ElevenLabs drops Turbo v2 with 400ms voice latency
ElevenLabs just rolled out Turbo v2, cutting speech generation latency down to around 400ms—more than twice as fast as its first-gen models. It keeps quality matching Multilingual V2 and adds mulaw 8kHz output for VoIP, making near-instant voice agents feel much closer to real time.
Anyone building conversational AI knows that if an agent takes more than half a second to reply, the illusion of a natural talk instantly breaks.
ElevenLabs just launched Turbo v2, introducing its fastest text-to-speech model yet. It delivers audio generation latency down to roughly 400ms, making it more than twice as fast as the company's original V1 models without sacrificing quality, which matches Multilingual V2.
Low latency meets telephony
Beyond raw inference speed, ElevenLabs is taking direct aim at voice calling infrastructure.
Turbo v2 adds support for mulaw 8khz output, tailor-made for VoIP services to deliver an even greater speed boost over standard telephony channels. While Turbo v2 currently leads with English, ElevenLabs confirmed that multilingual support is actively in the works.
Why it matters
Sub-second latency is the holy grail for real-time voice agents, customer support bots, and dynamic gaming NPCs. Shaving response times down to 400ms finally leaves enough headroom for LLM processing before a user feels an awkward pause.
Adding telephony-native mulaw 8khz support looks like a very sharp enterprise move. Instead of just chasing ultra-crisp studio quality, ElevenLabs is actively removing friction for traditional call centers looking to plug in AI. The main limitation right now is the lack of full multilingual capability, meaning global use cases will have to wait for the next update.
Sources (1)
- elevenlabs.io(opens in a new tab)Originalelevenlabs.io