Tuesday, November 17, 2026

Top 5 This Week

Related Posts

Scribe v3: ElevenLabs reduces transcription errors and better identifies who is speaking


A meeting with several voices, a client who interrupts in the middle of a contract number, a conversation that switches from one language to another: it is in these situations that automatic transcription stumbles the most. ElevenLabsthe start-up known for its synthetic voices, launched the November 10, 2026 Scribe v3 And Scribe v3 Livetwo new speech-to-text models designed for these real-world conditions. They become the default templates for its application.

What users were asking for

According to ElevenLabs, Scribe has already been used to transcribe hundreds of millions of hours of audio, for note-taking applications, call centers, medical documentation tools and voice agents. The company says it designed this version based on five recurring requests from its customers:

  • a better attribution of speech in rapid exchanges and with three or more speakers;
  • voice agents who no longer interrupt the floor to people who take a break;
  • a transcription live as precise than that of recorded files;
  • better support for conversations that mix two languageslike Hindi and English in India;
  • more control over the text producedfrom numbers to lists of business terms.

Two models, two uses

Model Use Progress announced compared to the previous generation
Scribe v3 Recorded audio files (meetings, podcasts, archives) 27% fewer word errors across 102 languages
Scribe v3 Live Real-time transcription (voice agents, live captioning) 43% fewer meaning errors, 31% fewer word errors

According to ElevenLabs, Scribe v3 is progressing in all languages ​​tested. Among the most common languages, German gains 42%, Spanish 35% and English 23%. The gains are even clearer for languages ​​considered difficult, such as Swahili (34%), Tamil (29%) or Thai (28%). Directly, Greek progressed by 77% and Danish by 54%. The company does not detail the result for French.

Know who is speaking, even live

In a meeting report, attributing each sentence to the correct speaker is as important as the words themselves. This is called the diarization. According to ElevenLabs, Scribe v3 commits 17% fewer attribution errors than the previous version on recordings with three or more speakers, where most models fail.

Notable new feature, Scribe v3 Live offers this identification in real time for the first time, with 39% fewer errors than the best competitor tested by the company. On the English part of its dataset, its attribution error rate is 3.1%, compared to 5.9% to 7.3% for the other live models evaluated.

Voice agents that interrupt less

For a voice agent, such as an automated telephone switchboard, knowing when the interlocutor has finished speaking is crucial. Classic systems rely on silence: too short, the delay cuts off someone who pauses; too long, the agent appears slow. Scribe v3 Live takes into account the context of the conversationin addition to sound, so that a customer who stops in the middle of a number is not interrupted.

The Dutch energy company Enecowhich supplies more than two million homes and businesses, tested the model before launching on its voice agent. According to ElevenLabs, it correctly recognized Dutch postal codes on the first try 40% more oftenand street numbers 15% more often.

More control over the text produced

Scribe v3 adds several settings useful to professionals:

  • a list of key terms can hold up to 1,000 entries (brand names, acronyms, jargon, drug names) to improve their recognition;
  • the choice of number formatin numbers or in full;
  • formatting rules written in everyday languagefor example write all dates in year, month, day format;
  • live, a multilingual mode for conversations that change language mid-sentence, and a library of terms for vocabularies too large for a simple list.

What is one less error point worth?

ElevenLabs offers a concrete benchmark. A one-hour meeting at 150 words per minute is approximately 9,000 words. Going from an error rate of 4% to 2% means avoiding 180 incorrectly transcribed wordsor three per minute of speaking. And these errors often focus on names, numbers, and technical terms, precisely those that an automated summary or agent needs most. At a financial conference, “fifteen” understood as “fifty” can distort an analysis.

However, the figures remain those of ElevenLabs. The company specifies its method: each language is scored separately before calculating an average, which makes its results difficult to compare directly with other publications. It announces that it will soon publish as open source the dataset of meetings and calls used for its tests.

Data, privacy and pricing

The templates work by default in the United States, but some isolated environments in EuropeIndia and Singapore are offered to businesses with data localization requirements. A no-preserve mode removes audio and text as soon as the query ends. ElevenLabs highlights its security certifications, including SOC 2 and ISO 27001, and offers a version of Scribe v3 Live adapted to medical conversations for its agents.

In terms of prices, several specialized media indicate prices identical to those of the previous generation for developers, i.e. approximately $0.20 per hour for Scribe v3 and $0.22 for Scribe v3 Live. For developers already using Scribe v2, migration is limited to changing the template ID.

Increasingly fierce competition

Transcription has become a hotly contested area, between AI giants and specialized players. At the end of October, NVIDIA released Nemotron 3.5 Transcribe, an open transcription model that includes French. ElevenLabs focuses on a specific niche: real, messy, multi-voice conversations, where automatic reports and voice agents come into play. It remains to be verified, on recordings in French, whether the announced gains are confirmed for French-speaking users.

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles