Sample transcripts, in six languages

Six short recordings in six languages, each run through the free preview and shown exactly as it came back, mistakes included. Every recording is openly licensed (CC0 or public domain) and the source is stated on its page. Taken 2026-10-27.

The error rate is for that one clip against the sentences the speaker was asked to read, after lowercasing and removing punctuation (Arabic also unifies spelling variants). It says nothing about how the tool performs on your recording, and a spelling variant that is acceptable in the language still counts as an error, so the true figure can be lower.

How to read the error rates

• ·The words of the transcript are compared with the sentences the speaker was asked to read, after lowercasing and removing punctuation. Arabic also drops diacritics and unifies alef, yeh and teh marbuta, the same function as our benchmark. Turkish also gets the Turkish rules for I and dotted I, and the other languages get no language-specific spelling rules.
• ·A spelling variant that is acceptable in the language still counts as a different word, and so does a number written as a digit in one text and as a word in the other, so the real error can be lower than the rate shown. Check any pair yourself with the WER calculator.
• ·Each rate is for one short clip of one speaker, so it shows the kind of mistakes made, not how the tool performs on your recording.

How each sample was made

• ·Common Voice clips were trimmed of leading and trailing silence, joined with 0.45 second pauses, loudness-normalised and saved as 40 kbps mono mp3; the Spanish recording was cut from the chapter and processed the same way.
• ·The audio was uploaded to a local TryScriptorium server as a guest with the language set, as the language page does, and ran on the small model the free preview uses. The segments and times on each page are what the job returned.
• ·Nothing was typed or corrected by hand. The scripts that build these pages are in the project's scripts/samples folder.

How the samples were chosen

• ·Common Voice (CC0) for Hindi, Urdu, Turkish, Indonesian and Arabic: one volunteer speaker per language, chosen by a fixed rule (most clips available, clips of 2.5 to 7 seconds, no down-votes), not by how well the engine did.
• ·Spanish is a LibriVox public-domain audiobook, cut at the first pauses; Spanish clips in Common Voice's test file come from many speakers with only a few clips each, so there was no single speaker to use.
• ·Each recording went through the free preview (the small model) with the language set, as on that language's page. Nothing was edited afterwards.
• ·Languages we could not source a clean, openly licensed recording for are not shown. The measured Arabic and English results are on the measurements page.

Have your own recording? Transcribe a file, or check a transcript against a reference with the WER calculator.