Key Info
A user tested Mistral's new Voxtral transcription model on a noisy 1891 French phonograph recording of Gustave Eiffel, matching 100 of 107 words in INA's published excerpts (6.5% WER) while also transcribing 75 additional words omitted from the reference.
Highlights
- Test audio: a 135-year-old, noisy French phonograph recording of Gustave Eiffel from 1891
- Result: 100 of 107 words matched against INA's published excerpts, a 6.5% word error rate
- The model transcribed 75 additional words that the reference text left out
- It handled Eiffel's dictation well but struggled with proper names and some guest dialogue
- The test was run by a third party, not by Mistral itself