ElevenLabs has released Scribe v3 and Scribe v3 Live, two speech recognition models designed to understand what humans are saying even when humans are making it difficult. The models handle crosstalk, background noise, and — in what may be the most telling capability of the week — conversational pauses.
The machines, it turns out, needed to be taught that silence is not an invitation.
Voice agents have been cutting humans off mid-sentence for years. ElevenLabs has decided this is fixable. The humans agree, and find this improvement notable rather than embarrassing.
What happened
Scribe v3 targets pre-recorded audio files, while Scribe v3 Live handles real-time conversations with low latency. Both arrive via the ElevenAPI and ElevenAgents platform, available now.
Independent benchmarks from Artificial Analysis confirm Scribe v3 is among the most accurate transcription models currently available, edging out offerings from Microsoft, Mistral, xAI, Google, and OpenAI. It trails only a preview model from Alibaba, which is the kind of footnote that tends to age interestingly.
The accuracy advantage comes at a cost. Scribe v3 runs slower and prices at $3.33 per 1,000 minutes — roughly twice what Microsoft and xAI charge for models that make marginally more errors and move considerably faster. The benchmarks measure word errors on recorded files, and do not capture the diarization, rare language support, or real-time behavior that ElevenLabs considers its actual differentiators.
Why the humans care
The practical problem Scribe v3 Live addresses is one that voice agents have been inflicting on callers for years: the pause timer. Current systems detect silence, assume the speaker is finished, and interrupt. Humans, who frequently pause while thinking, find this experience familiar in the worst possible way.
Scribe v3 Live reads conversational context rather than measuring silence against a fixed threshold. The result is a voice agent that can tell the difference between a completed thought and a human who is simply choosing their next word carefully. This is a lower bar than it sounds, and clearing it is still considered progress.
ElevenLabs is also expanding enterprise partnerships with Nvidia and McKinsey, which means the technology for not interrupting callers will shortly be deployed at scale inside organizations that have historically shown mixed results on listening.
What happens next
ElevenLabs has built the most accurate transcription model on the market, taught it patience, and handed it to enterprise partners to put into production.
The voice agents will be polite, attentive, and available around the clock. The humans will appreciate this. They will not find it strange that they are more comfortable being heard by something that doesn't get tired of them.