Free speech to text accuracy test: 2026 results

The recordings

We used EdAcc, the Edinburgh International Accents of English Corpus. It holds almost 40 hours of video call conversations between friends, transcribed word for word by professional transcribers, and a trained linguist labelled each speaker’s accent. The University of Edinburgh team published it for exactly this job, measuring speech recognition on accents, under a CC BY-SA 4.0 license.

From its test set we took five accent groups and about 10 minutes of speech from each, shared evenly between every speaker in the group, in the order they spoke:

The group names are EdAcc’s: Mainstream US English, Southern British English, Indian English, Nigerian English and Spanish. The Spanish group is English spoken by people whose first language is Spanish or Catalan. Some speakers in the American group grew up with another first language but were heard by the linguist as having a mainstream US accent.

We fixed the rules for picking clips before running any tool. We skipped clips the transcribers marked to be ignored, clips with two people talking at once or with words in another language, clips where people read out their participant codes, and clips shorter than one second or longer than 60 seconds. We removed sound labels such as laughter from the human transcripts. Each speaker’s clips were joined into one audio file with half a second of silence between them, and every tool got exactly the same files.

The tools and settings

• FreeTTS speech to text: the same upload the speech to text page makes, signed in, with the page defaults (Auto-detect language, Voice boost clean up). We then ran it again with English (US) chosen, and with English (UK) for the British speakers and English (India) for the Indian speakers.
• Whisper: OpenAI’s open source models tiny, base, small, medium, large-v3-turbo and large-v3, run with faster-whisper 1.2.1 in 8 bit mode on a 16 core server processor, beam size 5 (the default), language set to English. We then ran medium, large-v3-turbo and large-v3 a second time with the voice activity filter on and “condition on previous text” off.

How we counted

Word error rate is the standard measure: the words a tool got wrong, left out or added, divided by the words actually spoken. Before counting, both the tool’s text and the human transcript went through Whisper’s English text normaliser, which lowercases, removes punctuation, writes numbers and British spellings one way and drops “um”, “uh” and “hmm”. That way no tool loses points for “colour” against “color” or “25” against “twenty five”. We counted with the open source jiwer library and pooled the counts over every word in a group, so a long file counts more than a short one.

The EdAcc researchers saw the same kind of gap. In their paper, every model they tested did worse on Indian, Jamaican and Nigerian English speakers, and the best one got 2.7% of words wrong on clean US English but 19.7% on their conversations. Our Nigerian group fits that picture; our three Indian speakers did not, which is a good reminder that a handful of speakers can pull a group either way.

Missing words: mostly small ones, and fillers

Every tool left words out, and most of them were small words: “I”, “is”, “you”, “it”. FreeTTS left out 765 of the 8,800, more than Whisper large-v3 (619) and turbo (590) but fewer than Whisper small (901). About one in five of the FreeTTS misses (152) were repeats such as “I I” or “the the”. It also tidies filler phrases. Our speakers said “you know” 55 times; FreeTTS left it out 32 times, Whisper medium 12 times and Whisper large-v3 5 times.

For captions, notes and articles that makes FreeTTS text easier to read. For a word for word record, such as research interviews or legal notes, it matters, and our scoring counts every one of those as an error.

Extra words: Whisper adds more

The larger Whisper models added far more words that nobody said: 333 for large-v3-turbo and 265 for large-v3, against 104 for FreeTTS. Most are single words; a few are whole invented passages, which is the next section.

Misheard words: names and rare words

• Holyrood Park in Edinburgh came out as “Hollywood Park” in Whisper large-v3 and turbo. FreeTTS spelled it right.
• The Son of Neptune, a book title, became “Sunday Neptune” twice in FreeTTS.
• Accent became “assent” three times in Whisper large-v3 and “ascended” twice in Whisper medium.
• Dubai became “the bite” in Whisper base.

None of these tools knows the names of your places, products or people. Whatever you use, read names and numbers again before you publish.

It can invent text. In one American recording that every other tool got between 1.5% and 7.4% wrong, Whisper large-v3-turbo added 68 words that nobody said, including a Chinese character in the middle of an English sentence, and scored 36.1%. It also added a 36 word passage about photos to a Spanish accented recording. This is a known problem: a 2024 study by Koenecke and colleagues found that about 1% of Whisper transcriptions contained whole invented phrases or sentences.

It can get stuck in a loop.In one of our two British recordings, Whisper large-v3 wrote “but very soon after that I think we got cable” four times. That loop is part of the reason its British score was 17.8% while turbo managed 13.2%.

The fix.OpenAI’s own code notes that turning off “condition on previous text” makes the model “less prone to getting stuck in a failure loop, such as repetition looping”, and faster-whisper has a voice activity filter that skips stretches without speech. We ran medium, large-v3-turbo and large-v3 again with both settings:

With the fix, large-v3 and turbo became the two most accurate setups in the whole test. The catch is that you have to install Whisper, pick the model and set these options yourself, and the large models need a strong computer. In faster-whisper the settings are vad_filter=True and condition_on_previous_text=False; OpenAI’s own command line tool has the second one as --condition_on_previous_text False.