Is PlayHT still a viable option to evaluate?

No. Per Crail's data, PlayHT rebranded as Play AI, was acquired by Meta in 2025, and was shut down by early 2026 — it carries the lowest agent-readiness score in the category (6) for exactly that reason. Don't shortlist it for a new project.

Do all these vendors support voice cloning, and does that change what I need to check before using one?

ElevenLabs and Murf AI are both explicitly built around voice cloning (alongside TTS/STT and dubbing), per Crail's data. Cloning a voice — your own or someone else's — is a consent and licensing question, not just a technical one, and it's worth confirming directly with the vendor how they handle authorization before you clone anyone's voice, especially for commercial use.

Choosing a voice AI provider

Voice AI splits along two axes that matter more than any feature checklist — whether your use case is real-time (a live voice agent) or not, and whether voice cloning is involved, which raises consent and licensing questions the rest of the category doesn't.

Voice AI covers text-to-speech, speech-to-text, and increasingly full conversational voice agents, and the six vendors Crail tracks here overlap heavily on paper — most do some mix of TTS, STT, and agent tooling. The real differentiators are whether your use case can tolerate any latency at all, and whether you’re cloning a voice, which brings in consent and licensing considerations the rest of the category doesn’t have to deal with.

Real-time latency is a named design priority for some, not others

Cartesia is built specifically around low latency — its Sonic TTS and Ink STT models use a state-space model architecture chosen for speed, per Crail’s data. Deepgram is likewise positioned around real-time streaming transcription and voice agents. If you’re building a live voice agent (phone calls, live captioning, real-time assistants), these two are worth benchmarking against your actual latency budget before anything else. ElevenLabs, AssemblyAI, and Murf AI all support streaming/real-time use cases too, but latency isn’t the headline of their positioning the way it is for Cartesia and Deepgram.

ElevenLabs and Murf AI are both explicitly built around voice cloning alongside TTS/dubbing, per Crail’s data. Voice cloning is different from generic TTS in one important way: you’re reproducing a specific person’s voice, which raises the same kind of consent and licensing question that recording someone’s likeness does — worth resolving with the vendor and your own legal/compliance process before cloning anyone’s voice for commercial use, rather than assuming the platform enforces this for you.

Pricing has one real outlier

Entry pricing across most of this category is low — Cartesia at $5/month, ElevenLabs at $6/month, PlayHT at $9/month, Murf AI at $19/month, per Crail’s data. Deepgram is the clear outlier at $4,000/annual, reflecting an enterprise-committed pricing model rather than a self-serve monthly plan — worth knowing going in if you were expecting Deepgram to fit a small-team budget the way the others do. AssemblyAI has no published cheapest price on record.

Deployment flexibility if you can’t send audio to a third-party cloud

Deepgram and Cartesia both offer on-prem or BYOC deployment alongside cloud, per Crail’s data (Deepgram adds self-hosted too) — relevant if voice data can’t leave your own environment. ElevenLabs, AssemblyAI, and Murf AI are cloud-only.

Where to start

Full pricing, deployment options, and compliance certifications for all six vendors are on the AI Voice & Speech category page, or see a direct comparison in ElevenLabs vs Deepgram.