ElevenLabs Scribe v3 and v3 Live: Price, Accuracy, API Setup and Limits (2026)

ElevenLabs has replaced its speech-to-text models again, and this time the realtime one can finally tell speakers apart. Scribe v3 and Scribe v3 Live launched on 10 November 2026.

The headline claims are big: 27% fewer word errors than Scribe v2 across 102 languages, and 43% fewer for live transcription. The price did not go up.

This guide covers what changed, the full price table, every API limit, working code for files and live audio, how it compares with OpenAI, Deepgram, AssemblyAI and free local options, and the catches.

What ElevenLabs Released

ElevenLabs is best known for voice generation, but its Scribe speech-to-text line has become a big part of the business. Scribe v1 arrived in February 2025, then v2 and v2 Realtime followed.

On 10 November 2026 the company published “Introducing Scribe v3 and v3 Live, our most accurate speech to text models”. Both models went live the same day.

ElevenLabs also said a version of Scribe v3 Live tuned for medical conversations is coming to ElevenAgents. The batch medical model today is still Scribe v2 Medical.

The Accuracy Numbers (and How They Were Measured)

All figures below are vendor-reported by ElevenLabs. No independent leaderboard had published Scribe v3 results when we wrote this.

Scribe v3 vs Scribe v2 (word error rate reduction)

Scribe v3 Live vs Scribe v2 Realtime

How ElevenLabs tested

• Accuracy: an internal version of the FLEURS benchmark (short, clean read speech) in 102 languages. Each language was scored separately and then averaged, so small languages count as much as English.
• Diarization: an internal set of real meetings, agent calls and voice notes in 16 languages. ElevenLabs says it will open-source this set “soon”.
• Rivals: charts compare against “other popular” models, but the blog does not name them.

Note: FLEURS is clean read speech. Real phone calls, noisy rooms and heavy accents usually score worse, so treat these as relative gains, not the error rate you will see on your own audio.

ElevenLabs also gave one real-world number. Dutch energy firm Eneco saw its ElevenAgents voice agent get postcodes right on the first try 40% more often, and house numbers 15% more often.

Every New Feature, Batch vs Live

Tip: Keyterms are the cheapest accuracy win for brand names and jargon. Put product names, people’s names and acronyms in the list instead of fixing them after transcription.

Pricing and Included Hours by Plan

Scribe v3 costs the same as Scribe v2. Billing is per hour of audio sent, and silence counts.

Each ElevenLabs plan includes a block of transcription hours. The pricing page lists these allowances:

Warning: The API reference lists extra surcharges: transcript_edit adds 30% (10-second minimum), keyterms add 20% (20-second minimum above 100 terms) and detect_speaker_roles adds 10%. Multichannel bills every channel at full length.

Example: a one-hour podcast with diarization costs $0.20. Add keyterms and the same hour costs $0.25.

API Limits You Need to Know

• File size: up to 5 GB per upload (2 GB for the old cloud_storage_url field).
• Length: up to 15 hours; in multichannel mode, all channels combined under 15 hours. Minimum 100 ms.
• Speakers: num_speakers from 1 to 32.
• Channels: up to 5 in multichannel mode.
• Formats: AAC, AIFF, OGG, MP3, Opus, WAV, FLAC, M4A, WebM audio, plus MP4, AVI, MKV, MOV, WMV, FLV, MPEG and 3GPP video.
• Remote files: source_url accepts hosted files, including YouTube and TikTok links.
• Speed-up for long files: files over 8 minutes are split into up to 4 parts and transcribed in parallel.
• Languages: 90+ on the docs page (the blog benchmarks 102).

How to Use Scribe v3, Step by Step

Without code (ElevenLabs app)

1. Sign in at elevenlabs.io. A free account is enough to try it.
2. Open Speech to Text in the left menu. Scribe v3 is now the default model.
3. Upload a file or paste a link. Set the language or leave auto-detect on.
4. Turn on speaker labels if more than one person speaks, and add key terms.
5. Export the transcript as text or subtitles once it finishes.

With the API (Python)

1. Create an API key in your ElevenLabs settings and save it as ELEVENLABS_API_KEY.
2. Install the SDK: pip install elevenlabs
3. Send a file with model_id set to scribe_v3, as below.

import os
from elevenlabs.client import ElevenLabs

client = ElevenLabs(api_key=os.environ["ELEVENLABS_API_KEY"])

with open("meeting.mp3", "rb") as audio:
    result = client.speech_to_text.convert(
        file=audio,
        model_id="scribe_v3",
        diarize=True,              # label speakers
        num_speakers=3,            # 1-32, or leave out
        tag_audio_events=True,     # (laughter), (applause)
        keyterms=["Freakyjolly", "ComfyUI", "RTX 5080"],
    )

print(result.text)
for w in result.words[:10]:
    print(w.speaker_id, w.start, w.text)

With cURL

curl -X POST "https://api.elevenlabs.io/v1/speech-to-text" \
  -H "xi-api-key: $ELEVENLABS_API_KEY" \
  -F file=@meeting.mp3 \
  -F model_id="scribe_v3" \
  -F diarize=true \
  -F transcript_edit="Write every date as YYYY-MM-DD"

Useful parameters

Moving from v2 is just replacing scribe_v2 with scribe_v3. Nothing else in the request has to change.

