Best Arabic Text to Speech for Voice Agents

If you are building a voice agent for the Arab world, the text-to-speech layer is usually the part that makes or breaks the product.

Speech recognition has become good enough. Large language models handle Arabic reasoning well enough. But the moment your agent opens its mouth, the caller decides in about two seconds whether they are talking to something worth their time — or whether they should press zero and ask for a human.

This guide explains what actually matters when you pick an Arabic TTS engine for a voice agent specifically (not for audiobooks, not for video voiceover), how the current options compare, and why SILMA TTS v2 has become a common choice for companies running Arabic agents at scale.

Why voice agents need a different kind of TTS

Most TTS benchmarks measure the wrong things for this use case.

A voiceover tool can take three seconds to render a sentence. Nobody notices, because the audio is generated once and played back later. A voice agent has no such luxury. It generates audio live, inside a turn-taking conversation, while a human being sits on the other end of a phone line waiting.

That changes the requirements completely:

A voiceover engine is judged on how beautiful the audio is. A voice agent engine is judged on how fast, how predictable, and how correct it is — and only then on how beautiful.

Here is what that means in practice.

1. Time to first token (TTFT), not total render time

The number that matters is how long it takes for the first chunk of audio to come back, because the rest streams while the earlier part is already playing.

Your total response latency is roughly:

Every one of those stages eats into the same budget. Humans start to perceive an unnatural pause somewhere past 500 ms of silence, and by 800 ms they will usually talk over your agent. If your STT and LLM already spend 400 ms, a TTS engine that needs 700 ms has broken the conversation on its own.

This is why "seconds per sentence" is a useless metric here, and why TTFT is the one to negotiate on.

2. Dialect, not just language

"Arabic support" on a vendor page usually means Modern Standard Arabic — the formal register used in news broadcasts and official documents. It is the right choice for a government service announcement. It is the wrong choice for a food delivery app in Riyadh.

A Saudi customer who hears formal Fus'ha from a support agent does not hear "professional". They hear "foreign", or "robot", or "this company does not actually operate here". Dialect is not a nice-to-have feature for a consumer voice agent in the Gulf; it is the difference between a product people use and a product people abandon.

Generalist multilingual models can technically produce Arabic, but they were trained to cover a hundred languages at once. Coaxing an authentic Najdi delivery out of them usually means heavy prompt engineering, custom pronunciation dictionaries, and results that drift between calls.

3. Code-switching that does not break

Real Gulf conversations mix registers and languages constantly, in a single sentence:

"تمام، رح أرسل لك الـ tracking number على الواتساب."

A voice agent that has to switch models — or worse, switch voices — between the Arabic and the English parts sounds broken. You need one voice, one pipeline, that handles Najdi, MSA and English inside the same utterance without changing its identity mid-sentence.

4. Correct handling of the things agents actually say

Voice agents do not read essays. They read order numbers, phone numbers, dates, prices, email addresses, tracking codes and URLs. This is the single most common source of embarrassing failures in production, and it is almost never covered in demos.

Arabic makes this harder than English, because numbers are read differently depending on context, and because a raw string like [email protected] has no obvious pronunciation at all. If your engine cannot be told "this is a phone number, read it digit by digit," you will end up writing brittle text preprocessing yourself.

5. Diacritics (tashkeel) you should not have to write

Arabic script normally omits short vowels. Without them, علم could be ʿilm (knowledge), ʿalam (flag), or ʿallama (he taught). Some TTS systems demand fully diacritized input, which means you now need a separate diacritization model in front of your TTS — another model, another latency hop, another thing that fails.

A good Arabic TTS engine takes plain, undiacritized text and resolves it correctly from context.

6. Concurrency, not just speed

A single fast request means nothing if your engine falls over at peak. A call centre agent product does not need 1 request; it needs 20, 200 or 2,000 simultaneous streams, at the same latency as request number one. Ask every vendor what their concurrency ceiling is on your plan, and what happens when you exceed it — queueing, throttling, or hard errors.

7. Cost per minute at real volume

TTS is often priced per character or per credit, which hides the real number. Voice agents are billed by the world in minutes, so convert everything to cost per minute of generated audio before comparing.

The spread in this market is enormous — roughly 14× between the cheapest and the most expensive providers. At 100,000 minutes a month, the difference between $0.025/min and $0.10/min is $2,500 versus $10,000. Per month. For the same product.

8. Deployment and data residency

If you sell to banks, telcos, healthcare providers or government entities in the GCC, "our API is hosted in the US" ends the conversation. On-premise or VPC deployment inside the region is frequently a hard procurement requirement, not a preference.

The current Arabic TTS landscape

Broadly, your options fall into two groups.

Generalist multilingual providers — ElevenLabs, Cartesia, Deepgram, OpenAI, Google, Azure, AWS. These are excellent engineering products with very fast infrastructure and huge language coverage. Arabic is one of many languages they support, which means MSA is usually decent, dialects are usually weak, and Arabic-specific controls (tashkeel handling, dialect selection, entity tagging) mostly do not exist.

