AI voice technology company ElevenLabs has launched two new speech-to-text models, Scribe v3 and Scribe v3 Live, designed to improve transcription accuracy across recorded and real-time audio. Announced on November 10, 2026, the models introduce improvements in multilingual transcription, speaker identification, conversational turn detection and the handling of specialised terminology. ElevenLabs says Scribe v3 produces 27% fewer word errors than its predecessor, while Scribe v3 Live reduces transcription errors by 43% compared with Scribe v2 Realtime.
The launch expands ElevenLabs’ speech technology beyond its established text-to-speech and voice-generation products. The company is targeting businesses building meeting assistants, contact-centre automation, live voice agents, healthcare documentation and multilingual applications. Both models are available through the ElevenLabs API, while Scribe v3 Live is also available through ElevenAgents. The upgrades are intended to make AI-powered conversations more reliable, particularly in situations involving multiple speakers, background noise, accents and people switching languages mid-conversation.
What Are ElevenLabs Scribe v3 and Scribe v3 Live?
Scribe v3 and Scribe v3 Live are speech recognition models that convert spoken audio into written text.
Although they share the same underlying goal, they are designed for different types of applications.
Scribe v3 is intended for recorded audio. It can process meetings, interviews, podcasts, lectures, customer-service calls and other audio files before producing a transcript.
Scribe v3 Live is designed for real-time transcription. It processes incoming audio streams and generates text while a conversation is still taking place.
The distinction is important for developers. A company building a meeting transcription service may prioritise accuracy and detailed speaker labels after a recording finishes. A voice-agent developer, meanwhile, needs a system that can recognise speech quickly enough to respond naturally during a live conversation.
| Feature | Scribe v3 | Scribe v3 Live |
|---|---|---|
| Primary purpose | Recorded audio transcription | Real-time speech recognition |
| Main use cases | Meetings, interviews, subtitles and recorded calls | Voice agents, live captions and dictation |
| Accuracy improvement | 27% fewer word errors than Scribe v2 | 43% fewer errors than Scribe v2 Realtime, according to ElevenLabs |
| Speaker identification | Improved speaker diarization | Real-time speaker diarization |
| Multilingual support | 102-language benchmark evaluation | Real-time multilingual transcription |
| Availability | ElevenLabs app and API | ElevenLabs app, API and ElevenAgents |
Scribe v3 Delivers 27% Fewer Transcription Errors
One of the biggest improvements in Scribe v3 is its ability to recognise spoken words more accurately.
According to ElevenLabs, the model reduces word error rate by 27% compared with Scribe v2 across its evaluation of 102 languages.
Word error rate (WER) measures how frequently a transcription system incorrectly inserts, deletes or replaces words when converting speech into text. A lower score indicates fewer transcription mistakes.
The improvements were not limited to English. ElevenLabs reported better results across every language tested, including languages where speech recognition systems have traditionally faced difficulties.
Some of the reported improvements include:
- German: 42% fewer word errors.
- Spanish: 35% fewer word errors.
- English: 23% fewer word errors.
- Swahili: 34% fewer word errors.
- Tamil: 29% fewer word errors.
- Thai: 28% fewer word errors.
The improvement in Tamil is particularly relevant for multilingual markets such as India, where businesses increasingly need transcription systems that work across regional languages.
ElevenLabs also says Scribe v3 delivers more accurate speaker labels, making it easier to identify who said what during a conversation.
Better Speaker Identification for Meetings With Multiple People
Identifying individual speakers is one of the most important challenges in meeting transcription.
A transcription system may recognise every spoken word correctly but still produce an unreliable transcript if it assigns statements to the wrong person.
This problem becomes more difficult when several participants speak quickly, interrupt one another or have similar voices.
Scribe v3 introduces improvements to speaker diarization, the process of determining which speaker produced each section of audio.
ElevenLabs reports 12% fewer diarization errors overall compared with Scribe v2. For recordings involving three or more speakers, the improvement increases to 17%.
The company says this makes the model particularly useful for:
- Corporate meetings involving multiple participants.
- Interviews and panel discussions.
- Conference recordings.
- Customer-service conversations.
- Podcasts and other recorded media.
More accurate speaker identification can also improve downstream AI applications. For example, a meeting assistant that correctly identifies participants can generate more useful summaries, assign action items to the right employees and organise conversations more accurately.
Scribe v3 Live Improves Real-Time Transcription
Scribe v3 Live focuses on one of the biggest limitations of real-time speech recognition: balancing speed with accuracy.
Streaming transcription systems must generate text before a speaker has finished talking. This creates a challenge because the model may not yet have enough context to understand a sentence or identify when someone has completed a thought.
ElevenLabs says Scribe v3 Live reduces transcription errors by 43% compared with Scribe v2 Realtime.
The company also reports a 31% reduction in conventional word error rate in its real-time evaluation.
The model introduces several improvements designed for live applications.
Context-aware turn detection
Traditional voice systems often rely on silence to determine when someone has finished speaking.
This approach can create problems when callers pause briefly in the middle of a sentence, read out a long number or speak slowly.
If the silence threshold is too short, an AI agent may interrupt the person. If it is too long, the system can feel slow and unnatural.
