whisper-browser-benchmarks

Measurements from a Whisper transcription pipeline running entirely inside a browser tab: which audio and video containers the browser will actually decode, how accurate the smallest usable Whisper size is on clean synthetic speech, how long transcription takes relative to the length of the clip, what the first load pulls over the wire, and what happens to clips longer than the model's 30-second window.

Everything here was measured, not quoted from a model card. The raw records the tables are built from are in data/raw/, and the audio the measurements were taken on is in data/audio/.

What is in here

Path	What it is
data/decode-matrix.csv	Nine files fed to the browser decode path — which ones decoded, at what rate, how many channels, and what they weighed
data/speed.csv	Five transcription runs with wall time and realtime ratio, including two that failed
data/model-download.csv	The ten transfers that make up a cold first load, byte-exact
data/accuracy-word-level.csv	Word-by-word source-vs-transcript comparison for a 46-word clip
data/chunking-params.md	Chunking parameters plus a 98-second clip run twice, once with the overlap removed
data/audio/	Both WAV clips the measurements were taken on
data/raw/	The untouched measurement records, one JSON per run
COLUMNS.md	Field-by-field dictionary for every CSV
scripts/	The scripts that produced the audio and that regenerate the tables

Fields

Full field-by-field documentation is in COLUMNS.md. Two things worth knowing before you read the numbers:

• reported_rate_hz in the decode matrix is what the decode path reported, not necessarily what is stored in the file. The 8 kHz test file reports 44100 because that is what it was resampled to on the way through.
• The two failed runs in speed.csv are left in on purpose. Taking them out would make the table tidier than the session actually was.

Reproducing it

The tables are generated from the raw records, not maintained by hand:

python scripts/export_tables.py

That reads data/raw/*.json and rewrites every CSV in data/. It needs only the Python standard library.

The audio can be regenerated with scripts/make_accuracy_clip.py and scripts/make_chunking_clip.py. Both drive the Windows SAPI5 speech synthesiser through comtypes, so they are Windows-only, and the exact voice available on your machine will differ — which means a re-run will produce different audio and therefore different accuracy numbers. The clips committed in data/audio/ are the ones these measurements were actually taken on.

To re-measure from scratch, drive a browser at a page that runs Xenova/whisper-base through transformers.js v3 and capture what it returns.

Where the numbers came from

Every measurement was taken by driving headless Chrome 153.0.0.0 (64-bit) against a live page and recording what came back — no numbers were estimated and none were copied from documentation.

Date	2026-09-15
OS	Windows 11 (build 28000)
CPU cores seen by the browser	2
Browser	headless Chrome 153.0.0.0, 64-bit
Model	Xenova/whisper-base, ONNX int8
Runtime	transformers.js v3 on ONNX Runtime Web (WebAssembly)

Those are the limits of the numbers. A modern multi-core machine will beat the timings here comfortably; the decode matrix is browser- and platform-specific, so a phone or a different browser will accept a different set of containers.

Headline results, for anyone who wants them without opening the CSVs:

• Decoding: 7 of the 9 test files decoded. A fake .mp4 containing plain text and a 0-byte file were both rejected, which is the useful part — a wrong extension fails loudly rather than producing garbage.
• Accuracy: on a clean 20.8-second clip of synthetic speech — 46 words, no background noise — 44 of the 46 words came back exactly as spoken. Two came back different: review → reviews, and the spoken ten → the digit 10.
• Speed: between 0.49× and 0.77× real time on this two-core test machine, i.e. slower than real time throughout.
• First load: 78.2 MB, of which 73 MB is the two ONNX weight files.
• Long clips: a 98-second clip with 12 uniquely numbered sentences came back as 21 timestamped segments in order. A short fragment on a seam surfaced three times; running the same clip with the chunk overlap removed produced the same fragment the same number of times, so the overlap is not the cause.

The site these were measured on

These measurements were taken on the free browser-based transcription tool at https://www.ruanjiange.com/, which runs this pipeline client-side so files never leave the device.

Licence

Code in scripts/ is MIT — see LICENSE.

Everything under data/ — the tables, the raw records and the audio — is released under CC0 1.0 Universal (public domain dedication) — see data/LICENSE. Cite it, remix it, disagree with it; no attribution required.