This post contains affiliate links. If you purchase through them, we may earn a commission at no extra cost to you. Read our How we review page to see how we check prices and features.
ElevenLabs Scribe V2 Realtime is a speech-recognition model that transcribes live audio over WebSockets with about 150ms latency. It covers 90+ languages, returns word-level timestamps, and detects 65 entity types with optional redaction. It is API-only and bills per audio minute. Checked 5 Nov 2026.
Scribe works well alongside other realtime tools: our Forced Alignment guide covers word-level timestamps, the streaming guide explains live audio output, and the Dubbing V2 guide walks through multilingual video.
QUICK VERDICT
Best for: developers building live captioning, voice agents, meeting transcription, or any app that needs speech-to-text as it happens.
Not for: anyone who wants a no-code dashboard — Realtime runs through the API over WebSockets only.
Starting price: $6/mo for Starter with 30,000 credits/mo (checked 5 Nov 2026); Free $0 tier includes 10,000 credits/mo (checked 5 Nov 2026).
What is Scribe V2 Realtime?
Scribe V2 Realtime is ElevenLabs’ live speech-recognition model. Unlike the standard Scribe v2 API, which transcribes uploaded files, the Realtime model processes streaming audio over a WebSocket connection and returns transcripts as the words are spoken — with roughly 150 milliseconds of latency. It is built for anything where waiting for a file upload is not an option.
The model transcribes accurately in 90+ languages, produces precise word-level timestamps, and includes entity detection across 65 entity types — names, dates, locations, and more — with optional entity redaction for privacy-sensitive use cases. That combination makes it suitable for live captioning where you need both speed and structure.
Realtime is API-only. There is no dashboard button for live transcription; you connect to the WebSocket endpoint with your API key and stream audio. If you need a key, our guide to getting an ElevenLabs API key covers the steps.
How to use Scribe V2 Realtime in 5 steps
Step 1: Get an API key. Create an ElevenLabs account and generate an API key from your profile settings. The Free tier includes 10,000 credits per month for testing. Cost: 0 credits to generate the key.
Step 2: Open a WebSocket connection. Connect to the Scribe v2 Realtime WebSocket endpoint, authenticating with your API key. Configure the session: choose the language (or enable auto-detection across the 90+ supported languages), and decide whether you want entity detection, redaction, or word timestamps in the output. Cost: 0 credits to connect and configure.
Step 3: Stream audio. Send audio chunks over the WebSocket as they are captured — from a microphone, a call, or a media stream. The model accepts common audio encodings; keep chunks small and frequent for the lowest latency. Cost: billing accrues per audio minute streamed (checked 5 Nov 2026).
Step 4: Receive live transcripts. The API returns partial transcripts as words are recognized (around 150ms behind the audio), then finalizes segments with word-level timestamps and detected entities. Handle both partial and final messages in your client — partials drive live displays, finals drive records. Cost: included in the per-minute billing from Step 3.
Step 5: Close the session and store results. When the audio ends, close the WebSocket. Save the finalized transcript, timestamps, and entities to your database. If you enabled redaction, sensitive entities are already masked in the output. Cost: 0 credits to close; you paid for the streamed minutes.
Realtime vs standard Scribe v2
Both models transcribe speech, but they serve different architectures. Standard Scribe v2 takes a complete audio file and returns a transcript — ideal for recorded content. Realtime takes a live stream and returns words as they are spoken — ideal for live content.
| Scribe V2 Realtime | Scribe v2 (standard) | |
|---|---|---|
| Input | Live audio stream | Uploaded audio file |
| Protocol | WebSockets | REST API |
| Latency | ~150ms | Batch (minutes) |
| Languages | 90+ | 90+ |
| Word timestamps | Yes | Yes |
| Entity detection | 65 types + redaction | Yes |
| Best for | Live captioning, agents, calls | Recorded meetings, media files |
Choose Realtime when the transcript must appear while the audio is still playing — live streams, phone calls, voice agents. Choose standard Scribe v2 when you have a recording and want the highest-accuracy batch transcript with diarization (up to 32 speakers), which Realtime does not emphasize.
Key features in detail
~150ms latency. The model returns partial transcripts about 150 milliseconds behind the audio, which is fast enough for live captions that feel simultaneous. Finalized segments follow with word timestamps.
90+ languages. The same broad language coverage as the batch model, with automatic language detection so multilingual audio does not need manual switching.
Word-level timestamps. Every word carries start and end times, enabling karaoke-style caption highlighting, click-to-seek transcripts, and precise subtitle generation from live streams.
Entity detection and redaction. The model identifies 65 entity types — person names, organizations, dates, locations, and more. Enable redaction to mask sensitive entities in the output automatically, which matters for healthcare, finance, and customer-service recordings.
