Who this is for
Technical founders and support operators piloting inbound calls on real telephony.
Research Toplists editorial desk Reviewed November 10, 2026 Public documentation review
How we evaluated
These are editorial weights for this decision. Score the candidates with your own pilot results.
- Time to audible reply and interruption40%
- Names, numbers and locale25%
- Full-call cost20%
- Integration effort15%
| Provider | Best for | Starting price | Key strength | Main limitation | Data and commercial use | Sources |
|---|---|---|---|---|---|---|
| Cartesia | Live customer support and interactive voice flows where the first playable audio and barge-in behavior matter. | Free plan for evaluation. Production plans include usage credits. | A streaming speech API designed for conversational playback. | You still have to integrate and test the rest of the call stack, including recognition, inference and telephony. | Confirm retention and the treatment of voice samples for your plan and integration before using real caller data. The pricing page describes commercial-use licensing by plan. Confirm that your subscription covers the intended use. | |
| ElevenLabs | Creator-led narration and customer-facing voices where delivery and voice selection are a major part of the brief. | Free plan for evaluation. API prices vary by model and billing route. | A voice library and separate models for expressive delivery and lower-latency synthesis. | The model, plan, voice and billing route can change your cost and what you are allowed to publish. | Zero-retention options are described for eligible API configurations. Confirm eligibility and settings for your account before sending customer data. Check the commercial license for your plan and the rights to the specific voice. Treat the free plan as an evaluation option. | |
| OpenAI speech and Realtime | Teams that want to prototype a full speech interaction rather than buy a standalone synthesis layer. | Usage-based. Realtime audio input and output have separate prices. | A Realtime API path for audio input and output with tools in the same conversation. | OpenAI lists January 6, 2027 as the shutdown date for tts-1, tts-1-hd and gpt-4o-mini-tts. Moving to Realtime changes both the interface and the way you measure cost. | Consult the data-controls table for the exact endpoint and feature. Retention and zero-retention eligibility depend on the setup. Review the business terms and the voice disclosure instructions before you deploy. |
Where each option fits
Cartesia
Best for Live customer support and interactive voice flows where the first playable audio and barge-in behavior matter.
Reconsider when You still have to integrate and test the rest of the call stack, including recognition, inference and telephony.
Try this Interrupt a reply midway, then resume with a new question. Check how quickly the caller hears the next useful audio.
ElevenLabs
Best for Creator-led narration and customer-facing voices where delivery and voice selection are a major part of the brief.
Reconsider when The model, plan, voice and billing route can change your cost and what you are allowed to publish.
Try this Audition your longest paragraph and your shortest response using different model settings. Cost both paths using the plan you would actually buy.
OpenAI speech and Realtime
Best for Teams that want to prototype a full speech interaction rather than buy a standalone synthesis layer.
Reconsider when OpenAI lists January 6, 2027 as the shutdown date for tts-1, tts-1-hd and gpt-4o-mini-tts. Moving to Realtime changes both the interface and the way you measure cost.
Try this Prototype on the recommended replacement before committing. Check disclosure, interruption, tool calls and the cost of a complete conversation.
Run the check
Run the same ten scripted calls per candidate on the target carrier. Include a person speaking over the agent, a name with an accent and a forced handoff. Log time from end of caller speech to first audible reply, then report median, slowest call and failure count. Use a larger sample before treating p95 as reliable.
Risk to check Use synthetic test calls. Reject any configuration that cannot meet your data handling or human handoff requirements before you score voice quality.
Questions buyers ask
What should a small support team test first?
Start with a separate streaming TTS layer if your current stack already handles recognition and turn-taking. Evaluate a speech-to-speech API separately if you are replacing the whole loop.
Is a fast synthesis demo enough?
Measure when sound reaches the caller. Recognition, inference, streaming, encoding and the carrier all contribute to the delay.
Which service is cheapest?
It depends on the architecture and the call. Measure real audio and request usage, then include recognition, inference, carrier charges and retries.
What rules out a provider early?
A missing locale, unsupported deployment requirement, unclear data terms or failed interruption behavior should stop the pilot until it is resolved.
Source notes and updates
Research Toplists editorial desk Reviewed November 10, 2026 Public documentation review
First published research edition. Use provider links to confirm details that depend on your plan, voice or region.