ElevenLabs unveils scribe v3 series, boosting transcription accuracy across 102 languages

ElevenLabs has launched two advanced speech-to-text models, Scribe v3 for recorded audio and Scribe v3Live for real-time applications, both available via ElevenAPI and ElevenAgents. Scribe v3 supports 102 languages and reportedly reduces the word error rate by approximately 25% compared to its predecessor, while also enhancing German recognition. The new Scribe v3Live introduces real-time speaker diarization. Independent testing by Artificial Analysis positioned Scribe v3 as highly accurate, surpassing offerings from Google, OpenAI, Microsoft, and Mistral, though it ranked second only to a preview model from Alibaba. Despite this accuracy, the model's benchmark price of $3.33 per 1,000 minutes is noted as higher than some competitors. For customer service, Scribe v3Live improves turn-end detection by integrating conversational context. Beyond transcription, ElevenLabs is expanding its enterprise footprint. The company is deepening its integration with NVIDIA by incorporating Nemotron3.5Super VL into ElevenAgents for robot voice control, and it has partnered with McKinsey, launching a voice agent marketplace alongside system integrators like Salesforce, IBM, and Zendesk.