ElevenLabs Scribe v3 Live listens to context, not just silence, so voice agents stop cutting callers off mid number
The most common complaint about AI phone agents is that they interrupt. ElevenLabs is going after exactly that.
What happened
ElevenLabs launched two speech to text models. Scribe v3, for recorded audio, makes 27% fewer word errors than Scribe v2 across 102 languages, the company says. Scribe v3 Live, the low latency version, makes 43% fewer errors than Scribe v2 Realtime and adds speaker diarization, labeling who said what, in real time for the first time.
The fix voice agent builders will care about is turn detection. ElevenLabs says silence timers force a trade off: a short threshold interrupts people who pause mid sentence, and a long one makes the agent feel slow. Scribe v3 Live uses the conversation's context as well as the audio, so a caller who pauses halfway through a number is not cut off.
There are also output controls: keyterm prompting for up to 1,000 terms like brand names and jargon, plain English formatting rules such as writing every date as YYYY-MM-DD, and a multilingual mode for callers who switch languages mid sentence. Switching from Scribe v2 in the API is a change to the model ID.
The Decoder reports that Artificial Analysis benchmarks put Scribe v3 among the most accurate transcription models, but mid pack on speed and price, with models from Microsoft and xAI running many times faster at roughly half the cost with slightly higher error rates.
My take
For voice agents that book appointments or qualify leads, the transcript is the input to everything: the CRM note, the booked slot, the follow up text. A misheard phone number or a caller who hangs up because the agent kept interrupting is a lost lead.
Where I would use this: inbound call agents, especially with callers reading out numbers, addresses or reference codes, and teams with mixed language callers. Here in the Philippines that means Taglish, so the multilingual mode is worth testing.
Where I would not: bulk transcription of call recordings for reporting. If speed and cost matter more than a few points of accuracy, the benchmarks point elsewhere. Pick the model per job, and load your client's product names into the keyterm list either way.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
