Voice Cloning & Text-to-Speech: Production Guide 2026

Voice cloning and TTS in 2026 split into three tiers: hyper-realistic commercial APIs (ElevenLabs, OpenAI, Google), low-latency real-time engines (Cartesia, PlayHT, Hume), and open-source models you self-host (XTTS-v3, F5-TTS, OpenVoice v2, Sesame CSM, Spark-TTS). Latency, voice fidelity, language coverage, and licensing each pull you toward a different stack.

ElevenLabs v3 (alpha) is the most expressive TTS model on the market with audio tags, 70+ languages, multi-speaker dialogue, and instant voice cloning (IVC) from 1 minute of audio. Professional voice cloning (PVC) requires 30+ minutes for studio fidelity.

OpenAI gpt-4o-mini-tts brings steerable instructions ("speak like a calm therapist") at $0.60/1M input tokens. gpt-realtime is a unified speech-in/speech-out model with sub-300ms latency. Google Gemini 2.5 Flash/Pro TTS supports controllable multi-speaker dialogue at low cost.

Five open-source families dominate self-hosted TTS in 2026: Coqui XTTS-v3 (multilingual cloning, CPML license), F5-TTS (flow-matching, MIT), OpenVoice v2 (cross-lingual cloning, MIT), Sesame CSM-1B (conversational, Apache 2.0), and Spark-TTS (BiCodec, Apache 2.0).

For voice agents and IVR, sub-200ms time-to-first-audio matters more than absolute fidelity. Cartesia Sonic-2 hits ~40ms TTFB, PlayHT Play 3.0 mini ~143ms, and Hume EVI 2 adds emotional prosody and empathic turn-taking on top of an LLM.

A single RTX 4090 (24 GB) or RTX 5090 (32 GB) handles every open-source TTS model in real time with room to spare. The 5090 doubles memory bandwidth (1.79 TB/s) and adds FP4 support, cutting XTTS-v3 latency by ~35% versus 4090.

Commercial pricing varies 200x between providers. Below is the April 2026 list price per 1 million characters for the standard tier of each major provider, normalized for direct comparison.

Two minimal but production-ready snippets: a Python ElevenLabs streaming TTS call, and a local F5-TTS inference using the official CLI on an RTX 4090.