Scribe v3 Live introduces turn prediction, which uses conversational context and audio information to estimate when a speaker has finished.
This can help voice agents respond at more appropriate moments instead of relying entirely on fixed silence intervals.
Real-time speaker diarization
Scribe v3 Live also introduces speaker diarization for real-time transcription.
The system can identify different speakers while the conversation is taking place, helping applications organise dialogue more effectively.
ElevenLabs reports 39% fewer diarization errors than the next-best real-time model it tested.
This capability could be useful for contact centres, live meeting assistants and applications that need to maintain separate records of different participants.
New Multilingual Features and Custom Vocabulary
Both models introduce additional controls for developers working with specialised language and multilingual conversations.
Scribe v3 supports language hints and secondary language codes, allowing developers to indicate which languages may appear in a recording.
This is particularly useful for conversations where speakers switch between languages, such as English and Hindi.
The model also supports keyterm prompting with up to 1,000 terms. Developers can provide names, technical expressions, product terminology, acronyms or medical vocabulary that the system should recognise.
Another addition is greater control over transcript formatting.
For example, developers can specify whether numbers should appear as digits or words. They can also apply formatting instructions written in plain English, such as requiring dates to follow a particular format.
Scribe v3 Live adds a keyterm library that can hold thousands of specialised terms. Instead of treating every listed word as equally likely, the model consults the library when it detects a possible match.
Its multilingual mode also supports conversations in which speakers switch languages in the middle of a sentence, generating text in the appropriate writing system.
ElevenLabs Targets Voice Agents and Enterprise Applications
The new models are particularly relevant to companies developing AI-powered voice agents.
These systems need to understand customer requests accurately before generating responses or completing tasks.
A small transcription mistake can create significant problems. For example, mishearing a postcode, account number or house number could cause a customer-service agent to retrieve incorrect information or fail to complete a request.
ElevenLabs highlighted results from Eneco, a Dutch energy provider serving more than two million homes and businesses.
The company tested Scribe v3 Live in its voice-agent environment ahead of launch. According to ElevenLabs, the agent recognised Dutch postcodes correctly on the first attempt 40% more often, while recognition of house numbers improved by 15%.
The results illustrate how improvements in speech recognition can affect practical customer-service operations.
Other potential applications include healthcare documentation, live captioning, financial services, education, meeting transcription and media production.
ElevenLabs is also releasing a version of Scribe v3 Live optimised for medical conversations through ElevenAgents.
API Availability and Enterprise Data Protection
Developers can access both models through the ElevenLabs Speech-to-Text API.
Existing integrations can move from Scribe v2 to Scribe v3 by changing the model identifier to scribe_v3. Real-time applications can use scribe_v3_live instead of the previous real-time model.
Scribe v3 Live is also available through ElevenAgents, ElevenLabs’ platform for building voice and chat agents.
The company is offering additional enterprise deployment options for organisations with specific data-residency requirements.
According to ElevenLabs, both models are available in isolated European Union, India and Singapore environments for enterprise customers. The company also offers Zero Retention Mode, which deletes audio and text after a request has completed.
ElevenLabs says the models are covered by its SOC 2, ISO 27001 and PCI DSS Level 1 programmes and are HIPAA-eligible when used under an appropriate Business Associate Agreement.
These capabilities are important for businesses processing customer conversations, healthcare information and other sensitive audio.
The Bigger Picture
The launch of Scribe v3 and Scribe v3 Live highlights the growing importance of speech recognition in the AI industry. As businesses move from text-based chatbots towards voice agents that can handle conversations and complete tasks, accurate transcription becomes a foundational technology. A system that misunderstands a customer’s words can generate incorrect responses, make mistakes in automated workflows or fail to complete an interaction.
ElevenLabs is competing in a market that includes established speech recognition providers and major AI companies. Its latest models combine transcription accuracy with speaker identification, multilingual support and real-time conversation management. The company’s challenge will be demonstrating consistent performance across different accents, noisy environments and specialised industries while making the technology practical for large-scale deployments.
Looking Ahead
Scribe v3 and Scribe v3 Live give developers new options for building transcription services and real-time voice applications. The improvements in speaker identification, language handling and turn prediction could make AI-powered meeting assistants and customer-service agents more reliable. Enterprise adoption will depend on real-world accuracy, pricing, latency and how easily the models integrate with existing software systems.
As voice becomes a more important interface for AI, competition will increasingly focus on systems that can understand conversations naturally rather than simply convert speech into text. ElevenLabs’ latest launch is another step towards that direction, positioning speech recognition as an essential part of the infrastructure powering conversational AI.
SEO Metadata
SEO Title: ElevenLabs Launches Scribe v3 and Scribe v3 Live With Better Accuracy
Meta Description: ElevenLabs launches Scribe v3 and Scribe v3 Live, delivering fewer transcription errors, improved speaker identification and real-time multilingual speech recognition.
URL Slug: elevenlabs-scribe-v3-scribe-v3-live-launch
Focus Keyphrase: ElevenLabs Scribe v3
Secondary Keywords: ElevenLabs Scribe v3 Live, AI speech recognition, speech-to-text models, real-time transcription AI, ElevenLabs transcription, speaker diarization, multilingual speech recognition, AI voice agents
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
