Serbian transcription, in the right alphabet.
Serbian is the clearest example of why published accuracy figures matter: the gap between engines here is not a few percent, it is the difference between a usable transcript and an unusable one.
4.4%
Canto AI
92.2%
Industry-standard engine
21× fewer errors · word error rate on FLEURS, lower is better
How Canto handles this language
Serbian is written in both Cyrillic and Latin, and that is where most engines come apart: they transcribe into whichever script dominated their training data, so the benchmark counts every affected word as an error even when the sounds were recognised correctly. Canto keeps the script consistent instead of switching midway, which is what turns the measured gap below into a practical one.
Competing engines frequently transcribe this language in the wrong alphabet, which the benchmark counts as errors on every affected word.
What the benchmark measures
Word error rate counts every insertion, deletion and substitution against a human reference transcript, then divides by the number of words spoken. Lower is better, and 5% means roughly one wrong word in twenty. We measure on FLEURS, a public benchmark of read speech across many languages, so anyone can repeat the test rather than take our word for it.
How we measure
The same audio goes through both engines with no per-language tuning, and the output is scored with the same normalisation on both sides. We publish the competing engine's figure alongside ours rather than only our own, because a number with nothing to compare it to says very little.
What you get
- Automatic language detection, including conversations that switch language mid-sentence
- Speakers separated automatically, so the transcript reads like a script
- AI titles, summaries and action items written in the language you choose
- Search and chat across everything you record, with answers that cite the moment
- A notetaker that joins Teams, Meet and Zoom, plus phone and watch recording in person
Questions
How accurate is Serbian transcription?
Canto measures 4.4% word error rate on the public FLEURS benchmark. The engine most other notetakers build on scores 92.2% on the same audio, largely because it transcribes into the wrong alphabet.
Does it use Cyrillic or Latin?
Canto transcribes Serbian in Latin script and keeps it consistent through the recording, rather than switching partway.
Can I get summaries and notes in Serbian?
Yes. The interface and all AI-written output ships in Serbian, and transcription covers 99 languages regardless of interface language.
Try it on your own audio.
300 minutes a month, free forever. The fastest way to check an accuracy claim is to run your own recording through it.