The voice capability of Saj Sense
Voice AI that never leaves the device.
Transcription, speech synthesis, wake words and more. Every engine, from wake-word detection to speech synthesis, runs on the device with zero cloud dependency. Arabic, Hindi and Urdu are built into the foundation alongside English.
Built in Rust by SGH WLL, with hundreds of tests passing and a filed patent portfolio.
One wake word per language. Pick one and follow its thread through the pipeline.
saj-wakeUnder 1 ms inference, model under 1 MBRuns in Englishsaj-detectRuns in Englishsaj-listenRuns in Englishsaj-intentRuns in Englishsaj-speak-ttsRuns in EnglishA thread goes through the eye of every stage that runs in its language and passes under the rest.
Voice AI that respects your users.
Private from the first word
Your voice never leaves your device. There is no cloud processing, no data retention and no telemetry. GDPR, HIPAA and NESA compliance comes from the architecture itself.
Arabic from the foundation
Arabic, Hindi and Urdu are built into the foundation alongside English. None of them was added later as a translation layer.
Every platform
iOS, Android, Linux, Windows, macOS, Raspberry Pi, WASM and MCU. A Rust core with native bindings lets you deploy to phones, speakers, embedded devices and browsers.
Every engine runs on the device.
There is no cloud round-trip, so the voice features you ship keep working offline, everywhere. Each engine is its own Rust crate, and between them they carry 70+ trained models.
Listen
saj-wakeCustom keyword detection on a lightweight neural architecture, with models under 1 MB.
saj-detectNeural voice activity detection with sub-millisecond latency.
saj-listenReal-time transcription.
saj-scribeFile transcription with a split encoder and decoder.
saj-intentDirect intent extraction that turns voice commands into actions and skips the text.
Speak
saj-speak-ttsNatural synthesis on the device.
saj-cloneClones a voice from samples, on the device.
Who is speaking
saj-voiceSpeaker embeddings for verification and identification.
saj-whoReal-time speaker segmentation that tells you who spoke when.
saj-cleanLightweight neural noise removal.
Connect
saj-pipelineOne voice interaction that runs wake word, voice activity, speech-to-text, intent and text-to-speech in a single line.
saj-codecNeural speech codec with quality profiles from 1.2 to 6.0 kbps, real-time on CPU.
saj-senseUnified token streams with per-modality encryption.
saj-linkEnd-to-end encrypted messaging with sealed sender.
saj-serverREST and WebSocket with an OpenAI- and Deepgram-compatible API. Multi-tenant, and it runs in Docker.
Arabic in the foundation, dialect by dialect.
More than 420 million people speak Arabic, and 1.09 billion speak Arabic, Hindi or Urdu.
| Dialect | Coverage | Status |
|---|---|---|
| الفصحىModern Standard | Wake word, STT, TTS and intent. The formal register of 420M+ speakers. | Supported |
| الخليجيGulf | Wake word, STT and TTS. Tuned for UAE, Saudi, Bahrain, Kuwait, Qatar and Oman. | Supported |
| المصريEgyptian | 100M+ speakers, and the most widely understood Arabic dialect worldwide. | Planned |
| الشاميLevantine | Syria, Lebanon, Jordan and Palestine, with 30M+ speakers. | Planned |
Arabic wake words available
- يا ساجYa Saj
- يا بيلاYa Bella
- اوكي ساجOk Saj
- Custom
Enterprise voice AI for MENA
Banks, telcos and government agencies across the Gulf need voice AI that processes Arabic on the device, with dialect awareness, sovereign data processing and NESA compliance included.
Contact enterprise salesWhy on-device?
Cloud voice APIs charge per minute, leak data and add latency. Running on the device fixes all three.
Total privacy
Zero audio data leaves the device. No cloud processing and no data retention.
Sub-millisecond wake word
Wake-word inference in under a millisecond, against a 200 to 500 ms cloud round-trip, so voice interaction happens in real time without waiting on the network.
No per-minute charges
There is no audio-hour billing. Your models run on your hardware.
How to get Saj Speak.
Saj Link is the application you use every day. Saj Speak, Saj See and Saj Codec are internal capability modules that surface through Saj Sense Business+, the hosted developer API and OEM or Embedded engagements. They are never sold as separate brands.
Saj Speak
Saj Sense Business+
Saj Speak surfaces in the Saj Sense Business+ tier.
Hosted developer API
Saj Speak is also available through the hosted developer API.
OEM / Embedded
Saj Speak inside your own device or product, through an OEM or Embedded engagement.
Built on Saj Speak
Get Saj Link.
Encrypted team communication with Bella AI built in. The post-quantum protocol is built and rolling out. Available now for macOS and the web.
iOS on TestFlightAndroid and Windows are coming soon.
- End-to-end encryption
- Sealed sender, and end-to-end encryption for messages and shared media. Call audio is relayed and is not yet end-to-end encrypted.
- Voice intelligence
- On-device transcription, translation and search that understand dialect.
- Arabic-first design
- Native right-to-left layout, dialect awareness, and voice messages as rich cards.