The Hebrew model behind IvreetMeet

Behind every upload to IvreetMeet is our own speech model, trained on spoken Hebrew, with speaker separation beside it and a structured summary after it. This post explains what that gives you in practice: a transcript that understands how people speak in Israel, and a summary ready minutes after the meeting.

Our own model, for Hebrew as it is spoken

A real meeting is not dictation: people interrupt each other, two speak at once, English terms appear mid-sentence, and names and numbers fly by. That is exactly what IvreetMeet's model was trained on, and it keeps improving: every new version is compared against the one in service on the same real meeting recordings, and replaces it only if it is more accurate. On public Hebrew test sets it ranks among the leading systems; the table is on the Hebrew benchmark page.

What happens when you upload

• Hebrew transcription. Every word, with a timestamp, including English terms as they were said.
• Speaker separation. The system notices when the speaker changes; you name the speakers, and the names appear in the transcript and the summary.
• Results in minutes. An hour-long meeting usually comes back within minutes, and a two-hour one is quick too.
• Yours. The recording and the transcript stay in your account and are not used to train models.

The summary

The summary is written from the full transcript in a fixed structure: key points, decisions, and action items with owners. You can attach instructions and ask for a new summary with a different emphasis, without transcribing again. More in How an AI meeting summary is written.

Why it matters to you

• Accuracy: a model built for spoken Hebrew and tested on real meetings, not only on read speech.
• Time: transcript, speakers and summary in minutes instead of days of manual transcription.
• Everything in one place: search across meetings, exports to Word and PDF, and sharing by link.

Try it on your most challenging recording: the free plan covers 7 meetings and 120 minutes a month.