ElevenLabs is expanding its real-time Speech-to-Text capabilities as developers increasingly use voice as an interface for AI agents and other interactive applications. While Scribe v2 Realtime itself was introduced in November 2025, the latest updates around its realtime API and capabilities show how ElevenLabs is continuing to develop the technology for live applications.
Scribe v2 Realtime is designed to transcribe speech as a conversation happens, rather than processing an audio recording after the fact. The platform supports more than 90 languages and offers approximately 150 milliseconds of realtime transcription latency, according to ElevenLabs.
The distinction is important: Scribe v2 is designed primarily for batch transcription, while Scribe v2 Realtime is built for applications where transcription needs to happen continuously and feed information into another system, such as an AI voice agent.
How Scribe v2 Realtime Fits Into Live AI
Scribe v2 Realtime is designed for applications where speech needs to be converted into text while a conversation is still happening. ElevenLabs says the model can deliver transcription with approximately 150 milliseconds of latency and supports more than 90 languages.
It is available through the Speech-to-Text API and is aimed at use cases including voice agents, meeting assistants, and other live applications. Unlike batch transcription, realtime Speech-to-Text has to continuously process incoming audio and provide usable text while the conversation is still taking place.
That makes it fundamentally different from conventional transcription systems that process a recording after the conversation has ended. For a voice agent, the transcription is not the final product. It is an intermediate layer between what a person says and what an AI system does next.
Live audio → Scribe v2 Realtime → transcript → AI agent → response. Reducing the delay in that chain can help an application respond more naturally.
Scribe v2 vs Scribe v2 Realtime
The two models serve different transcription workflows. Scribe v2 is designed for batch processing of recorded audio and video, while Scribe v2 Realtime is designed to process streaming audio during a live interaction.
That difference affects how developers build around them. A recorded interview, podcast, or video can be uploaded and processed after the fact. A customer-service voice agent, by contrast, needs transcription while the caller is still speaking so that the AI system can determine what to do next. Scribe v2 supports capabilities such as speaker diarization, word-level timestamps, keyterm prompting, and entity detection. Scribe v2 Realtime focuses instead on continuous streaming, voice activity detection, and controls for handling partial and committed transcript segments.
Partial Transcripts Give AI Systems an Early Signal
One of the more important characteristics of realtime transcription is the difference between partial and committed text.
Scribe v2 Realtime can provide a partial transcript for the segment currently being spoken, while committed transcripts represent finalized portions of the conversation. ElevenLabs notes that early recognition can change as additional context arrives, making the commit mechanism useful for producing a more stable conversation history.
For developers, this creates an opportunity to separate speed from final accuracy. An application can use partial speech recognition to anticipate what a user is saying while relying on committed segments for downstream processing where stability matters more.
Entity Detection Moves Transcription Beyond Plain Text
Another relevant development is the addition of entity detection to Scribe v2 Realtime. ElevenLabs’ documentation says realtime transcription can detect supported entity categories and return their timestamps. The current API also exposes an entity_detection option for realtime sessions.
That matters in applications where recognizing what was said is only part of the task. A customer-service agent, for example, may need to identify names, account-related information, or other structured entities from a conversation. Feeding those signals directly into an AI workflow can reduce the amount of separate processing required after transcription.
Why the Architecture Matters for AI Agents
The significance of realtime Speech-to-Text goes beyond transcription. In an AI voice application, transcription is one of the first processing layers between human speech and an AI-generated response. The bigger significance of Scribe v2 Realtime is therefore not simply faster transcription. It is the role transcription plays inside a real-time AI stack. A voice agent needs to capture audio, determine when someone is speaking, convert speech into text, interpret the request, generate a response, and potentially turn that response back into speech. Each stage adds latency.
ElevenLabs positions Scribe v2 Realtime as the speech-recognition layer in that pipeline. Its streaming architecture, partial transcripts, VAD, and manual commit controls give developers mechanisms for managing the interaction as it happens. The approximately 150ms figure should also not be interpreted as the total response time of a finished voice application. Network conditions, audio capture, AI inference, text-to-speech generation, and application architecture can all add additional delay.
What Developers Can Build With It
The technology can support more than conventional live captions. Potential applications include customer-service voice agents, meeting assistants, live translation workflows, accessibility tools, and other conversational interfaces.
The general significance of this technology is the fact that realtime Speech-to-Text is transforming from a tool for making transcripts into an infrastructure layer for voice-based AI applications. In such an architecture, Scribe v2 Realtime may operate between the live audio feed and the AI application that needs to make sense of what is being said and respond to it.
Scribe v2 Realtime was introduced in 2025, so the story is not about the arrival of realtime transcription at ElevenLabs. Instead, the continuing development of its realtime capabilities shows how the company is building Speech-to-Text into the wider stack required for interactive AI applications.
Conclusion
ElevenLabs’ Scribe v2 Realtime highlights the growing importance of Speech-to-Text as a core component of real-time AI systems. Rather than simply generating transcripts, realtime transcription can serve as the link between live human speech and the AI systems that interpret and respond to it.
With support for more than 90 languages, approximately 150ms transcription latency, streaming input, partial and committed transcripts, VAD and entity detection, Scribe v2 Realtime gives developers greater control over how speech is processed during an ongoing interaction.
These capabilities can support applications ranging from voice agents and meeting assistants to accessibility and live translation tools. At the same time, the approximately 150ms figure should be understood as the latency of the transcription layer, not the total response time of a voice AI application. Audio processing, network conditions, AI inference, application architecture, and text-to-speech generation can all add to the overall delay.
As ElevenLabs continues to develop its realtime capabilities, the broader trend is becoming increasingly clear: Speech-to-Text is evolving from a standalone transcription tool into an infrastructure layer for interactive, voice-driven AI applications.
(Source)