Voice models compared: ElevenLabs, Whisper, OpenAI, Cartesia

This is a four-product evaluation guide, not a record of a private benchr test. It separates provider-published capabilities and pricing from the latency, naturalness, Arabic quality, and consent checks your team must reproduce.

Pick your voice stack from these four serious production models. The picks below tell you which one belongs in which path of your app.

This piece covers four common roles: English narration, Arabic narration, speech recognition, and interactive conversation. It also flags voice-cloning consent as a separate product and legal decision. A multi-provider architecture can reduce lock-in, but whether it is worth the complexity depends on your measured workload.

Covered: ElevenLabs for speech generation, Whisper Large V3 for recognition, OpenAI's Realtime API for native audio interaction, and Cartesia Sonic for streaming text to speech. Open-source alternatives are outside this page. No comparative Arabic-quality result is claimed without a published evaluation.

The editorial shortlist starts with ElevenLabs for prerecorded narration and with Cartesia or OpenAI Realtime for interactive paths because that is how the products are positioned. That is not a measured rank. Latency varies enough by region, connection, selected voice, request shape, and provider load that a borrowed millisecond figure should not decide your architecture.

English narration

For prerecorded narration, ElevenLabs belongs on the shortlist because its product emphasizes voice selection and expressive speech. Do not translate that positioning into a universal naturalness result: have target listeners review long clips, names, numbers, abbreviations, and the punctuation patterns your content uses.

OpenAI's Realtime product belongs on the interactive shortlist because it combines native audio input and output. Evaluate pronunciation of technical terms, interruption handling, tool calls, and end-to-end turn latency; this page does not claim a private comparison of voice texture or first-audio time.

Cartesia Sonic is positioned for low-latency streaming text to speech. Include it when responsiveness matters, then score naturalness and rhythm separately from speed. Provider positioning alone does not establish that every connection or selected voice will be faster.

Whether the Cartesia naturalness gap will close as the model trains on more data, or whether it's a fundamental architectural difference from the way ElevenLabs handles prosody, is an open question the public reports can't resolve yet.

Latency checks to reproduce

No millisecond values here are benchr measurements. Use the same script, region, connection, voice, and concurrency.

One uncertainty: voice-cloning consent. The technical capability of voice cloning has run ahead of the consent infrastructure. This piece does not make a policy recommendation; the working assumption in production should be that the legal regime will tighten, possibly retroactively. Build the consent flow into your app now.

Arabic narration

Provider support pages establish which products expose Arabic voices, but support is not the same as native-quality narration. Evaluate Modern Standard Arabic separately from each dialect, and include numbers, names, vowel ambiguity, and mixed Arabic-English text in review.

For OpenAI voice, use the same native-speaker review rather than assuming quality from a language-support label. Record the exact model or voice version because provider updates can change pronunciation without changing your application code.

For Cartesia, confirm current Arabic availability in its documentation and test the specific voices you can deploy. Do not infer dialect fidelity from MSA support; Khaleeji, Egyptian, and other spoken varieties need separate samples and reviewers.

Whisper Large V3 is a common open recognition baseline, but its cited research does not establish performance on your dialect mix or audio conditions. Compare it with current managed alternatives on MSA and each spoken dialect you serve, including noise, code-switching, names, and numbers. For more on Arabic model evaluation generally, see AI for Arabic content.

On most production voice work the top models are closer than the marketing suggests, and the few places they diverge are the places that decide your architecture.

Real-time conversation latency

For an interactive product, measure end-to-end time from the end of user speech to the start of model audio. The table below is a measurement worksheet, not a benchr result.

Voice-stack latency worksheet for your deployment

Stack	Measure	Keep constant	Report
Cartesia Sonic (TTS only)	Text to first audio	Voice, region, text, connection	P50, P95, errors
OpenAI Realtime API	Speech end to response audio	Prompt, tools, region, connection	P50, P95, interruptions
ElevenLabs streaming	Text to first audio	Voice, model, text, connection	P50, P95, errors
Custom ASR → LLM → TTS	Full turn and each stage	All versions and concurrency	P50, P95, failure stage

The custom pipeline is the realistic baseline if you are building your own voice agent. It's what comes out when transcription, LLM, and TTS get chained sequentially. The full-stack solutions are far faster than yours will be because they pipeline the steps and start synthesizing audio before the LLM response is complete.

