logo

Speech & Voice

Your Arabic voice channel, finally searchable and automatable

You cannot search what was said, and you cannot automate what you say back.

Your Arabic voice channel is a black box: you cannot search what was said, and you cannot automate what you say back. Contact-centre and meeting audio pile up unprocessed because the tools that transcribe and generate Arabic speech were built for other languages first.

Delivered by

Sada turns meeting audio into decisions, actions and risks on-premises, understanding both directions of the voice channel - transcription and generation - in one place. Our speech research also produced Nabr, a Quranic recitation model analysing 8 acoustic properties per sound at 0.59% character error rate - the level our acoustic modelling operates at.

Why the channel, not the technology

  • Both directions in one place

    Understanding audio and generating it are one procurement for a contact centre or a CX team, so they are one solution here rather than two product listings.

  • Speech models built at the phoneme level

    Our acoustic modelling is trained on Arabic sound, not fine-tuned from someone else's English model, which is what makes dialect and recitation tractable at all.

  • The most sensitive audio an organisation produces

    Board minutes, HR investigations and legal strategy are exactly the recordings a cloud-only tool cannot take, which is why Sada deploys on-premises.

What it does with the channel

  • Meeting audio into decisions, actions and risks

    Sada runs multi-agent analysis over the transcript, so a meeting ends with a structured record rather than a wall of text.

  • Arabic transcription that handles dialect

    Speech recognition trained for Arabic rather than adapted to it, so the transcript is usable input for search and for analysis.

  • Voice generation with a real regional accent

    Sawt Najd is a Najdi Arabic TTS model trained on 41 hours of Najdi broadcast speech with a fixed male persona: a Saudi voice, not an MSA model approximating one.

  • Diacritization before synthesis

    Sadeed restores Tashkeel so a TTS system receives unambiguous text instead of guessing which word it is looking at.

Depth of capability

  • 0.59%Character error rate for Nabr, our Quranic recitation model, at the phoneme level
  • 8Acoustic properties analysed per sound, which is the level our acoustic modelling operates at
  • 22.4%Sawt Najd intelligibility WER: 9.7 points better than the human recording floor

What changes

  • The voice channel becomes searchable

    Calls and meetings join the rest of the organisation's searchable record instead of piling up as unprocessed audio.

  • Meetings end with a record, not a recording

    Decisions, actions and risks are written down as structure, which is the part anyone actually needs afterwards.

  • Sensitive audio never leaves the building

    The recordings an organisation is least able to send to a cloud service are the ones this handles on-premises.

How it deploys

  • On-premises

    Sada runs meeting intelligence entirely inside your environment.

  • Cloud API

    Speech models through Kawn Console for teams building voice features into their own applications.

  • Dedicated in-Kingdom

    Hosted inside Saudi Arabia where residency is required but running the hardware is not practical.

Let's talk about what you're trying to build.

Tell us the problem. We'll tell you honestly whether AI is the right answer.

Speech & Voice | Misraj AI | Next-Gen Arabic AI Lab