ElevenLabs has released two next-generation speech-to-text models, Scribe v3 and Scribe v3Live, aimed respectively at transcribing pre-recorded audio and low-latency real-time voice applications. They focus on improving recognition in complex speech environments, including multi-person conversations, interruptions, and background noise. Both models are now available to developers through the ElevenAPI and ElevenAgents platforms.
ElevenLabs says Scribe v3 supports 102 languages and reduces word error rate by about 25% compared with the previous-generation model, while also significantly improving German recognition accuracy. Its speaker diarization is also more reliable in large group conversations, and Scribe v3Live supports real-time speaker diarization for the first time.
Independent benchmarking by Artificial Analysis shows that Scribe v3 is one of the most accurate speech transcription models currently available, outperforming products from Microsoft, Mistral, xAI, Google, OpenAI, and others, and ranking second only to a preview model from Alibaba.
However, the model is not competitive in processing speed or price: its benchmark price is $3.33 per 1,000 minutes, while some products from Microsoft and xAI offer faster processing and lower costs. The tests mainly measure word error rate in recorded-audio transcription and do not yet cover strengths such as speaker diarization, less common languages, or real-time performance.
For AI voice customer service scenarios, Scribe v3Live also improves turn-end detection. By combining conversational context and audio signals, it reduces false judgments caused by traditional fixed pause timers and prevents users from being interrupted when they pause briefly or read numbers slowly. Developers can also preload specialized terminology, set formatting rules for numbers and dates, and handle mid-sentence language switching through multilingual mode. Enterprise customers can use dedicated EU servers and a zero-data-retention mode.
In addition, ElevenLabs is expanding its enterprise ecosystem. This includes deepening its collaboration with NVIDIA by integrating Nemotron3.5Super VL into ElevenAgents and developing local voice control capabilities for robots; it has also established a partnership with McKinsey and launched a partner program for system integrators and a voice agent marketplace, with companies such as Salesforce, IBM, and Zendesk already offering related products. This update shows that ElevenLabs is expanding beyond voice generation into high-accuracy transcription, real-time interaction, and enterprise-grade voice agent infrastructure.
