Trained on Lebanese speech, not textbook Arabic
Lebanese Arabic,
written as it was said.
Upload a recording. Read it back with every word timed, the uncertain ones marked, and the conversation cut into turns. Online in the Studio, or on a server that never touches the internet.
No card. Your recordings stay yours and are never used for training without your say.
What comes back
A transcript you can check, not just read.
Every word carries its time and the engine's own confidence in it. You see where it was sure and where you should listen.
Word timings
Click any word and the audio plays from there. Follow the text while it plays; it follows you back.
Confidence marks
Words the engine is unsure of are highlighted — review goes where it matters, not everywhere.
Turns
Long recordings arrive cut into turns with timestamps, so a forty-minute call has handholds.
Exports
Plain text, SRT subtitles, or JSON with every word, time and confidence for your own tools.
Speaker identification is not part of the engine yet. The Studio says so rather than guessing.
Why general models fail here
Lebanese is not Modern Standard Arabic.
Ask a general model to transcribe a Beirut phone call and it writes textbook Arabic back at you — clean, confident and not what was said. Auralis is trained on the spoken language: podcasts, street interviews, and calls recorded in the field.
A general model hears
كل شيء غالٍ في السوق
Textbook Arabic. Nobody says this.
Auralis hears
كلشي غالي عالسوق
What was actually said.
How it works
Drop it in. Read it back.
- 01
Upload or record
Audio or video, up to an hour. Drop a file, pick one, or speak into the microphone.
- 02
Transcribe
Auralis Levant decodes it on our GPUs — or on your own server in the on-premise edition.
- 03
Read and check
Words light up as the audio plays. Uncertain ones are marked. Click to jump.
- 04
Export
Download TXT, SRT or JSON, or re-run it when a newer model ships.
Where it runs
In the Studio, or inside your building.
Studio
Online, in a minute.
Sign in, upload, read. Your recordings are stored under your account, served only to you, and deleted when you say.
- Files up to 600 MB and 60 minutes
- Word timings, confidence and turns
- TXT, SRT and JSON export
- Re-run on newer models as they ship
On-premise
Nothing leaves the network.
The same engine installed on your hardware, licensed to that machine, fully operable with the network unplugged. Built for evidence rooms where data sovereignty is not negotiable.
- Air-gapped by design — zero network calls
- Perpetual, hardware-bound licence
- Signed model packages, installed offline
Hear it written down.
Your first recordings are free. The on-premise edition is a conversation.