LMNT

Low-latency streaming TTS. Sub-100ms first audio.

Streaming-first speech synthesis built around the Sonic family of state-space models. Aims at real-time agent voices where latency between turns is the product. Strong choice for sub-200ms voice loops.

Enterprise text-to-speech built for real-time voice agents.

Rime builds AI voice models for high-stakes business conversations like IVRs, contact centers, and AI phone agents. Its Arcana and Mist models target ultra-low latency and natural, conversational delivery, with deterministic pronunciation control so terms are spoken consistently without retraining. Rime can be deployed on-prem, in a VPC, or via cloud API, and is offered directly or through voice-AI partner platforms.

Frontier TTS, voice cloning, and dubbing. Industry default.

Hosted speech synthesis at near-human quality — TTS, voice cloning, multilingual dubbing, and conversational voice agents. Default choice when you need a voice that sounds like a person, not a robot.

Ultra-low-latency text-to-speech that runs on-device.

Neuphonic is a voice-AI company building text-to-speech and voice cloning that run locally with very low latency. Its cloud API targets real-time voice agents, and in October 2025 it open-sourced NeuTTS Air, a 748M-parameter speech language model that runs on CPU via llama.cpp and clones a voice from a few seconds of audio. Aimed at private, offline, and voice-agent use cases.