ElevenLabs Scribe v2: The Most Accurate Speech to Text Model Explained

ElevenLabs Scribe v2 is the company’s most accurate speech-to-text model, transcribing audio in 99 languages with word-level timestamps, speaker diarization, and audio-event tagging. Scribe v2 Realtime streams transcription with under 150 milliseconds of latency. The Free plan includes 12 transcription minutes per month (checked 4 Nov 2026).

What Scribe v2 is

Scribe v2 is ElevenLabs’ second-generation speech-to-text model, and the official page calls it the most accurate speech-to-text model available. Upload audio or video and it returns a transcript with word-level timestamps, so every word is pinned to its exact moment. It also tags audio events (laughter, applause, pauses), detects up to 32 speakers in one file, and identifies entities like names and places.

Our ElevenLabs speech-to-text guide covers the general transcription workflow. This post is about the v2 model specifically: what changed, what it costs, and when the Realtime variant is the right pick.

Scribe v2 vs Scribe v2 Realtime

There are two flavors. Scribe v2 is the batch model: upload a file (MP3, WAV, MP4, MOV up to 3GB) and get the full transcript back. Scribe v2 Realtime is the streaming model: it transcribes live audio with under 150 milliseconds of latency, using a streaming-first architecture with Voice Activity Detection. Realtime supports over 90 languages and runs live in the API.

Pick batch v2 for files you already have: recorded interviews, podcast episodes, video archives. Pick Realtime for live use: meeting captioning, call transcription, voice agents that need to hear and respond. Realtime’s sub-150ms latency is what makes it usable for conversation; batch v2’s deeper analysis is what makes it more accurate on difficult audio.

The features that set v2 apart

• Keyterm prompting (up to 1,000 words). Feed the model a list of names, terms, and jargon before transcribing, and it will recognize them correctly instead of guessing. This is the single biggest accuracy lever for technical content.
• Dynamic audio tagging. The transcript marks non-speech events: laughter, applause, music, pauses. You get a readable record of what happened, not just the words.
• Speaker and entity detection. Up to 32 speakers are separated and labeled, and entities (people, places, organizations) are identified. Multi-person interviews come back structured.
• PII redaction. Personally identifiable information can be redacted from transcripts, which matters for customer calls and any workflow with privacy requirements.
• 99 languages. Transcription covers 99 languages, with Realtime streaming in over 90 of them.

How to transcribe with Scribe v2, step by step

1. Prepare the audio (0 credits). Export or record your audio. MP3 and WAV for audio-only; MP4 and MOV for video. Files can be up to 3GB. Clean audio transcribes far better than noisy audio, so do basic noise reduction first if you can.
2. Write your keyterms (0 credits). List up to 1,000 words of names, product terms, and jargon the model must get right. This step takes five minutes and prevents most mis-transcriptions of unusual words.
3. Upload the file (0 credits). Drop the file into the Scribe interface or send it through the API. The upload itself costs nothing; transcription minutes are what count.
4. Transcribe (minutes from your plan). The Free plan includes 12 minutes per month; Starter includes 1 hour 31 minutes; Creator 6 hours 7 minutes; Pro 30 hours 18 minutes. A 45-minute interview consumes 45 minutes.
5. Review speakers and tags (0 credits). Check the speaker labels (up to 32 supported) and audio-event tags. Fix any speaker merges before exporting; it is faster now than after.
6. Export (0 credits). Download the transcript with word-level timestamps for captions, or as plain text for articles and show notes.

Scribe v2 pricing and plan minutes

Prices checked 4 Nov 2026. Through the API, extra transcription is billed at $0.22 per additional hour. Concurrent transcriptions scale with the plan: 8 on Free, 12 on Starter, 20 on Creator, 40 on Pro. If you transcribe more than a few hours a month, the Creator plan’s 6 hours 7 minutes is the sweet spot; daily podcasters and call centers belong on Pro.

Choosing the right plan is mostly about minutes. If you transcribe one podcast episode a week at 45 minutes each, that is about 3 hours a month, which exceeds Starter’s 1 hour 31 minutes; weekly shows need Creator’s 6 hours 7 minutes. Teams transcribing customer calls daily should start on Creator and watch usage for a month before committing to Pro. The $0.22 per extra hour API rate means occasional overages are cheap, so there is no penalty for starting small. Remember that the Free plan’s 12 minutes reset monthly, which makes it useful for ongoing testing even after you upgrade.

Scribe v2 vs the original Scribe

The original Scribe (launched February 2025) was already strong: 99 languages, word-level timestamps, speaker diarization, audio-event tagging. V2 keeps all of that and adds the accuracy improvements plus the Realtime streaming variant with sub-150ms latency. If you used v1, the upgrade is invisible: same upload flow, better transcripts, and Realtime available when you need live captioning.

Getting the best accuracy from Scribe v2

Accuracy starts before you press transcribe. Record in a quiet room with the microphone close to the speaker; Scribe v2 is accurate, but it cannot invent words buried under noise. For interviews, use separate microphones when possible, because isolated tracks give the diarization cleaner signals to separate. Write your keyterm list before uploading, not after the first bad transcript: up to 1,000 words of names, brands, and technical terms. If the audio has heavy accents or overlapping speech, run a short test clip first to calibrate expectations. For video files, remember that Scribe transcribes the audio track; burned-in captions or on-screen text are ignored. Finally, review the transcript while listening to the audio at 1.25x speed: you will catch the few errors v2 makes, and the word-level timestamps make jumping to each correction instant.

Who should NOT use it

Skip Scribe v2 if you need text turned into speech; that is the opposite direction, and ElevenLabs’ text-to-speech models are the tool for it. Also skip batch v2 for live conversation: its accuracy comes from analyzing complete files, so use the Realtime variant (or a different live-captioning tool) when latency matters more than depth.

Frequently asked questions

What is ElevenLabs Scribe v2?

Scribe v2 is ElevenLabs’ most accurate speech-to-text model. It transcribes audio and video in 99 languages with word-level timestamps, speaker diarization for up to 32 speakers, audio-event tagging, and keyterm prompting for jargon.

What is the difference between Scribe v2 and Scribe v2 Realtime?

Scribe v2 is the batch model for uploaded files; Scribe v2 Realtime streams live transcription with under 150 milliseconds of latency. Use batch for recordings, Realtime for meetings, calls, and live captioning.

How many languages does Scribe v2 support?

99 languages for batch transcription, and over 90 for the Realtime streaming variant. Language coverage is one of the model’s headline strengths.

How much transcription do I get for free?

The Free plan includes 12 transcription minutes per month. Starter includes 1 hour 31 minutes, Creator 6 hours 7 minutes, and Pro 30 hours 18 minutes. Extra API transcription costs $0.22 per hour.

What file formats does Scribe v2 accept?

MP3 and WAV for audio; MP4 and MOV for video. Files can be up to 3GB, and concurrent transcriptions range from 8 on Free to 40 on Pro.

Can Scribe v2 handle multiple speakers?

Yes, up to 32 speakers are detected and labeled, with entity detection for names and places. Check the speaker labels before exporting, since similar voices can occasionally be merged.

Sources checked

• ElevenLabs official speech-to-text page (Scribe v2 accuracy, Realtime latency, keyterm prompting, plan minutes) – https://elevenlabs.io/speech-to-text – checked 4 Nov 2026
• ElevenLabs official Scribe launch blog (99 languages, diarization, audio-event tagging) – https://elevenlabs.io/blog/meet-scribe – checked 4 Nov 2026