The company ElevenLabs has released the Scribe v3 and Scribe v3 Live speech recognition models
The company ElevenLabs has released the Scribe v3 (for recordings) and Scribe v3 Live (for real time) models. Scribe v3 reduces transcription error rate by 25% compared to v2 across 102 languages; according to the Artificial Analysis benchmark, it is the most accurate on the market, but slower and more expensive than models from Microsoft and xAI.
The company ElevenLabs has released two new speech recognition models: Scribe v3, designed for transcribing recorded audio files, and Scribe v3 Live, for low-latency real-time applications. Both models are available via the ElevenAPI interface and the ElevenAgents platform and are targeted at difficult audio conditions, such as meeting participants talking over each other or phone calls with disruptive background noise.
According to ElevenLabs, Scribe v3 reduces word error rate (WER) by approximately 25 percent compared to the previous v2 version across 102 languages, with accuracy for German expected to improve significantly. The company also states that it offers more reliable speaker differentiation (diarization) in larger group conversations, which Scribe v3 Live now also offers in real time. Independent benchmarks from Artificial Analysis confirm that Scribe v3 ranks among the most accurate transcription models on the market and outperforms competitors from Microsoft, Mistral, xAI, Google, and OpenAI in accuracy, with only a preview model from Alibaba scoring better.
The source states that Scribe v3 is slower and more expensive compared to the competition — according to the benchmark, models from Microsoft and xAI run many times faster at only a slightly higher error rate and at roughly half the price. Moreover, the benchmark only measures word error rate on recorded files and does not account for improvements in diarization, rare languages, or the real-world behavior of the Scribe v3 Live model.
Scribe v3 Live is meant to address the problem where voice agents, due to fixed pause timers, interrupt callers mid-sentence — the new model analyzes conversational context together with the audio signal to recognize when an utterance is actually complete. The company offers developers the option to pre-load thousands of specialized terms, set number and date formatting using natural language, and use a multilingual mode for switching languages mid-sentence. Enterprise clients have access to isolated servers in the EU and a zero data retention mode after request processing.
Why it matters
For companies deploying call transcription or voice agents, this is a concrete choice between accuracy and speed/price: according to the benchmark, Scribe v3 offers the highest accuracy on the market, but at higher latency and cost compared to competing models from Microsoft and xAI, which is especially relevant for processing large volumes of audio where price and speed matter. Scribe v3 Live also addresses a specific technical problem with voice agents — interrupting callers during pauses in speech.
Relevant practical impact
What this means
For a business
Companies developing or deploying voice agents and transcription tools now have access to a model with the highest claimed accuracy on the market, but at the cost of higher latency and higher costs than competing models from Microsoft and xAI – the choice depends on whether the priority is accuracy, or speed and price.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.