How accurate is speech recognition in Latvian?

Latvian is one of the harder languages for on-device speech recognition: roughly 1 word in 6 comes out wrong on read speech. In our October 2026 benchmark on a Mac with an M2 Pro, Parakeet TDT 0.6B v3 had a word error rate of 17.27 percent ¹ and Whisper large-v3-turbo had 18.64 percent ², on 100 Latvian recordings from the FLEURS test set. We have no Apple SpeechTranscriber result for Latvian.

The two engines are close in Latvian, and Parakeet is faster

Parakeet has the lower error rate, but the gap to Whisper is small, about one point. With only 100 recordings, treat the two as roughly equal on accuracy. Speed separates them more. Parakeet transcribed the recordings about 21 times faster than real time ³, and Whisper about 15 times ⁴. Both are fast enough for dictation. The speed dataset covers every language.

Our measurement of the engines in Latvian

Lower is better in this table.

The transcripts behind each row are on the on-device benchmark page.

Published scores are worse than ours on some tests

NVIDIA reports 22.84 percent for Parakeet on FLEURS Latvian ⁵, higher than our 17.27 percent. OpenAI's paper gives 23.1 percent for the older Whisper large-v2 ⁶. On CoVoST 2 Latvian, a test built from volunteer recordings, Parakeet had 38.36 percent ⁷. The spread shows how much the result depends on the recordings and the scoring.

All published rows are on the published accuracy page.

Latvian ranks twenty-second of 25 languages

Taking the best engine in each language, Latvian is twenty-second of the 25 languages in the benchmark. Only Lithuanian, Slovenian and Maltese are worse. Estonian, its northern neighbor, is clearly better at 14.17 percent ⁸, and Finnish is better still at 7.73 percent ⁹. Latvian is a small language with little training audio, and it has rich word endings, though the benchmark does not test which factor matters more.

Punctuation and capitals add errors you will see

Word error rate ignores punctuation and capital letters. A reader does not. Counting them, Parakeet's text differed from the written sentences by 18.62 errors per 100 tokens ¹⁰, against 20.35 for Whisper ¹¹. Whisper loses a little more once formatting counts, so the gap between the engines widens slightly in the text you actually read.

What to expect when you dictate in Latvian

Expect to correct about 1 word in 6 even in a quiet room. Dictation works as a fast first draft that you proofread, not as finished text. Choose Parakeet if your tool offers it, since it is a little more accurate and faster. Names, numbers and inflected endings are the likely trouble spots, and a wrong ending counts as a whole wrong word. Speak in short, clear phrases and fix punctuation afterward. We found no published fact that confirms Latvian in a built-in Mac or Windows dictation list, so check your system's language list, and read our overview of dictation languages before you choose a tool.

What these numbers do not show

The recordings are people reading prepared sentences in a quiet room, scored on one Mac with 100 recordings. Spontaneous speech, dialects, noise and specialist vocabulary are not in the test, and we did not measure Apple's model or any cloud service ourselves. Real dictation will usually score worse than read speech. For the measure itself, see the guide on how accurate speech recognition is.

Our guide to European languages in speech to text helps you choose an engine, and the languages overview covers other languages. For comparison, read the same question for Lithuanian, Estonian and Slovenian.