How accurate is speech to text translation?

Speech to text translation is only as accurate as its weakest step, and errors in the first step carry into the second. Accuracy is not one number. It has a recognition part, which measures whether the words were heard correctly, and a translation part, which measures whether the translated sentence means the same thing. Research on each part gives different figures, and both depend on the language pair and the kind of speech.

Errors pass from the first step into the second

Speech translation usually runs in two steps. A recognizer turns speech into text, and a translator turns that text into another language. The translator sees only the transcript, so it cannot tell a misheard word from a correct one. Researchers call this error propagation, and a 2022 study describes it as a fundamental challenge for cascaded speech translation that still remains ¹.

A 2026 study on Vietnamese speech to English text found that most substitution errors came from sound-alike confusions, and that these phonetic errors significantly degrade translation quality ². So one word that sounds like another can change the translation.

Recognition accuracy depends on the language

Recognition is the first place to look. In Whisper's tests on the FLEURS read-speech benchmark, the large-v2 model had a word error rate of 4.2 percent for English ³. For Swahili, the same model had a word error rate of 39.3 percent ⁴. Word error rate counts substitutions, deletions and insertions as a share of the words in the reference, so a higher number means more of the transcript is wrong.

The gap is large, and it does not follow one rule. These numbers come from read speech in a test set, so do not expect the same numbers for your own voice and room. For the measure itself, see word error rate.

Translation scores are measured on a different scale

Translation quality is usually reported in BLEU, an automatic score that compares machine output with reference translations. It is a comparison score, not a percentage of correct sentences, so do not read it that way.

Whisper's paper reports a zero-shot score of 29.1 BLEU on the CoVoST2 speech to English translation benchmark ⁵. That is one benchmark with its own languages, so it does not tell you how well your pair will work. The effect of recognition errors on translation is also measurable. On the Vietnamese to English subset of FLEURS, translating the recognizer's output instead of clean transcripts cost between 6.79 and 10.64 BLEU points ⁶. Recognition errors alone can therefore take a large bite out of the translation score.

It depends on the speech, the language and the tool

Accuracy changes with the kind of speech you use. A 2024 Urdu benchmark found that Seamless-large was best on read speech, while Whisper-large was best on conversational speech ⁷. A tool that wins a read-speech test can lose a real conversation, so match the test to your use.

The tool matters as well. Compare tools on the same sentences in your language, and do not rely on a benchmark that used different speech.

Language support also varies. A tool may recognize a language but translate only some pairs. Check both the recognition support and the translation direction you need, and do not assume one implies the other.

What to do before you rely on a translation

1. Test your exact pair. Speak a few sentences in the language you will use, with the app you will use, and read the transcript before you read the translation.
2. Fix the transcript first. A wrong word in the transcript becomes a wrong word in the translation, so correct recognition errors before you translate again.
3. Keep turns short. Short sentences with a pause at the end give the recognizer less to get wrong. For how this works in one app, see the voice input steps for Google Translate.
4. Check names, numbers and instructions with a speaker of the target language. These are the words where one sound-alike error causes the most harm.
5. Use a person for anything important. Medical, legal or safety text needs a qualified human check, not a machine result alone.

The guide to speech to text with translation covers the two-step design in more detail. For the wider set of language questions, see the guide to dictation languages. If a term is new to you, the entry on speech translation explains the design, and the entry on low-resource language explains why some languages score much lower.