Best text-to-speech for AI voice agents: how to choose (2026)
For a voice agent, text-to-speech is judged on a phone line, mid-conversation, not on a studio demo. The shortlist depends on four things: how quickly the first audio arrives, how it sounds at phone quality, how it reads names and numbers, and how it is priced at your volume.
What matters for real-time agents
Streaming output (usually over WebSockets) so audio starts before the whole sentence is synthesized; low time-to-first-audio; clean 8 kHz output for phone calls; and word-level timestamps or alignment so your orchestration knows what was actually spoken when the caller interrupts. Vendors publish their own latency figures. Treat them as a starting point and measure from your own region.
Low-latency, conversation-first models
These vendors build primarily for real-time conversational agents: Cartesia's Sonic models, Rime's Mist and Coda models, and LMNT. Rime also offers self-hosting and VPC/on-prem deployment.
- Cartesia — Low-latency generative TTS models designed for realtime conversational agents.
- Rime — Realtime TTS voices optimized for natural conversational agents.
- LMNT — Low-latency neural TTS for conversational and media applications.
Broad voice libraries and cloning
If you need many languages, a large voice catalogue or a cloned brand voice, these platforms cover more ground. ElevenLabs offers separate fast models aimed at agents (Flash, v4 Turbo) alongside its more expressive ones, so check which model your platform actually calls.
- ElevenLabs — High-quality text-to-speech and voice cloning used widely in voice agents.
- PlayHT — AI voice generation and cloning platform with streaming TTS APIs.
- Resemble AI — Voice cloning and generative TTS platform with realtime options.
Cloud and model-provider TTS
If you already run on a cloud or model provider, its TTS can simplify billing, compliance reviews and data agreements. The trade-off is usually a smaller choice of conversational voices.
- OpenAI TTS — OpenAI text-to-speech API for generating spoken audio from text.
- Amazon Polly — AWS neural text-to-speech service for scalable voice output.
- Azure AI Speech — Microsoft Azure speech services including STT for enterprise voice systems.
Compare pricing in the same unit
Pricing units differ. The ElevenLabs API charges per 1,000 characters, Rime charges per minute of audio (and says a minute is about 1,000 characters), and Cartesia sells monthly credits at 1 credit per character. Convert everything to cost per minute of agent speech at your expected volume, and remember the agent usually speaks for only part of each call minute.
Using TTS through a voice agent platform
Hosted platforms often resell TTS at their own per-minute rates and let you choose per agent. Retell, for example, lists ElevenLabs, Cartesia, OpenAI and other voices at separate prices. Open-source frameworks such as Pipecat and LiveKit Agents let you plug in a TTS provider directly with your own API key.
- Retell AI — Voice AI platform for phone agents with low-latency conversation and telephony integrations.
- Vapi — Developer platform for building and deploying voice AI agents with telephony, STT/TTS, and orchestration.
- Pipecat — Open-source Python framework for building realtime voice and multimodal AI agents.
- LiveKit Agents — Framework for building realtime voice AI agents on LiveKit WebRTC infrastructure.
How to test before you commit
Use one script with the hard parts (a name, an address, an order number, an email, a price) and play it through each candidate on a real phone call. Listen for mispronounced alphanumerics, unnatural pauses and how quickly audio starts after the caller stops talking. Check voice-cloning and commercial-use terms if you need a custom voice.
Comparisons
Sources
Browse the categories
Building one of these tools? Submit it free, or get featured in front of builders reading this.