Set an interaction budget from user research for your product, then report median and tail latency rather than a single best run. Native realtime products can pipeline stages, while a custom ASR-to-LLM-to-TTS chain exposes more components to optimize. The table does not presume which option clears your threshold.

Naturalness review

Use blinded human review; the labels below are evaluation roles, not benchr scores.

Voice cloning

This page previously described a three-listener voice-cloning experiment without publishing the recordings, consent record, versions, or protocol. Those results have been withdrawn and must not be treated as evidence about any provider.

For a legitimate custom-voice evaluation, use only a speaker who has given documented informed consent, follow the provider's current eligibility and disclosure rules, secure the recordings, and define deletion and revocation procedures before upload.

Quality review should be blinded and should separate speaker similarity, intelligibility, prosody, artifacts, and misuse resistance. Publish no identifiable sample without explicit permission.

The ethical implications of voice cloning at this quality are genuine and not in scope for this review. If you are building with this tech, your feature needs a consent and watermarking story before it ships. Skip that and you're setting yourself up for an incident.

Pricing as of April 2026

ElevenLabs charges roughly $0.30 per minute of generated audio on the standard tier, dropping to about $0.15 per minute on the high-volume tier, per ElevenLabs' pricing page. OpenAI's Realtime API runs around $0.06 per minute for audio input and $0.24 per minute for audio output, per OpenAI's API pricing. Cartesia is the cheapest at about $0.04 per minute generated, with volume discounts going lower. Whisper for transcription is around $0.006 per minute, which is cheap enough that the cost of the recognition side is almost always rounding error compared with synthesis. For broader cost context across workloads, see price per use case.

Illustrative calculation only: multiplying one thousand minutes a day by thirty days and the listed per-minute rates yields roughly $9,000 versus $1,200 before volume discounts, taxes, retries, or related services. Verify current rates and replace every assumption with your own usage before making an architecture decision.

Splitting the stack pays off

Managed deployment is now a separate path. OpenAI Presence packages voice and chat agents with policy controls, evaluations, approved actions, escalation, and implementation support. It does not replace Realtime API model selection and it has no public price card. If you are buying an enterprise outcome rather than assembling the stack, use the Presence procurement guide and compare it with an API-built baseline on the same calls.

The right architecture for serious voice work in 2026 runs two providers, not one.

For latency-critical paths, shortlist a streaming synthesis product and a recognition layer, then measure the complete turn. Cartesia and OpenAI Realtime are reasonable starting candidates; Whisper is a common recognition baseline. This page does not claim a sub-300 ms result or a universal cost ratio.

For your narration paths (audiobooks, long-form spoken content, recorded podcasts) use ElevenLabs v3. The quality difference is audible after a few sentences and the latency does not matter when the audio plays back later.

For your multilingual paths where the target audience is Arabic-speaking, ElevenLabs handles MSA well enough to ship for premium content. Cartesia handles MSA well enough to ship for utility content. Neither handles dialect at the level a native speaker would produce. For dialect work that matters to your audience, budget for human voice talent.

The voice AI market has matured to the point where choosing a provider is now an architecture decision rather than a quality one. Each of the four serious players fits a different workload and falls short on the rest. The most common mistake teams make in 2026 is picking one provider for everything (usually ElevenLabs, because of name recognition) and absorbing the latency or cost penalty instead of splitting the stack across the paths in your app.

If you're building voice features in the next year, design your system to support multiple providers from the start. Wrap your synthesis and recognition layers behind clean interfaces. Use the right model for each of your paths.

Most production deployments you will encounter lean on ElevenLabs for narration and Whisper for recognition, with Cartesia handling the real-time paths. Pick what fits each task; the cost difference is too big to run one tool for everything.