The best AI models
for voice agents.

"Which model is best?" is the first question most teams ask about voice AI, and the answer changes every few months. This article is a dated snapshot, written in October 2026, of the models developers can use to build a phone agent. It sticks to what vendors document on their own pages. Where a claim is a vendor's own benchmark, we say so. Check the linked pages before you decide, because model names, prices and language lists change often.

Two ways to build the voice

There are two architectures. A speech-to-speech model hears audio and answers in audio, in one model. A cascaded pipeline chains three components: speech-to-text, a text language model, and text-to-speech. Speech-to-speech is usually quicker and handles interruptions and tone more naturally; a cascade lets you pick the best component for each job, including for languages a speech-to-speech model does not cover well. Our comparison article goes deeper. Below, we cover both.

Speech-to-speech models

OpenAI GPT-Live 1

OpenAI's gpt-live-1 was released in the API on 10 September 2026. OpenAI describes it as a full-duplex model that can listen and speak at the same time, handle pauses, interruptions and backchannels, and delegate reasoning and tool use to a backend model or agent. It runs on OpenAI's Live API, not the older Realtime API, and is priced per minute of session rather than per token, with backend model usage billed separately. It accepts G.711 μ-law and A-law at 8 kHz, so phone audio can go in without conversion. OpenAI reports a turn-taking latency of 0.798 seconds against 1.41 seconds for its previous realtime model; that is OpenAI's own measurement.

Clodia runs on GPT-Live 1. One caution for Arabic: OpenAI's GPT-Live documentation does not list supported languages, and none of its built-in voices is listed for Arabic. Arabic quality has to be tested, not assumed.

OpenAI GPT-Realtime 2.1

OpenAI's realtime family continues alongside GPT-Live. gpt-realtime-2.1 is a speech-to-speech model with configurable reasoning effort and tool use, on the Realtime endpoint, priced per audio token. Teams with an existing Realtime integration should note OpenAI's warning that a Realtime integration is not automatically compatible with GPT-Live.

Google Gemini 3.8 Live

Google announced Gemini 3.8 Live on 15 September 2026 as its default model for low-latency voice agents, with an extended-thinking variant that reasons in the background. Google cites support for 97 or more languages and asynchronous function calling. The Live API takes 16 kHz PCM input and returns 24 kHz PCM, so phone audio needs resampling. For Arabic, note a discrepancy: Google Cloud's language and voice configuration page lists only one Arabic locale, Egyptian Arabic (ar-EG), with no Gulf locale, while saying the live models can switch languages during a conversation.

Amazon Nova 2.5 Sonic

Amazon's Nova Sonic models run on Bedrock and integrate with Amazon Connect and several telephony providers. AWS's service card for Nova 2.5 Sonic lists English, Spanish, German, French, Italian, Portuguese and Hindi, and says AWS does not recommend using it for other languages at this time. It is not an option for Arabic callers today.

Microsoft Azure Voice Live

Microsoft's Voice Live API is a managed service that can run OpenAI realtime models or a cascaded pipeline using Azure speech recognition and Azure neural voices. Its value for UAE businesses is the Azure speech catalogue, which includes UAE Arabic locales (below).

Speech-to-text for a cascaded pipeline

A locale code tells you what a vendor has packaged, not how well it understands an Emirati caller on a mobile phone. Test with your own recordings.

Text-to-speech voices

• ElevenLabs lists Arabic (Saudi Arabia, UAE) on its multilingual models, and Arabic on its newer models, with fast variants aimed at conversation.
• Azure offers two UAE Arabic neural voices, ar-AE-FatimaNeural and ar-AE-HamdanNeural.
• Cartesia Sonic 3.6 supports Arabic among 44 languages, and publishes samples labelled Emirati Arabic, with the caveat that labels describe recordings rather than guaranteeing every regional accent.
• Deepgram Aura-2 does not offer Arabic.

How to choose

1. Start from your callers. If most calls are in English, a speech-to-speech model gives the most natural conversation. If many callers speak Gulf Arabic, shortlist components that document Gulf or UAE support, and test speech-to-speech models against them.
2. Test on phone audio. Phone calls are 8 kHz narrowband. A model that shines on studio audio can stumble on a mobile call.
3. Measure the full turn gap, from the caller stopping to the agent starting, on a real call. Vendor latency figures usually exclude the network.
4. Check the tool story. How does the model call your systems, and can you keep consequential actions behind approval?
5. Check data handling. Where audio is processed, how long it is kept, and whether it is used for training.
6. Plan to switch. This market moves quickly. Keep the model behind an interface so you can change it without rebuilding your business logic.

The honest answer

There is no single best model. For natural English conversation on the phone, the leading speech-to-speech models are strong and getting faster. For Emirati Arabic, the documented coverage is thinner and uneven, and research on Arabic dialect recognition shows much higher error rates for Gulf speech than for Modern Standard Arabic. See Arabic and Emirati dialect in voice AI. Whichever model you pick, the deciding evidence is a set of real calls, in your callers' languages, reviewed by someone who speaks them.