xcribe
checking…

Meeting transcription · 90+ languages

Transcribed twice.
Checked against the screen.

Two engines transcribe every recording. Every spoken figure is compared between them, and against whatever was on screen when it was said. Where they disagree, you are told — not handed one confident guess.

Tuned on Thai and Lao meetings. Runs on 90+ languages.

What comes back. Every figure checked between engines, and against the screen that was live.

merged.json · segments consecutive-unique 1.000
00:04:42S1 MOU มันแค่ 14 ร้าน ต่อสามรายการ
⚠ engines split ร้าน shops / ล้าน million · screen shows 14 → shops
00:04:51S0 ดูจากตัว Report View ก่อนครับ
00:04:58S1 ยอดรวมสองแสนห้าหมื่น
✓ 250,000 both engines agree · on screen 250,000
▣ 00:04:40 Stock on Hand · 2,000 rows status: read
ไทย ລາວ 한국어 Tiếng Việt ខ្មែរ မြန်မာ Bahasa English + 80 more

Both engines cover 90+ languages, and screen reading handles any script. The spoken-number cross-check is built for Thai and Lao today, with Korean, Japanese and Chinese next.

1.000 consecutive-unique the hallucination gate — a looping run scores 0.36–0.48
1,116 → 73 frames → screens perceptual-hash selection, threshold chosen off the curve
96.9 / 0.0 / 3.0 thai / lao / latin language mix measured per speaker before any pass is run
01 · screens Read, not guessed.

Distinct screens are selected by perceptual hash and read with the transcript that was live alongside them. Unreadable comes back as unreadable.

02 · record A ledger, not a wall of text.

Decisions, actions and figures, each with its timestamp, speaker and the verbatim words it came from. A short list of what to check before quoting.

03 · where it runs Your box, or the cloud. Your call, per meeting.

Run locally and nothing leaves the machine. Or send the audio out and have the result in minutes instead of hours.

04 · input Video, or just the audio.

Screen-share recordings, podcast-style audio from BigBlueButton, or a phone recording of a room. Upload and go.