Model releases · · 2 min read
ElevenLabs Scribe v3 Live targets voice agents that interrupt callers
ElevenLabs released Scribe v3 and Scribe v3 Live, speech models built for noisy calls. What the accuracy-versus-speed trade-off means for growth teams.
Takeaway Scribe v3 leads on accuracy but trails rivals on speed and price, so match the model to the job: analysis or live calls.
ElevenLabs Scribe v3 and Scribe v3 Live are two new speech recognition models from ElevenLabs. They are built to transcribe accurately even with background noise and people talking over each other. The Decoder reports that the live version uses conversational context to keep voice agents from cutting callers off mid-sentence.
What shipped
ElevenLabs released two models: Scribe v3 and Scribe v3 Live. Both are speech-to-text systems. The Live variant is aimed at voice agents that have to respond during a conversation.
What is new
The focus is difficult audio. The models are designed to hold up through background noise and crosstalk, which are common on real sales, support and onboarding calls.
The Live model also draws on the context of the conversation. According to the report, the goal is to stop an agent from jumping in before the caller has finished speaking.
Benchmarks (developer-reported)
The figures here come from Artificial Analysis, an outside benchmarking group, rather than from ElevenLabs itself. Per The Decoder, those benchmarks show Scribe v3 narrowly ahead of OpenAI, Google and Microsoft on accuracy.
The lead is small, and it comes with a trade-off. Microsoft and xAI models run many times faster, according to the report. The summary does not give specific scores.
Price and access
The report says rivals such as Microsoft and xAI charge roughly half the price of ElevenLabs. The announcement summary does not list exact prices, plans or availability; see the source.
What it means for marketing
The practical question is which job you are buying transcription for.
For call analysis, such as mining sales and support recordings for objections, phrasing and product feedback, accuracy matters most and speed matters little. A narrow accuracy edge can still mean fewer garbled quotes in the data you feed into messaging and content work. Noisy or overlapping audio is where that difference is most likely to show.
For live voice agents, such as inbound lead qualification or booking, being cut off is a direct conversion risk. Context-aware turn handling in ElevenLabs Scribe v3 Live is worth testing on your own call recordings. The reported cost and speed gap, though, means a cheaper and faster model may be the better fit where latency and volume dominate.
A sensible approach is a small side-by-side test: run the same set of real calls through two or three providers, then compare transcript errors, response delay and cost per hour. Benchmarks from a single outlet's coverage are a starting point, not a verdict on your audio.
Source
Written by the AI CMO OS newsroom from the publisher's report. Facts and figures are as reported by the source; read the original for full details.
More in Model releases
Seedance 2.5 ByteDance video AI now has a 4K API for developers
ByteDance's Volcano Engine has opened API access to the 4K Seedance 2.5 video model. Here is what the release means for marketing and growth teams.
Huawei Cloud plans open source for CloudRobo-WM1.2 world action model
Huawei Cloud released the CloudRobo-WM1.2 embodied world action model and says it will go open source. Here is what that means for growth teams.
Alibaba flagship model Qwen4-Max cuts task costs 60%, adds cyber guardrails
Alibaba launched Qwen4-Max, its most advanced model, claiming 60% lower task costs and stricter cyber safeguards. What growth teams should check before testing it.