Scribe v3 Live

ElevenLabs released Scribe v3 Live on November 10, 2026, targeting ultra-low-latency, real-time speech-to-text applications. Built specifically to handle live audio streams in interactive environments, the model cuts word and semantic errors by 43 percent compared to its predecessor, Scribe v2 Realtime. Rather than processing recorded audio files in post-production batches, Scribe v3 Live operates directly on incoming audio streams to provide live transcription for conversational agents, customer contact centers, and automated meeting software.

Under the hood, Scribe v3 Live utilizes a dense architecture with undisclosed parameters. Because ElevenLabs maintains the model as proprietary software accessible via ElevenAPI and ElevenAgents, exact architectural internals and raw parameter sizes have not been officially documented. However, its behavioral profile places it squarely in competition with enterprise streaming transcription engines such as Deepgram Nova-3 and OpenAI real-time audio endpoints.

For developers building responsive voice agents, Scribe v3 Live addresses the two historical friction points of live audio: speaker overlap and premature agent interruptions. By introducing streaming speaker diarization and context-aware turn detection, the model tracks who is speaking and predicts when a speaker has actually completed a thought, preventing voice bots from cutting off users who pause mid-sentence to think.

Scribe v3 Live operates on a dense architecture with undisclosed parameters. In a dense neural network, every parameter is activated for every compute pass, in contrast to mixture-of-experts (MoE) designs that route data to specific subsets of weights. For real-time speech-to-text, dense networks offer steady, predictable memory bandwidth consumption and uniform computational latency per audio chunk, which is essential for maintaining strict streaming budgets.

While the exact context length is not specified, Scribe v3 Live maintains conversational context across consecutive audio frames. Traditional real-time automatic speech recognition (ASR) engines operate on fixed time slices, relying on simple silence timers to decide when an utterance has finished. Scribe v3 Live incorporates semantic discourse context into its acoustic processing loop, allowing the model to distinguish between a mid-sentence pause and a conversational turn handoff.

Latency is the defining metric for Scribe v3 Live performance. The system is engineered for median transcription latencies around 150 milliseconds. This enables voice agent pipelines to ingest audio, complete text generation, and stream synthetic speech back to the caller within human conversational tolerances (typically under 500 milliseconds total pipeline latency).

Scribe v3 Live delivers targeted capabilities for live audio streams rather than general-purpose text generation:

• Real-Time Streaming ASR: Transcribes live multi-accent speech across more than 90 languages with a 43 percent error reduction over Scribe v2 Realtime.
• Live Speaker Diarization: Labels "who spoke when" directly within the live stream for the first time in the Scribe series, allowing multi-party conversations to be attributed in real time.
• Contextual Turn Prediction: Evaluates grammatical structure and conversational intent to prevent audio truncation when speakers pause to retrieve information or read digits.
• Low-Latency Agent Orchestration: Deep integration with ElevenAgents for automated call routing, interactive voice response (IVR) platforms, and real-time meeting transcription.

Primary use cases center on two-way spoken communication. In call center automation, the model captures overlapping dialogue and maintains accuracy through telephony background noise. In interactive voice agents, the improved turn detection ensures users can speak naturally without adapting to artificial, robotic cadence constraints.

Engineers frequently search for instructions to run Scribe v3 Live locally. Because ElevenLabs distributes the model strictly through ElevenAPI and the ElevenAgents cloud platform, local weights, quantization files, and runtime execution code are not publicly available. Scribe v3 Live cannot be run as a fully offline local AI model undisclosed parameters 2026 deployment using Ollama, vLLM, or llama.cpp.

However, developers designing on-premise hardware infrastructure to support streaming speech pipelines at this accuracy tier must understand the baseline hardware requirements for comparable models. If ElevenLabs ever issues an on-premise container, or if you are sizing hardware for comparable dense streaming ASR engines, consider the following parameters.

Dense models tuned for low-latency streaming ASR and diarization typically land between 1 billion and 3 billion parameters. Based on these architectural baselines, estimated memory footprints break down as follows:

• FP16 / BF16 (Unquantized): 6GB to 8GB VRAM minimum for weight storage, plus 2GB to 4GB for streaming audio buffers, KV cache, and diarization tracking.
• Q8_0 Quantization: ~4GB to 5GB VRAM. Preserves maximum acoustic resolution with negligible error rate degradation.
• Q4_K_M Quantization: ~2.5GB to 3.5GB VRAM. The best quantization for Scribe v3 Live class models when balancing minimal memory footprint against transcription accuracy.

When evaluating how to run undisclosed model on consumer GPU hardware in low-latency mode, memory bandwidth and tensor throughput determine latency far more than sheer VRAM capacity:

• NVIDIA GeForce RTX 4090 (24GB GDDR6X): The best GPU for Scribe v3 Live class real-time streaming pipelines. Its 1,008 GB/s memory bandwidth can handle multiple simultaneous real-time audio streams with processing overhead remaining well below the 150ms latency ceiling.
• NVIDIA GeForce RTX 4080 / 4070 Ti Super (16GB GDDR6X): Fully capable of running dense streaming ASR pipelines alongside downstream small language models (SLMs) in unified voice agent stacks.
• Apple Silicon (M4 Max / M3 Max with Unified Memory): Excellent for local development due to high unified memory bandwidth (up to 400 GB/s), allowing continuous streaming ingestion without PCIe bus transfer bottlenecks.

In speech-to-text models, inference speed is measured by Real-Time Factor (RTF) rather than raw Scribe v3 Live tokens per second. An RTF of 1.0 means an engine takes one second to transcribe one second of audio. To achieve Scribe v3 Live's 150ms streaming response target locally, your hardware must achieve an RTF of 0.1 or lower, transcribing incoming audio chunks in less than a tenth of their playback duration.

If your project strictly mandates local hardware execution without external API dependencies, you must look to open-weights alternatives like Whisper Large v3 Turbo or Distil-Whisper, managed through frameworks like Faster-Whisper or TensorRT-LLM on dedicated GPUs.

Deepgram Nova-3 has been the enterprise default for high-speed streaming ASR. Scribe v3 Live competes directly on latency, matching Nova-3 around the 150ms mark while leveraging ElevenLabs' conversational context modeling. While Deepgram excels at raw throughput and keyword boosting across noisy industrial audio, Scribe v3 Live provides better conversational pacing for voice agents by natively preventing mid-sentence cutoffs.

OpenAI Whisper remains the standard for local, self-hosted transcription. When running on an RTX 4090 using Faster-Whisper, Whisper Large v3 achieves excellent accuracy across hundreds of languages. However, Whisper was architected for batch transcription on 30-second windows. Adapting it for live streaming requires artificial chunking, external Voice Activity Detection (VAD), and auxiliary models for diarization (such as PyAnnote). Scribe v3 Live eliminates this pipeline complexity by unifying streaming transcription, diarization, and turn detection into a single engine, though it sacrifices the data sovereignty and zero-cost scaling of local open-weights deployments.