Arabic Text-to-Speech: Natural, Multi-Dialect Voice Synthesis

In the case of a bilingual Arabic-English test text, the various commercial TTS systems record word error rates exceeding 0.38, and this problem is evident the moment a caller changes languages during a sentence – since that is the way most business conversations in the Gulf and Levant regions actually take place. 

The real problem is not accuracy but identity: when a synthetic voice switches from Arabic to English and back, most engines let the tone change, so the agent sounds as if they were handed over during the call. If you stick with Modern Standard Arabic instead, the voice will be correct but won’t feel authentic to anyone who speaks Najdi, Egyptian, or Levantine Arabic day to day. Rawi was designed specifically to overcome that kind of failure: it maintains a single, stable voice identity throughout the code-switch, using the dialect that the caller actually speaks.

What is Arabic text-to-speech?

Text-to-speech (TTS) turns written text into spoken audio, a problem largely solved for most languages. Arabic is more difficult for three structural reasons:

• Ambiguity of script. In written Arabic, short vowels are frequently left out, so the model has to work out the pronunciation based on the context.
• Diglossia means that formal MSA is separate from the various regional dialects, such as the Gulf dialect, the Egyptian dialect, the Levantine dialect, and others, all of which differ from one another and also from formal MSA in vocabulary and in rhythm.
• Code switching. People in the Middle East often switch between Arabic and English within the same sentence, particularly when giving figures or discussing technical terms.

Why Most Arabic TTS Engines Fall Short

Three failure patterns show up repeatedly in commercial Arabic voice engines:

• Distortion when code-switching. The way words are pronounced changes as the input shifts from Arabic to English and then back to Arabic, especially with regard to numbers and account information.
• Mid-sentence identity drift. There is a noticeable change in the voice’s timbre when switching languages.
• The output is rigid and limited to MSA only. Although it sounds correct in writing, it sounds foreign to someone from the Gulf, Egypt, or the Levant.

For a live voice agent who is dealing with collections, KYC, or support calls, merely one of these situations is sufficient to damage trust during the conversation.

Introducing Rawi: ConvoZen’s Arabic Voice Synthesis Model

Rawi v1 is the neural voice synthesis model developed by Convozen in order to stream high-fidelity Arabic-English code-switched speech across the major regional dialects. This model is based on the architectural method used in Ragini, Convozen’s voice system for Indian languages, which addressed the issue of code-switching for Hindi-English and Tamil-English speech by separating speaker embedding from phonemic language representation. 

Rawi applies this same approach to Arabic, enabling native-like code-switching together with fine-grained dialectal conditioning, so that the model maintains a consistent speaker identity as it switches between Arabic and English within a single sentence.

Dialect Coverage: From MSA to Regional Speech

Rawi conditions its output across four core dialectal pillars rather than defaulting to a single formal register:

It is thus possible for one deployment to provide support for a Gulf retail bank and an Egyptian D2C support line without requiring either of them to have a one-size-fits-all MSA voice.

Built for Real-Time Voice Agents: Latency That Doesn’t Break Conversations

For a live voice agent, natural-sounding speech is only one half of the requirement; it also has to be fast enough in order to allow a real-time conversation. When tested on standard off-the-shelf hardware, Rawi v1 delivers:

By feeding the audio stream into continuous neural buffers rather than waiting until the entire sentence has been generated, the AI agent can start speaking almost immediately, thus maintaining natural turn-taking rather than having to pause, which would reveal that the voice is synthetic.

How Rawi Is Benchmarked: A Scientific Evaluation Framework

Convozen Research evaluates Rawi against five objective dimensions rather than subjective listening tests:

• Intelligibility (WER, CER, PER). The accuracy of transcription at the level of words, characters, and phonemes.
• Naturalness and quality (DNSMOS, SQUIM). Estimates of audio clarity and signal quality based on neural and mathematical methods.
• Consistency of speaker’s voice. The voice remains stable throughout a clipped passage in which the language is switched.
• Prosody and dynamics. The variation in pitch and the expressive quality.
• Spectral purity. The presence of metallic or buzzy audio artefacts is checked.

Arabic Text-to-Speech Benchmark Results

The v1 version of Rawi was tested against Google Chirp, Cartesia, and ElevenLabs in the case of bilingual Arabic-English code-switched text.

Rawi achieves the highest ASR-confidence score (TTScore-Int) of the four models and exceeds both Cartesia and ElevenLabs in terms of accuracy at both the word and character levels when it comes to code-switched text, even though Google Chirp currently has lower error rates on this metric.

With regard to naturalness and speaker consistency, Rawi does better than ElevenLabs in both cases and is overall superior to Cartesia in terms of audio quality, although it comes very close to Google Chirp, which leads in naturalness and speaker consistency in this round of the benchmark.

Use Cases: Where Arabic Text-to-Speech Powers Real Conversations

Rawi is built for the same categories of high-volume, dialect-sensitive voice interactions that Ragini text-to-speech already handles across Indian languages:

• Outbound and inbound contact centre calls
• Collections and KYC verification flows.
• Customer support where a caller may switch between a regional dialect and English within the same sentence

In all these situations, there is a voice agent which has to sound native to the dialect of the person making the call, not one that has learned a formal style from a textbook.