Fact · Languages and accents
The 2023 Interspeech paper says that most speech corpora and Arabic ASR models are undiacritized, because diacritics are often omitted.
Checked against its source on October 17, 2026
In the source's own words
“so most speech corpora and ASR models are undiacritized”
Diacritic Recognition Performance in Arabic ASR, Hanan Aldarmaki and Ahmad Ghannam, Interspeech 2023, 2023, Abstract.
Open the original- What it measures
- The abstract explains that the vowel marks are often left out of the transcripts used to train speech models, so the models learn to output text without them.
- What it does not show
- The authors describe the typical training data and the typical model output. Individual products may do better; the paper does not test dictation apps.