Fact · Languages and accents
6.79 to 10.64BLEU points
A 2026 study found that on the Vietnamese to English subset of FLEURS, translating recognizer output instead of clean transcripts cost between 6.79 and 10.64 BLEU.
Checked against its source on October 17, 2026
In the source's own words
“6.79 to 10.64 BLEU drops when translating ASR outputs instead of clean transcripts”
PiDA: Phonetically-Informed Data Augmentation for Robust Vietnamese Speech Translation, Giang Son Nguyen et al., arXiv, 2026, Introduction.
Open the original- What it measures
- The comparison is between translating the recognizer's transcript and translating a clean, human transcript of the same speech. The difference is the cost of the recognizer's errors.
- What it does not show
- This is one language pair, Vietnamese to English, on read speech. It shows the size of the effect, not the error rate of any app.
Source
Cited on
- Speech to Text With Translation on Phones and ComputersGuide
- How accurate is speech to text translation?Question
- How do you translate a voice note to text?Question
- How do you translate an audio file into another language?Question
- How do you voice type in Hindi and get English text?Question
- How do you voice type in Urdu?Question