ElevenLabs released Scribe v3 and Scribe v3 Live speech-to-text models

ElevenLabs released two speech-to-text models on 2026-11-10: Scribe v3 for recorded audio and Scribe v3 Live for low-latency realtime audio, available in ElevenAgents and ElevenAPI under the IDs scribe_v3 and scribe_v3_live. Existing integrations upgrade by changing the model ID.

The company reports that Scribe v3 makes 27 percent fewer word errors than Scribe v2 across 102 languages on the FLEURS benchmark, with larger gains in some languages, such as 42 percent in German and 35 percent in Spanish. Scribe v3 Live is said to make 43 percent fewer errors than Scribe v2 Realtime. On speaker diarization, Scribe v3 has 12 percent fewer errors than v2, and Scribe v3 Live posts a 3.1 percent diarization error rate against 5.9 to 7.3 percent for the other realtime models ElevenLabs tested.

Scribe v3 Live adds realtime speaker diarization for the first time, turn detection that uses conversational context as well as silence, and a multilingual mode for conversations that switch languages mid-sentence. Both models support keyterm prompting for up to 1,000 terms. ElevenLabs cites the Dutch energy supplier Eneco, whose voice agent recognized postcodes correctly on the first try 40 percent more often in testing.

Why it matters: transcription for voice agents is shifting from raw word accuracy to conversational features - who is speaking, when a turn ends, mixed languages - that decide whether a live agent can hold a real phone call.

What it does not show: all accuracy numbers are ElevenLabs’ own, the diarization comparison uses an internal dataset, and the post does not state prices.

Sources

Last verified November 16, 2026