Arabic specialists — SILMA, Hamsa, Munsit and a handful of others. Narrower language coverage, but the Arabic itself is the product rather than a checkbox, and the tooling is built around Arabic-specific problems.

If your agent operates only in Arabic, or in Arabic plus English code-switching, the specialist path usually wins. If you need Arabic as one of thirty languages in a global product, a generalist may be simpler to standardise on — with the dialect trade-off that implies.

Latency and price at a glance

From SILMA's 2026 low-latency benchmark, measured end-to-end from the EU with warmed connections:

Two honest observations about this table.

First, Cartesia and ElevenLabs Flash are genuinely faster on raw network round-trip. They have spent years on globally distributed edge infrastructure, and it shows.

Second, these figures include the network path from Europe. SILMA's server-side TTFT is around 170 ms; the remainder is transit. For traffic originating in the Gulf, or for an on-premise deployment sitting in the same data centre as your telephony stack, that transit cost largely disappears.

The practical takeaway: everything in the top four rows is inside the budget for natural conversation. ElevenLabs v3 is not, at least not for live agents. Once you are inside the budget, the deciding factors become dialect quality, control, concurrency and price — which is where the comparison gets more interesting than a latency table.

Why SILMA TTS v2 API is one of the best choices for voice agent companies

SILMA TTS v2 was not built as a general-purpose speech product with Arabic added later. It was built for Arabic conversational agents, and the design decisions reflect that. Here is the concrete case.

It is trained on conversational speech, not narration

The v2 models were trained exclusively on conversational speech. That sounds like a small detail, but it changes the output character completely. Narration-trained models produce measured, even, slightly performative delivery — perfect for an audiobook, subtly wrong for a phone call. Conversation-trained models produce the shorter phrasing, quicker turns and natural hesitancy that people expect from someone actually talking to them.

For an agent, this is the difference between "a nice voice reading a script" and "someone on the phone".

Real Saudi Najdi, not MSA with an accent

SILMA ships two dialect models you select per request:

• silma-tts-v2-msa — Modern Standard Arabic (Fus'ha)

• silma-tts-v2-ksa — Saudi Najdi dialect

This is dialect as a first-class model choice, not a prompt hint. As SILMA puts it: where others generalize, these models specialize. For any company selling into Saudi Arabia — by far the largest voice-agent market in the region — a native Najdi voice is a commercial advantage, not a technical one.

One pipeline for Najdi, MSA and English

Code-switching is native. A single voice moves between Saudi Najdi, MSA and English inside one utterance, without a model swap and without the voice identity changing mid-sentence. Practically, this means you can write agent responses the way your customers actually speak — product names in English, conversation in dialect — and not build a language-detection router in front of your TTS.

Latency that is engineered for turn-taking

~170 ms TTFT on the server, with sub-300 ms typically observed end-to-end. That leaves real room in the budget for your STT and LLM stages, which is where most of your latency actually lives.

Just as important: the audio is natively streamed, not generated and then chunked. You get three transport options, so you can match your existing stack rather than rebuild it:

If you are already running LiveKit, Pipecat, Twilio Media Streams or an in-house orchestrator, the WebSocket path drops in without architectural changes.

Controls that exist because agents need them

This is where the Arabic-specialist design shows most clearly. The v2 API exposes parameters you will otherwise have to fake with preprocessing:

• Entity tags. Wrap text in <STAG_PN>, <STAG_EMAIL> or <STAG_LINK> and phone numbers, email addresses and URLs are pronounced correctly. No custom number-to-words logic, no regex pipeline in front of your TTS.

• Pronunciation overrides. Server-side rules (enable_server_pronunciation_overrides) so your brand name, product names and local place names are said correctly on every call, permanently — fixed once, not patched per prompt.

• creativity. Controls prosodic variance. Turn it down for deterministic, compliance-sensitive flows like reading back an account balance; turn it up for conversational warmth in a sales or support agent.

• speed. Straightforward rate control, useful for matching a brand's existing IVR pacing.

• Tashkeel-optional input. Send plain, undiacritized Arabic. No separate diacritization model in your critical path.

Voice cloning and eight native voices

Eight production voices ship out of the box — four female (sarah, salma, salwa, saja) and four male (sultan, salman, sulaiman, salim) — which is enough to give different clients or different agent personas distinct identities.

Beyond that, high-quality voice cloning lets you build a brand voice, or let each of your customers have their own. For a voice-agent platform reselling to enterprises, this matters: your customers increasingly want their voice, not a stock one shared with their competitor.

Pricing that survives contact with scale

From $0.025 per minute, which sits at or below every mainstream competitor in the latency table above — including ones that are substantially worse at Arabic.