Real-world use cases
Live-streaming platforms add Realtime captions to broadcasts for accessibility and engagement. Contact centers transcribe calls as they happen, flagging entities and feeding live text to agent-assist tools. Meeting apps generate running transcripts with timestamps so participants can follow along and search later.
Voice agents pair Realtime transcription with a language model: the agent hears the user (via Realtime), reasons, and speaks back (via text-to-speech). Broadcasters generate live subtitles for news and sports. In each case the requirement is the same — words now, not after the recording ends.
Tips for reliable live transcription
Send small, frequent chunks. The ~150ms latency figure assumes you stream audio in small pieces without long gaps. Buffering several seconds before sending defeats the purpose of a realtime model — keep chunks under a second where possible so transcripts stay current.
Handle reconnects. WebSocket connections drop. Build automatic reconnection with backoff, and decide whether to resume the session or start a new one. Track which audio you already sent so a reconnect does not duplicate transcript segments in your records.
Use partials for display, finals for records. Partial transcripts update as the model refines its guess; finalized segments are stable. Show partials in live captions but only persist finalized segments with their word timestamps, so you never store text that later changes.
Pros
- ~150ms latency — transcripts feel simultaneous with the audio.
- 90+ languages with automatic detection for multilingual streams.
- Word-level timestamps enable live caption highlighting and click-to-seek.
- 65 entity types with optional redaction for privacy-sensitive audio.
- WebSocket protocol fits naturally into live-app architectures.
Cons
- API-only over WebSockets — no dashboard, so non-developers cannot use it directly.
- Requires managing a persistent connection, reconnection, and partial-vs-final message handling.
- Billed per streamed audio minute — long always-on streams add up.
- Live accuracy can trail the batch model on difficult audio; no multi-speaker diarization emphasis.
- Network interruptions break the stream — your app must handle reconnects gracefully.
Who should NOT use it
Skip Realtime if you are transcribing recordings — the standard Scribe v2 batch API is simpler and supports diarization. Skip it if you want a no-code tool, because there is no dashboard. Skip it if your app cannot maintain a WebSocket connection reliably. And if you need speaker labels on multi-speaker audio, the batch model with diarization (up to 32 speakers) is the better fit.
Frequently asked questions
What is ElevenLabs Scribe V2 Realtime?
It is a live speech-recognition model that transcribes streaming audio over WebSockets with about 150ms latency. It covers 90+ languages, returns word-level timestamps, and detects 65 entity types with optional redaction. It is API-only. Checked 5 Nov 2026.
How low is the latency?
Around 150 milliseconds from speech to partial transcript, according to the official docs. That is fast enough for live captions that feel simultaneous. Finalized segments with word timestamps follow as the model confirms each phrase.
How is Realtime different from standard Scribe v2?
Realtime processes live audio streams over WebSockets and returns words as they are spoken. Standard Scribe v2 processes uploaded audio files via REST and returns a complete transcript. Use Realtime for live captioning, calls, and agents; use standard for recordings where batch accuracy and diarization matter.
What is entity redaction?
Entity redaction automatically masks sensitive detected entities — such as person names, organizations, or locations — in the transcript output. The model detects 65 entity types. Redaction is useful for healthcare, finance, and customer-service applications with privacy requirements.
How much does Scribe V2 Realtime cost?
It bills per audio minute streamed. The Free plan includes 10,000 credits per month (checked 5 Nov 2026), Starter is $6/mo with 30,000 credits (checked 5 Nov 2026), and Creator is $22/mo with 121,000 credits (checked 5 Nov 2026). Check the official API pricing page for current per-minute Realtime rates.
Can I use Realtime without coding?
No. Scribe V2 Realtime is API-only over WebSockets with no dashboard interface. You need an API key and the ability to manage a streaming connection. Non-developers should use the standard Speech to Text dashboard tools or a no-code platform that wraps the ElevenLabs API.
Does Scribe V2 Realtime support speaker diarization?
Not as a headline feature. The standard Scribe v2 batch API supports diarization for up to 32 speakers, but the Realtime model prioritizes speed over speaker labeling. If you need to know who said what, transcribe the recording afterward with the standard API, or pair Realtime with a separate diarization step.
What audio formats does Realtime accept?
Realtime accepts streaming audio over WebSockets in common encodings — send small, frequent chunks for the lowest latency. The batch Scribe v2 API supports MP3, WAV, FLAC, M4A, OGG, OPUS, AIFF, AAC, WebM, and MP4/MOV files. Check the official docs for the current streaming encoding list, as Realtime requirements differ from file uploads.
Sources checked
- ElevenLabs documentation: Speech to Text — https://elevenlabs.io/docs/overview/capabilities/speech-to-text (checked 5 Nov 2026)
- ElevenLabs pricing — https://elevenlabs.io/pricing (checked 5 Nov 2026)
Last updated 5 Nov 2026. Prices checked on the official pricing page on that date; confirm at checkout.