ConvoZen.AI launches Alif and Rawi, native Arabic speech models built for the Middle-East’s code-switched, multi-dialect reality
ConvoZen.AI announced the launch of Alif and Rawi, two proprietary speech models purpose-built for the Arabic language as it is actually spoken: bilingual, multi-dialect, and heavily code-switched with English. The launch marks ConvoZen.AI’s entry into the Middle East, extending a speech AI approach the company has spent nearly a decade refining in India.
A playbook nearly a decade in the making
In 2017, years before generative AI entered the mainstream vocabulary, ConvoZen.AI began building its own speech models rather than relying on off-the-shelf transcription and voice systems. That work produced Akshara, a speechto- text engine, and Ragini, a text-to-speech engine, both engineered specifically to handle the code-switching, accents, and linguistic diversity of Indian languages. The two models are now deployed across 40-plus enterprises in nine Indian languages, including HDFC Bank, Tata AIG, Cars24, Apollo, Maruti Suzuki, and more.
While deploying voice AI solutions for enterprise and government partners across the Middle East, ConvoZen.AI encountered a strikingly similar problem. The MENA region shares the same structural challenges the company had already solved for in India: a large, geographically scattered population, sharp dialectal variance from country to country and even city to city, and a population that code-switches fluidly between Arabic and English as a default mode of communication rather than an exception. Standard Arabic voice AI, built for Modern Standard Arabic and formal, broadcast-style speech, breaks down the moment real, mixed-language conversation begins.
Alif: Purpose-built for code-switched Arabic speech recognition
Alif is ConvoZen.AI’s bilingual Arabic-English speech-to-text model, engineered to transcribe the messy, multidialect, code-switched reality of everyday MENA conversations rather than the clean, formal Arabic that most commercial ASR systems are trained on.
In benchmark testing against leading proprietary and open-source alternatives, Alif v1 recorded the lowest overall average Word Error Rate (WER) among all models evaluated across the full set of eight conversational and open Arabic benchmarks.
One deliberate choice throughout: ConvoZen build Small Language Models, not large ones. SLMs are faster, cheaper to run at scale, and easier to deploy in real-time voice pipelines, which is exactly what enterprise and government use cases actually need, not a bigger model doing more than the job requires.
Overall average WER is shown only for models evaluated across all 8 listed benchmarks. Lower WER indicates higher transcription accuracy.
On SAWTARABI, a benchmark built specifically to test real-world, conversational Arabic-English code-switching, Alif recorded a 6.04% WER, less than half the error rate of the next-best model tested (ElevenLabs Scribe v2, at 14.20%).
Rawi: Natural, code-switching text-to-speech
Rawi is ConvoZen.AI’s native Arabic Code-Switching text-to-speech model, addressing the speaking side of the same problem. Where standard Arabic TTS systems are rigidly built for Modern Standard Arabic and sound overly formal or robotic, particularly when asked to render a mixed Arabic-English phrase or a regional dialect, Rawi is engineered to speak with the natural rhythm, intonation, and multi-dialect flexibility that real-world B2C and enterprise voice interactions demand.
Rawi was built using the same voice-synthesis approach ConvoZen.AI developed for Ragini in India: aggregating disparate datasets, mapping localised phonemes, and engineering specifically for consistent emotional tone and natural-sounding language mixing, rather than treating it as an edge case.