The arithmetic at agent volumes:

For a voice-agent company, TTS is a direct cost of goods sold. Halving it does not just save money — it changes what margin you can offer, and therefore what price you can win deals at.

Enterprise concurrency and regional deployment

Enterprise plans support up to 20 concurrent requests, with higher ceilings available, and SILMA offers on-premise and VPC deployment for GCC data residency requirements.

That last point is worth dwelling on. If you are selling voice agents to Saudi or UAE banks, telcos, insurers or public-sector bodies, procurement will ask where the audio is processed. A vendor who can deploy inside your VPC — or inside the customer's — turns a blocked deal into a signed one. Most global TTS providers simply cannot do this.

It is verifiable before you commit

SILMA publishes an open-source Arabic TTS benchmark and has released open-weight models under a permissive licence. You can also generate your first five minutes of audio free in the playground.

For a technical buyer, this matters more than any marketing claim: you can hear the dialect quality yourself, on your own scripts, in an afternoon.

A practical evaluation checklist

Before you commit to any Arabic TTS vendor, run this test with your own copy:

1. Take 20 real utterances from your agent's script. Not marketing sentences — actual production strings, including the awkward ones.

2. Include the hard cases: a phone number, an order ID, a price with decimals, a date, an email address, and one sentence that code-switches into English.

3. Generate all 20 on every shortlisted vendor with default settings.

4. Play them to native speakers from your target market — Saudis for a Saudi product — without telling them which vendor is which.

5. Measure TTFT from your own infrastructure, in the region your traffic originates from, not from the vendor's benchmark page.

6. Ramp to your expected peak concurrency and re-measure. Latency under load is the number that matters, and it is rarely the one advertised.

7. Ask about deployment options and data residency before you build, not after your first enterprise deal stalls in procurement.

Most vendors look similar at step 3 and separate sharply at steps 4 and 6.

Getting started with SILMA TTS v2

1. Create a free account at app.silma.ai — the first five minutes of audio generation are free.

2. Try the eight voices in the playground against your own Arabic scripts, in both MSA and Najdi.

3. Generate an API key and review the API specification for the streaming endpoints.

4. Point your agent framework at the WebSocket endpoint and measure TTFT from your own environment.

5. Talk to the team about concurrency limits, volume pricing, or on-premise deployment if you need them.

Frequently asked questions

What is the best Arabic text to speech for voice agents? For agents operating in Arabic — particularly Gulf dialects — an Arabic-specialist engine generally outperforms a generalist multilingual one. SILMA TTS v2 is a strong default because it combines conversational training data, native Saudi Najdi and MSA models, ~170 ms server-side TTFT, native streaming over WebSocket/SSE/REST, and pricing from $0.025 per minute. Global providers like Cartesia and ElevenLabs Flash have lower raw network latency but weaker dialect performance and fewer Arabic-specific controls.

How low does TTS latency need to be for a voice agent? Aim to keep your total response time — STT, LLM, TTS and network combined — under about 800 ms, and ideally under 500 ms. Since TTS is only one stage, a TTS TTFT in the 150–300 ms range leaves workable headroom. Anything above roughly 700 ms for TTS alone will produce noticeable, conversation-breaking pauses.

Does Arabic text need diacritics (tashkeel) for text to speech? Not with a modern Arabic-native model. SILMA TTS accepts plain undiacritized Arabic and resolves pronunciation from context, so you do not need a separate diacritization model in your pipeline. You can still supply tashkeel where you want to force a specific reading.

Can one Arabic voice handle both dialect and English in the same sentence? Yes. SILMA TTS v2 supports native code-switching across Saudi Najdi, Modern Standard Arabic and English within a single utterance and a single voice identity, so English product names and technical terms do not require a model or voice change.

How much does Arabic text to speech cost? The market ranges from roughly $0.007 to $0.100 per minute of generated audio. SILMA TTS v2 starts from $0.025 per minute. Always convert per-character or per-credit pricing into cost per minute before comparing, because voice-agent economics are measured in minutes.

Can Arabic TTS be deployed on-premise for data residency? SILMA offers on-premise and VPC deployment to meet GCC data residency requirements — commonly a hard requirement for banking, telecom, healthcare and government buyers in Saudi Arabia and the UAE. Most global TTS providers offer cloud API access only.

Does SILMA support voice cloning? Yes. SILMA TTS v2 supports high-quality voice cloning in addition to its eight built-in voices, so you can build a distinct brand voice or provision separate voices per customer.

Final thoughts

For a voice agent working in Arabic, "best" is not the model with the prettiest demo reel. It is the model that answers fast enough to hold a conversation, speaks the dialect your customers actually speak, reads an order number correctly, handles your peak concurrency, deploys where your compliance team needs it to, and does all of that at a price that still works when you are running a million minutes a month.

SILMA TTS v2 was built specifically against that list. Start with five free minutes in the playground, run your own scripts through it, and let native speakers decide.