a secure context (HTTPS or localhost).
Save the following as index.html:
<!doctype html>
<html lang="en">
<head><title>parakeet.wgsl</title></head>
<body>
<input id="audio" type="file" />
<button id="transcribe" type="button">Transcribe</button>
<pre id="output" style="white-space: pre-wrap"></pre>
<script type="module">
import { createTranscriber } from "parakeet.wgsl";
const output = document.querySelector("#output");
const transcriber = createTranscriber({
onLoadProgress: ({ phase, fraction }) => {
if (phase === "weights") {
output.textContent = `Loading model... ${Math.round(100 * fraction)}%`;
}
},
});
document.querySelector("#transcribe").onclick = async () => {
output.textContent = "Transcribing...";
const result = await transcriber.transcribe(
document.querySelector("#audio").files[0],
{
onProgress: ({ fraction }) => {
output.textContent = `Transcribing... ${Math.round(100 * fraction)}%`;
},
},
);
output.textContent = `Transcription time: ${(
result.metrics.totalMs / 1000
).toFixed(2)} seconds\n\n${result.text}`;
};
</script>
</body>
</html>Run:
npx vite --openSee DOCS.md for the complete API, including progress and live transcripts, cancellation, model-cache management, timestamp rendering, and self-hosting.
The repository also includes a complete local demo that consumes the built package through the same boundary as an external application.
transcribe() accepts an audio File or Blob. Its canonical input format is
a 16 kHz mono signed PCM16 WAV.
Other supported inputs are decoded, downmixed, and resampled by the package's reduced, audio-only FFmpeg core. Supported inputs include:
- MP3
- AAC or ALAC in M4A, MP4, or CAF
- Opus or Vorbis in WebM, Matroska, or Ogg
- FLAC
- WAV, AIFF, AIFF-C, and CAF PCM
- WavPack
These inputs require the Origin Private File System (OPFS) and sufficient temporary origin storage for the decoded audio.
Model inference is implemented with custom WebGPU compute shaders. A purpose-built WebAssembly frontend performs filter-bank extraction.
Canonical WAV files stream directly into the inference frontend. For noncanonical inputs, audio decoding and model inference run in a pipelined fashion. FFmpeg decodes the recording into bounded, immutable PCM segments in OPFS, and each completed segment is made available to inference as soon as it is ready. Temporary segments are removed when the operation finishes.
WebGPU capability selection is automatic. An adapter with shader-f16 uses
the FP16 model; all others use FP32. Compatible fixed-width subgroups use the
subgroup kernels, while other supported adapters use portable workgroup
kernels. These decisions are fixed for a transcriber's lifetime, and only the
selected model package is downloaded.
Inference currently runs fastest in Chrome and Safari. Firefox is supported but substantially slower (at least on Mac), and performance in other browsers may vary. This is likely explained by differences in WebGPU implementations of browsers.
A cold initialization downloads approximately 405 MB for FP16 or 441 MB for FP32. The packages reach these sizes with a custom k-means palette format, informed by FluidInference's Möbius model-conversion work. Most model weights are stored as bit-packed 5- or 6-bit indices with precision-specific lookup tables.
Model weights are loaded incrementally into GPU buffers and verified before use. Cache Storage can retain verified model assets for later visits, while later transcriptions on the same transcriber reuse the GPU-resident model.
The long-form transcription policy is derived from the excellent FluidAudio project. It uses fixed inference windows, timestamp-aware overlap merging, adaptive speech-energy gating, and bounded seam repair.
The reproducible
LibriSpeech benchmark
evaluates the complete test-clean and test-other splits with a pinned
dataset and scorers.
pnpm install
pnpm test
pnpm build
pnpm verify:package
pnpm demoSee model/README.md to reproduce or mirror the canonical model packages.
Native audio component architecture and rebuild instructions are in src/cpp/README.md.
The project's source is MIT licensed. The separately hosted model packages are modified forms of NVIDIA Parakeet TDT 0.6B v2 under CC BY 4.0. Complete third-party attribution and license terms are in THIRD_PARTY_LICENSES.
Thank you to my friend Akki for running and sharing the result of the 1 hr audio benchmark on his Apple M5 MacBook.