Realtime Transcription With Scribe v3 Live

Scribe v3 Live streams audio over a websocket and returns text as people speak. ElevenLabs’ own example uses the JavaScript SDK and the new turn prediction:

import { ElevenLabsClient } from "@elevenlabs/elevenlabs-js";

const elevenlabs = new ElevenLabsClient();
const connection = await elevenlabs.speechToText.realtime.connect({
  modelId: "scribe_v3_live",
  commitStrategy: "turn_prediction",
});

• turn_prediction: decides when someone has finished speaking using the words, not just a pause. This helps voice agents avoid cutting people off.
• diarize=true: labels each speaker in the live transcript.
• Browser apps: use a single-use token instead of exposing your API key; tokens expire after 15 minutes.

If you are building a full voice agent instead of captions, compare it with our OpenAI Realtime API guide, which handles speech in and speech out in one session.

Scribe v3 vs OpenAI, Deepgram, AssemblyAI and Local Whisper

Rival prices below come from Soniox’s public comparison page (updated 23 September 2026). Soniox sells a competing API, and it adds paid add-ons to some vendors to match features, so treat it as a rough guide.

Which one should you pick?

• Podcasters and YouTubers: Scribe v3 in the ElevenLabs app. The free hours cover a few episodes and speaker labels are good.
• Voice agent builders: Scribe v3 Live if you need speaker labels or mixed-language callers; it is already wired into ElevenAgents.
• Lowest cloud cost at volume: compare Soniox and xAI’s transcribe models, which list lower per-hour prices.
• Private or offline audio: run it yourself. See how to run Whisper locally or NVIDIA Nemotron 3.5 Transcribe.

Privacy, Compliance and Data Residency

• Default region: US. Enterprise customers can use isolated EU, India or Singapore environments.
• Zero Retention Mode: deletes audio and text once a request finishes. In the API it is enable_logging=false, and it is for enterprise customers only.
• Certifications: SOC 2, ISO 27001 and PCI DSS Level 1.
• Healthcare: HIPAA-eligible once you sign a Business Associate Agreement with ElevenLabs sales.
• Private deployment: offered for large customers.

Note: On normal self-serve plans, your audio is logged. Do not upload recordings you have no right to share, such as calls recorded without consent.

Known Catches and Troubleshooting

• 422 Unprocessable Entity: usually a bad parameter combination. transcript_edit cannot be used with entity_detection, entity_redaction or use_multi_channel.
• diarization_threshold ignored: it only works when diarize=true and num_speakers is empty.
• Keyterm errors: each term must be 50 characters or less, five words or less, and cannot contain < > { } [ ] or backslash.
• Too many speakers detected: raise diarization_threshold toward 0.4, or set num_speakers.
• Bills higher than expected: silence is billed, multichannel bills each channel, and add-ons stack. Trim long silences before upload.
• Realtime cuts people off: switch commitStrategy to turn_prediction instead of a silence timer.
• Webhook never arrives: set up the webhook in workspace settings first, then send webhook=true.

My Take

The most useful change is not the 27% figure. It is realtime diarization at the same $0.22 price. Live captions for meetings and support calls usually needed a second pass to label speakers.

The accuracy claims need independent checks. FLEURS is clean read speech, the rivals are unnamed, and the diarization set is not public yet. I would wait for Artificial Analysis or Open ASR numbers before switching a large pipeline.

For anyone already on Scribe v2, though, there is no reason to wait. Same price, same API, one changed string. Test it on 30 minutes of your own audio and compare.

If you are shopping more widely for voice tools, our ElevenLabs alternatives guide covers cheaper text-to-speech and voice cloning options.

FAQ

When was Scribe v3 released?

ElevenLabs released Scribe v3 and Scribe v3 Live on 10 November 2026, on the API and in the ElevenLabs app.

How much does Scribe v3 cost?

$0.20 per audio hour for Scribe v3 and $0.22 per hour for Scribe v3 Live, the same as v2. Keyterms and transcript edits cost extra.

Is Scribe v3 free to try?

Yes. The free plan includes about 4 hours 30 minutes of batch transcription and 2 hours 30 minutes of live transcription.

What is the model ID for Scribe v3?

Use scribe_v3 for files and scribe_v3_live for realtime streams.

How many speakers can Scribe v3 identify?

Up to 32 speakers in one file. You can set num_speakers or let the model decide.

How many languages does Scribe v3 support?

ElevenLabs benchmarked 102 languages, and the docs list 90+ supported languages with auto-detection.

Can Scribe v3 transcribe YouTube videos?

Yes. The API’s source_url field accepts hosted files, including YouTube and TikTok links. Only transcribe content you have the rights to use.

Is Scribe v3 better than Whisper?

ElevenLabs says it beats every batch model it tested, but it did not name them. Whisper is free and runs offline, so it still wins on privacy and cost.

Sources

• ElevenLabs: Introducing Scribe v3 and v3 Live (10 Nov 2026)
• ElevenLabs API pricing (Scribe rates and plan hours)
• ElevenLabs docs: Transcription capabilities and limits
• ElevenLabs API reference: Create transcript
• NoCode MBA AI release tracker: Scribe v3
• Soniox: Speech-to-text API pricing comparison (23 Sep 2026)