“Voice AI” has become an umbrella over four different businesses: speech model providers, platforms that assemble those models into agents, cloud speech services, and open source projects. This guide maps who is who in late 2026, so you can shortlist the two or three that match your job instead of demoing a dozen.
We build one of the companies on this list, and the piece says so. Where a claim matters, it links to a leaderboard or a product page you can check.
Speech model providers
These companies train their own speech models and sell them through APIs. If you are building voice into a product, this is the layer you buy.
Cartesia (that’s us) builds real-time voice models on State Space Model architecture, the research direction our founding team pioneered at Stanford AI Lab. Sonic is our text-to-speech model: it streams audio as it generates, with sub-90ms model latency, 40+ languages, and the #1 rank in blind listening tests on VoiceArena and Artificial Analysis. Ink is our streaming speech-to-text model with turn detection built in, and Managed Agents is a builder for production voice agents. Decagon, ServiceNow, Quora, Retell, and EliseAI build on the platform; see customers and pricing.
ElevenLabs is the largest consumer-facing brand in the space, with a voice library, cloning, dubbing, and a studio aimed at creators. Its homepage now leads with voice agents too. Its strength is creative production; for real-time agent workloads, compare time-to-first-audio directly (our side-by-side with ElevenLabs covers the platform differences).
Deepgram sells speech-to-text and text-to-speech APIs with an enterprise focus. Its reputation is built on fast transcription at scale; evaluate its TTS separately from its STT.
PlayHT has offered TTS APIs for years. Its site did not load when we checked on September 24, 2026, so confirm the service is operating and supported before you build on it.
Speechify is best known as a reading assistant for listening to documents; the company says it has 60M+ users. Its studio and API products are separate from the reading app.
Murf AI and WellSaid Labs both target voiceover production: Murf with a script-to-voiceover editor (it now also pitches conversational agents), WellSaid with enterprise narration workflows.
Voice agent platforms
These companies do not (mostly) train speech models. They orchestrate speech recognition, language models, and speech synthesis into phone- or app-based agents, and sell you the plumbing: telephony, orchestration, observability.
Vapi is a developer platform for composing voice agents from provider-agnostic parts. Retell AI builds production voice agents and uses Cartesia’s voices in its stack. Bland AI focuses on agents that place and handle phone calls at scale. LiveKit sells the real-time audio and video infrastructure, the WebRTC plumbing underneath many voice products.
If you want an agent without assembling the pipeline yourself, start here. If you need control over the models themselves, buy from the providers above and wire your own.
Cloud speech services
Google Cloud (Speech-to-Text and Text-to-Speech), Microsoft Azure AI Speech, and Amazon Polly are the incumbent speech services. They have the widest distribution, deep compliance tooling, and a long track record; voice quality and conversational latency are where startups have been outpacing them. If your organization already runs on one cloud, these are the path of least resistance.
OpenAI ships speech through its APIs, including a Realtime API for speech-to-speech sessions, and its models power voice mode in ChatGPT.
Open source
Fish Audio publishes open-weight speech models (Fish Speech) alongside a hosted platform. Whisper remains the standard open speech-to-text model, and projects like Piper and F5-TTS cover self-hosted text-to-speech. Open source buys you control and data locality at the cost of running the infrastructure yourself; latency and voice quality vary a lot by setup.
How to choose
Match the vendor to the job first, the leaderboard second. A real-time agent needs streaming endpoints and low time-to-first-audio; test with your own script and your own network path, from your own deployment region. Recorded narration needs editing workflow and commercial rights more than it needs 90 milliseconds.
Then test for real: same scripts in both systems, including account numbers, abbreviations, and interrupted replies. Our voice agent guide covers the evaluation checklist for agent workloads, and our comparison hub has side-by-sides with the named competitors.
If you want to hear what the #1-ranked TTS sounds like before reading another comparison, the voice generator runs Sonic in your browser, no sign-up.