# Whisper Web transcription samples

These short recordings let you check the local workflow before selecting your own media. Listen to the audio, select the matching spoken language, transcribe it and compare the result with the dataset reference.

| Sample | Duration | Audio | Reference |
| --- | ---: | --- | --- |
| English | 10.56 seconds | [WAV](https://whisperwebfree.com/samples/en.wav) | [Reference text](https://whisperwebfree.com/samples/en-reference.txt) |
| Spanish | 12.84 seconds | [WAV](https://whisperwebfree.com/samples/es.wav) | [Reference text](https://whisperwebfree.com/samples/es-reference.txt) |
| Arabic | 8.16 seconds | [WAV](https://whisperwebfree.com/samples/ar.wav) | [Reference text](https://whisperwebfree.com/samples/ar-reference.txt) |

Each localized tool lists its matching sample first. Selecting a sample also sets the spoken language. All three recordings are 16 kHz, mono, 32-bit float WAV files. The recorded outputs below are separate from the dataset references.

The reference files are dataset transcripts, not guaranteed Whisper output. The clips contain read speech; they do not represent noisy meetings, overlapping speakers or a one-hour recording. Review names, numbers and timestamps before using any transcript.

## Source and license

Audio and reference text: **FLEURS contributors / Google**, from the [FLEURS dataset](https://huggingface.co/datasets/google/fleurs), under [Creative Commons Attribution 4.0](https://creativecommons.org/licenses/by/4.0/).

The English sample is `en_us`, test split, row ID `1904`, original file `1003119935936341070.wav`. The Spanish sample is `es_419`, test split, row ID `2001`, original file `1001146817223348054.wav`. The Arabic sample is `ar_eg`, test split, row ID `1993`, original file `10019390257075855641.wav`. Original WAV recordings and reference text are unchanged; all three use the same CC BY 4.0 license.

The [manifest](https://whisperwebfree.com/samples/manifest.json) records source archives, filenames, duration, byte size and SHA-256 checksums. Sample preparation does not use visitors' recordings or transcripts.

## Local workflow

- [Standard transcription](https://whisperwebfree.com/): up to 300 MiB and 20 minutes.
- [Large-file transcription](https://whisperwebfree.com/large-file-transcription): local files up to 1 GiB and 1 hour, processed in sections.
- [Transcribir audios largos](https://whisperwebfree.com/es/transcribir-audios-largos): the Spanish large-file tool.
- [Arabic transcription](https://whisperwebfree.com/ar): load the Arabic sample and compare the result with its reference.

The browser still downloads the application and model. Selected local media is processed on the current device. Keep the tab open during processing and export results you need to keep. Sample attempts are identified separately in optional usage measurements.

## Observed browser tests — 3 October 2026

Chromium 154 on macOS, Whisper Tiny (q8), WebAssembly. The English clip completed in the standard tool; the Spanish clip completed in the large-file tool. Both WAV inputs are 32-bit float, 16 kHz, mono. Both results were stored in local history, and the TXT/SRT/VTT export contents were inspected.

| Sample | Actual text | Actual subtitles |
| --- | --- | --- |
| English | [TXT](https://whisperwebfree.com/samples/en-output.txt) | [SRT](https://whisperwebfree.com/samples/en-output.srt) · [VTT](https://whisperwebfree.com/samples/en-output.vtt) |
| Spanish | [TXT](https://whisperwebfree.com/samples/es-output.txt) | [SRT](https://whisperwebfree.com/samples/es-output.srt) · [VTT](https://whisperwebfree.com/samples/es-output.vtt) |

These are observed model outputs, including recognition mistakes. Compare them with the dataset references above. Results may change with model, backend or settings. This is a reproducible workflow check, not an accuracy or performance benchmark.

A separate 21-minute-1-second synthetic recording, made from the public English clip repeated with silence, completed in five sections. A second run paused after section 1, resumed at section 2 and completed. SRT/VTT exports contained 144 cues with the final cue ending at 20:54. This input was selected through a developer-injected File in a local preview because the native file chooser was unavailable. The synthetic fixture is not a real long meeting or an accuracy benchmark. These tests do not prove the full 1 GiB / 1 hour limit. Safari, Firefox, mobile devices, WebGPU and other codecs are not covered. Operating-system download completion was not verified. [Machine-readable test notes](https://whisperwebfree.com/samples/browser-tests.json).

## Observed Arabic test — 11 November 2026

The 8.16-second Arabic sample completed in the standard tool on a local development build. It was selected through the sample button on `/ar`, with Arabic and the original-language transcription task selected. The run used Whisper Tiny Q8 / WebAssembly in the Codex in-app browser on Apple M4, 16 GB RAM and macOS 15.7.3. The browser's Chromium version was not recorded.

| Actual text | Actual subtitles | Full exported record |
| --- | --- | --- |
| [TXT](https://whisperwebfree.com/samples/ar-output.txt) | [SRT](https://whisperwebfree.com/samples/ar-output.srt) · [VTT](https://whisperwebfree.com/samples/ar-output.vtt) | [JSON](https://whisperwebfree.com/samples/ar-output.json) |

These files were downloaded through the application's export menu, verified on disk and copied here unchanged. The result has two subtitle cues; the last ends at 7.48 seconds. It contains substantial recognition errors, including the Latin text `2a`. Compare the raw output with the [dataset reference](https://whisperwebfree.com/samples/ar-reference.txt) and replay the audio before using any wording.

This checks that this short sample can produce downloadable files in that browser. It is not an accuracy or speed benchmark, and does not establish support quality for Arabic dialects, noisy conversations, WebGPU, other browsers, mobile devices or long recordings. [Machine-readable test notes and output checksums](https://whisperwebfree.com/samples/browser-tests-2026-11-11.json).

## Large-file timing

A separate October 20, 2026 run processed a 23-minute 6-second MP4 in 3 minutes 24 seconds on Apple M4, 16 GB RAM, macOS 15.7.3 and Chrome 155, using Whisper Tiny Q8 / WebAssembly with English selected. That time includes decoding and local saving, but excludes the first model download. The input joined two MP4 files from the same source.

The large-file tool's 20–60 minute waiting-time table scales that single result by recording length and rounds up to minutes. Those rows are estimates, not independent tests. The original MP4 and its transcript are not included in this public sample set, so the files above cannot reproduce that timing result. See the [recorded test conditions and limits](https://whisperwebfree.com/blog/how-to-transcribe-large-audio-files#tested-example).

## Compare backends

For a repeatable check on your own device, follow the [same-recording comparison steps](https://whisperwebfree.com/blog/webgpu-vs-webassembly-whisper#compare-backends). Whisper Web uses Q8 on WASM and FP16 on WebGPU, so the comparison covers two product configurations rather than the backend alone. Record the actual backend: a failed WebGPU initialization can fall back to WASM. The tests on this page do not establish a WebGPU speed advantage.
