ElevenLabs Scribe v3 makes 27% fewer errors, and Live labels speakers

ElevenLabs has two new speech to text models. Scribe v3, for recorded audio, makes about a quarter fewer word errors than Scribe v2. Scribe v3 Live, for realtime audio, cuts errors by more than 40 percent, and labels who is speaking.

Slide text

Scribe v3 makes 27% fewer errors, and Live labels speakers

Scribe v3 Recorded audio −27% word errors against Scribe v2, across 102 languages

Scribe v3 Live Realtime audio −43% errors against Scribe v2 Realtime · speaker labels, a first for realtime

elevenlabs.io/blog/scribe-v3

On recorded audio in 102 languages, Scribe v3 gets about one word in eight wrong, where Scribe v2 missed one in six, and every language improved. It also beats Gemini 3.5 Transcribe, GPT-Transcribe and Deepgram Nova-3.

Slide text

Scribe v3 gets about one word in eight wrong, down from one in six

FLEURS, 102 languages · fewer word errors than every batch model tested

elevenlabs.io/blog/scribe-v3 · elevenlabs.io/docs/overview/models

A silence timer cuts off a caller who pauses halfway through a number. Scribe v3 Live predicts the end of a turn from the conversation as well as the audio. In Eneco’s voice agent, Dutch postcodes came through right on the first try 40 percent more often.

Slide text

Scribe v3 Live lets a caller finish the number

Silence timer

“My postcode is 3011cut off AB”

Scribe v3 Live

“My postcode is 3011 … AB”

+40% Dutch postcodes right on the first try, in Eneco’s voice agent

elevenlabs.io/blog/scribe-v3

Scribe v3 labels who said what more accurately than any model tested, and gains most in meetings with three or more people. Scribe v3 Live brings those speaker labels to realtime transcription for the first time.

Slide text

Scribe v3 puts fewer words under the wrong name

17% fewer speaker errors than Scribe v2 with three or more people talking

Scribe v3 Live: speaker labels in realtime

elevenlabs.io/blog/scribe-v3

Scribe v3 takes up to a thousand key terms, like brand names and drug names, and writes numbers as digits or words. Scribe v3 Live adds a library of up to ten thousand terms, and a mode for speakers who switch languages mid-sentence.

Slide text

Scribe v3 learns your product names and your number style

Scribe v3 1,000 key terms Brand names, acronyms, drug names

Scribe v3 Digits or words “42” or “forty-two”, your choice

Scribe v3 Live 10,000-term library Used only when it hears a likely match

Scribe v3 Live Multilingual mode Follows a switch of language mid-sentence

elevenlabs.io/blog/scribe-v3 · elevenlabs.io/docs/overview/models

Scribe v3 Live is in ElevenAgents today. In ElevenAPI, moving from Scribe v2 is a one-line model ID change, and Scribe v2 stays available. The official SDKs don’t support turn prediction yet.

Slide text

Scribe v3 Live is in ElevenAgents today

Both models in ElevenAPI, a one-line model ID change · Scribe v2 stays available · SDKs lack turn prediction for now

elevenlabs.io/blog/scribe-v3 · elevenlabs.io/docs/overview/models

The upgrade steps are under Upgrading from Scribe v2, on the Models page of the ElevenLabs docs. If I ran a voice agent, I’d switch it to Scribe v3 Live first, because turn prediction stops cutting off callers mid-number.

Slide text

Upgrade steps: Upgrading from Scribe v2, on the Models page of the ElevenLabs docs

Running a voice agent? Switch it to Scribe v3 Live first, so callers stop getting cut off.

elevenlabs.io/docs/overview/models