Why a model trained on spoken Hebrew is more accurate
"Did he say fifty-nine, or fifty-nine thousand?" A transcription model picks the sequence of words that is most likely given the sounds, and what counts as likely depends on what it learned. A model trained mostly on English knows Hebrew the way a tourist does; a model trained on spoken Hebrew knows it like a local. Here is why that matters, and what we did about it.
How a transcription model picks its words
Every speech recognition system works roughly the same way: it takes a stretch of audio and searches for the word sequence that best explains both the sound and the words already written. That is the method, and it is why transcripts read like language. The difference between systems is what taught them which sequences are likely.
A model that learned from spoken Hebrew knows how real conversations sound: swallowed endings, names, numbers, English terms inside Hebrew sentences, and people finishing each other's sentences. On a clean news reading every model looks good; in a real team meeting, that knowledge is what makes the transcript right.
What spoken Hebrew asks of a model
- Swallowed endings. Spoken Hebrew compresses word endings that written Hebrew keeps.
- Names. People, companies and products, which a Hebrew-trained model meets every day.
- Numbers. Hebrew numerals carry gender agreement, sometimes in a single unstressed syllable.
- Code switching. A Hebrew sentence with three English words inside it is normal here.
- Lively conversation. Interruptions and quick back-and-forth, the way meetings really go.
What we did
We trained our own transcription model on spoken Hebrew, and we keep improving it. Each new version is compared against the one in service on the same recordings of real meetings, and only takes over if it is more accurate. The biggest gains in recent comparisons came in the most challenging audio — room echo, background noise, fast talk — which is to say, in real meetings. More in the Hebrew model behind IvreetMeet.
See for yourself
Upload your most challenging recording and compare. That is the test that matters. The free plan covers 7 meetings and 120 minutes a month.
Frequently asked questions
What is the difference between a general model and a Hebrew-trained one?
A general model learns dozens of languages, mostly English. A model trained on spoken Hebrew knows the slang, the prefixes and the switch to English mid-sentence, so it transcribes Hebrew meetings and calls more accurately.
Do mixed Hebrew and English sentences work?
Yes. Hebrew sentences with English product names and professional terms are how people actually speak at work here, and that mixture is part of what the model is trained on.
How do you know a new version is better?
Each candidate is compared against the version in service, on the same recordings of real meetings, and only replaces it if it is more accurate.
Try it on your own recording
IvreetMeet transcribes and summarizes Hebrew meetings. The free plan covers 7 meetings and 120 minutes per organization each month.
Start free