burncaps-app's picture
Upload 2 files
f56090e verified
|
Raw History Blame Contribute Delete
1.41 kB
metadata
license: cc-by-4.0
pretty_name: Speech service errors on FLEURS, 12 Indian languages
language:
  - hi
  - ur
  - bn
  - ta
  - te
  - gu
  - mr
  - kn
  - ml
  - pa
  - ne
  - or
task_categories:
  - automatic-speech-recognition
size_categories:
  - n<1K

Speech service errors on FLEURS, 12 Indian languages

Speech service errors on 240 FLEURS clips in 12 Indian languages. Each row is one Google FLEURS test clip (20 per language: Hindi, Urdu, Bengali, Tamil, Telugu, Gujarati, Marathi, Kannada, Malayalam, Punjabi, Nepali, Odia), the reference sentence, the text a speech service heard (Microsoft MAI-Transcribe through the Burncaps speech service, North Europe, measured 16 October 2026), the language it found by itself, and the words and letters it got wrong. Scoring: edit distance after lower-casing and removing brackets and punctuation, so a number written as a word or a joined compound counts as wrong. Read speech, not casual talk; FLEURS may be in a vendor's training data. Mean of the twelve languages: 17.0 words wrong in 100 and 5.1 letters wrong in 100; language found correctly in 240 of 240 clips. The method, the table and the on-device comparison: https://burncaps.app/indian-language-speech-to-text-accuracy. Source audio and sentences: Google FLEURS (https://huggingface.co/datasets/google/fleurs), CC-BY-4.0. Published by Burncaps (Jordan Labs Ltd, London, company no. 17358646), hello@burncaps.app.