ElevenLabs ships Eleven v4: inline-directed TTS, 150ms Turbo tier
The flagship reads [laughs] and [whispers] tags inline across 90+ languages; the Turbo tier answers in ~150ms, and Professional Voice Clones are back.

Copy markdown
Expression you type into the script
Eleven v4 reads inline audio tags — [laughs], [whispers], [door slams] — straight from your text, and tracks who is speaking across multi-speaker dialogue so each line lands in context. SSML break tags are gone; the tags replace them. That is the unlock for audiobooks, dubbed video, and game NPCs without hand-tuning every line.
v4 Turbo answers in ~150ms
The Turbo variant hits ~150ms median time-to-first-speech and ~100ms median inference — versus ~262ms for Cartesia Sonic 3.6 and ~814ms for GPT-4o mini TTS — with bidirectional streaming that takes audio mid-sentence. That is the latency budget a real-time voice agent actually needs.
Voice cloning is back, in 90+ languages
Professional Voice Clones return after being pulled in v3, now with speaker stability that resists drift across regenerations, a 17,500+ voice library, and native-accent preservation across 90+ languages. Clone once and hold the same voice consistent over a long project.
Free tier to start, from $6/mo to scale
Both models are live now on the ElevenLabs API over REST and streaming, with TypeScript and Python SDKs and a one-line model-ID switch between v4 and Turbo. Output is MP3, WAV/PCM, or µ-law for telephony, up to 10,000 characters per generation. Free tier is 10,000 credits (~10 min); paid plans start at $6/month.