Auto caption accuracy: measured in 6 languages

Updated November 2026

Auto caption accuracy on SubtitlePop, measured against human transcripts: between 89% and 98% of words right across English, Spanish, French, German, Arabic and Chinese.

Auto caption accuracy results by language

Tested on 9 November 2026. WER is word error rate; Chinese is scored by character (CER). Timing offset is the median gap between the moment a sentence starts in the audio and the start of its first caption word.

The spoken language was detected automatically and correctly in all six clips.

What these caption accuracy numbers mean for a video

A word error rate of 5.8% means about 6 words in every 100 differ from a careful human transcript. A 60-second short holds roughly 150 spoken words, so an English clip needs about 9 small corrections and a Spanish clip about 3. Each one is a click and a retype in the caption editor, and the timing of the other words stays where it is.

A timing offset of 0.16 to 0.31 seconds is under a third of a second: captions appear with the voice, and the word-by-word highlight follows the speaker.

How the auto caption accuracy test was run

1. Audio. Sentences from the test split of FLEURS, a public speech dataset from Google with a human-checked transcript for every recording. For each language, 10 to 17 recordings were joined into one clip of about two and a half minutes, with a 0.6-second pause between sentences.
2. Recognition. Each clip went through the same steps as an upload on SubtitlePop: the audio is cut into pieces of about 25 seconds at the quietest moment nearby, and each piece is transcribed by the Whisper large-v3-turbo model with word-level timestamps.
3. Scoring. The output was compared with the human transcript after lower-casing and removing punctuation. Word error rate is the number of substituted, missing and added words divided by the number of words in the human transcript. One recording per language was left out of scoring, so the figures cover the sentence counts in the table.

The scoring is strict about numbers: “25” in the captions against “twenty-five” in the transcript counts as an error, although both read correctly on screen.

Why SubtitlePop transcribes in 25-second pieces

The same English clip was also transcribed in one pass, as a single 2.7-minute file. The word error rate was 11.9%. Cut into 25-second pieces at quiet moments, the same audio scored 5.8% — half the errors. In one long pass the model added words during pauses and at the end of the clip; in short pieces those additions went away. Across the six languages, French, Arabic and Chinese improved as well, Spanish scored 1.9% both ways, and German went from 2.0% to 4.5%.

How to get the most accurate auto captions

• Record the voice close and clear. A phone held near the speaker or a clip-on microphone gives the recogniser the same kind of audio as this test.
• Set the language for mixed or very short clips. Automatic detection listens to the opening of the video; choosing the language yourself removes the guess.
• Fix names once. Brand and personal names are the most common misses. Find and replace in the editor corrects every occurrence at the same time.
• Read the captions against the video. Select Play on a caption to hear that moment, and correct the line in place.

Try it on your own audio with the auto subtitle generator, or turn the result into a file with the SRT generator.

Data source and licence

FLEURS: Conneau et al., “FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech”, Google Research. The dataset is published under the Creative Commons Attribution 4.0 licence (CC BY 4.0). The recordings were joined into longer clips for this test; the transcripts were used unchanged.