ElevenLabs Launches Scribe v3: More Accurate but Slower and More Expensive

Key Takeaways

1ElevenLabs launches Scribe v3 for pre-recorded audio and Scribe v3 Live for real-time use, designed for noisy environments and cross conversations.

2An independent benchmark credits Scribe v3 with leading accuracy but points out a lower speed and higher cost compared to some competitors.

3Scribe v3 Live analyzes context to prevent voice agents from interrupting callers, and the offering includes control tools, data protection options, and new partnerships with Nvidia and McKinsey.

Why it matters : These models aim to enhance the reliability of voice transcription in business while addressing specific needs for security and customization.

ElevenLabs unveils Scribe v3 for pre-recorded audio and Scribe v3 Live for real-time use. Third-party tests credit it with very high accuracy but highlight slower speed and a higher cost compared to some offerings. The Live version leverages context to prevent voice agents from interrupting callers. Partnerships with Nvidia and McKinsey accompany this enterprise-focused rollout.

Partnerships with Nvidia and McKinsey accompany the rollout

ElevenLabs is expanding its enterprise partnerships. With Nvidia, the company integrates the open model Nemotron 3.5 Super VL into ElevenAgents, secures additional computing capacity running on Nvidia, develops local voice control for robots, and manages a phone agent for Poland's public health system. ElevenLabs also announces an alliance with McKinsey and an expanded partnership program for system integrators. Simultaneously, the company is launching a marketplace where companies like Salesforce, IBM, and Zendesk offer ready-to-use voice agents.

High accuracy confirmed but speed and cost less favorable

An independent benchmark credits Scribe v3 with one of the best accuracies, surpassing Microsoft, Mistral, xAI, Google, and OpenAI, and only being outperformed by a preliminary model from Alibaba. However, the same test indicates a speed lag and a reference price of $3.33 for 1,000 minutes. According to these measurements, the models from Microsoft and xAI are several times faster, with slightly higher error rates, and cost about half. This benchmark evaluates only word errors on recorded files and does not cover diarization, rare languages, or the real-time behavior of Scribe v3 Live.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Real-time aims to avoid interruptions in voice agents

Scribe v3 Live targets a common issue with voice agents based on rigid pause timers, which interrupt speakers during brief silences. The model analyzes the conversational context, in addition to the audio signal, to determine whether a speaker has truly finished, aiming to prevent interruptions mid-sentence.

Control tools for developers and data guarantees

Developers have the option to preload thousands of domain-specific terms to limit vocabulary errors and can also specify formatting rules for numbers and dates expressed in natural language. A multilingual mode allows for handling language alternations within a single sentence. For businesses, ElevenLabs provides dedicated servers located in the European Union as well as a zero retention option that deletes data immediately after each request.

Two models, two uses, available via ElevenAPI and ElevenAgents

Scribe v3 and Scribe v3 Live target challenging listening conditions, in pre-recorded files and real-time streams. Both are accessible via ElevenAPI and the ElevenAgents platform. Scribe v3 is aimed at pre-recorded content and, according to ElevenLabs, reduces word errors by about 25% compared to its predecessor across 102 languages, with a notable improvement claimed in German. Speaker diarization improves for large groups, and for the first time at ElevenLabs, Scribe v3 Live offers this in real-time.