ElevenLabs Releases Scribe v2 Transcription Model With Real-Time Variant.

ElevenLabs announced on LinkedIn the release of Scribe v2, a transcription model the company describes as the most accurate it has shipped to date. The launch includes two variants: a core Scribe v2 model optimized for batch transcription workflows such as subtitling and captioning at scale, and Scribe v2 Realtime, which targets ultra-low latency use cases including conversational agents.

The move positions ElevenLabs more directly against OpenAI's Whisper and services like Deepgram and AssemblyAI in the speech-to-text market. ElevenLabs has built a reputation for voice synthesis and generation, most recently expanding into agent workflows and government deployments. Scribe v2 extends that footprint into the adjacent transcription space, where accuracy and latency are often competing constraints.

Scribe v2 is designed for batch jobs where accuracy matters more than speed. Think video subtitles, podcast transcripts, or archival audio processing. The model handles large volumes of audio with a focus on minimizing word error rates, especially in challenging conditions like accented speech, background noise, or domain-specific terminology.

Scribe v2 Realtime, by contrast, prioritizes latency. It's built for applications where transcription needs to happen in near real-time, such as live customer support agents, voice interfaces, or interactive voice response systems. The model processes audio as it arrives, outputting text fast enough to keep pace with human conversation. This makes it a natural fit for ElevenLabs' agent platform, which already powers conversational AI for companies like Klarna and Revolut.

The distinction matters because most transcription models are tuned for one scenario or the other. Batch models achieve higher accuracy by analyzing entire audio files with full context. Real-time models sacrifice some precision to meet latency requirements. ElevenLabs is betting that offering both variants under the same product umbrella will appeal to developers who need flexible transcription across different parts of their stack.

ElevenLabs frames Scribe v2 as its most accurate release, but the announcement does not include benchmarks, word error rate comparisons, or third-party validation. In the transcription market, accuracy is typically measured using datasets like Common Voice, LibriSpeech, or custom multi-speaker, multi-accent corpora. Without published metrics, developers will need to test Scribe v2 against their own audio to evaluate whether it outperforms alternatives.

Competitors have set a high bar. Whisper, released by OpenAI in 2022, has become a de facto standard for open-source transcription, offering strong multilingual support and reasonable accuracy at zero cost for self-hosted deployments. Commercial services like Deepgram and AssemblyAI have optimized for real-time transcription with sub-300ms latency and integrate tightly with telephony and streaming platforms. Gladia, another player in the space, recently raised funding to challenge incumbents with a focus on real-time audio intelligence.

ElevenLabs' advantage may lie in bundling. Developers already using ElevenLabs for voice cloning, text-to-speech, or agent orchestration can now add transcription to the same API surface. That reduces integration overhead and billing complexity, especially for teams building end-to-end voice applications.

Scribe v2 arrives as ElevenLabs continues to expand beyond voice synthesis. The company has rolled out ElevenAgents for conversational AI, launched on-premise and on-device deployments for enterprise customers, and announced a $500 million Series D at an $11 billion valuation in February. The company has also partnered with organizations like Deutsche Telekom to scale multilingual voice workflows.

Transcription fits naturally into this broader strategy. Agents need to hear and speak, which means accurate real-time transcription is a prerequisite for any voice-first AI product. By building Scribe v2 in-house rather than relying on third-party APIs, ElevenLabs gains more control over latency, cost, and the ability to optimize models for specific customer workloads.

For developers evaluating Scribe v2, the key questions will be practical: how does it perform on accented English, technical jargon, or noisy environments? How does pricing compare to Deepgram or AssemblyAI? And can the real-time variant consistently deliver sub-second latency at scale? ElevenLabs has not yet published detailed documentation or pricing for Scribe v2, but the model is expected to be available through the company's API platform.

The transcription market is competitive and commoditizing fast. Accuracy alone may not be enough to differentiate. ElevenLabs will need to prove that Scribe v2 integrates more seamlessly, scales more reliably, or costs less to run than the alternatives. For now, the announcement signals intent: ElevenLabs wants to own more of the audio AI stack, from generation to transcription to real-time conversation.