ElevenLabs Scribe v2 Realtime is a fast, well-priced streaming speech-to-text model: ElevenLabs quotes about 150 ms latency, 90+ languages and $0.39 per hour of audio. What it returns is text in the language that was spoken. There is no translation option on the realtime endpoint, and no live speaker labels. The LecSync Realtime Speech API costs $0.006 per audio minute ($0.36 per hour) and returns captions, a paired translation, speaker labels and running minutes on the same WebSocket.
So the choice mostly comes down to one question: do your live captions need to be translated, or to show who is talking?
Everything below comes from public pages read on October 14, 2026. For ElevenLabs: the API pricing page, the Scribe v2 Realtime product page, the models reference, the transcription overview, the realtime WebSocket reference and the realtime event reference. For LecSync: the developers page, the pricing docs and the connect docs.
The short answer
For transcription alone, Scribe v2 Realtime is a strong pick. It covers 90+ languages against LecSync's 60, quotes lower latency (about 150 ms against 0.28 s to first word), and on annual Business plans ElevenLabs lists it at $0.28 per hour or less, below LecSync's $0.36.
For live captions that someone reads in another language, LecSync does more for less. Translation, speaker labels, glossary and live minutes are in the $0.36 hourly rate. With Scribe you add a translation service yourself, and speaker labels only come from re-running the recording through the batch model after the session.
How each session works
Both APIs keep your key off the client. Your server asks for a short-lived credential, and the browser or app opens the socket with it.
With LecSync, your server calls POST /v1/realtime/connect with the audio format, a language pair and the features you want. It gets back a stream URL and a one-time token. The client pushes 16 kHz mono PCM. Transcript segments, their translations (matched by segment ID), speaker labels with millisecond timestamps (speakers: true) and sections of minutes (digest: true) all come back on that one connection.
With ElevenLabs, your server creates a single-use token, and the client connects to wss://api.elevenlabs.io/v1/speech-to-text/realtime with model_id=scribe_v2_realtime. Audio goes up as input_audio_chunk messages; text comes back as partial_transcript and then committed_transcript. A few details shape the integration:
- Commits. You either commit segments yourself or let voice activity detection do it. In manual mode the model commits on its own after about 36 seconds of audio, and the docs suggest committing every 20 to 30 seconds.
- Silence. Keep sending audio, even silent frames. The server closes the connection after 15 seconds without client messages, and that limit can't be changed.
- Startup. The docs note that the server buffers roughly the first 2 seconds of audio before it starts transcribing.
- Extras. Word timestamps, language detection, up to 50 keyterms (20 characters each), entity detection and transcript editing are query parameters. Keyterms, entity detection and transcript editing each cost extra, and editing can't be combined with entity detection.
Side-by-side facts (checked October 14, 2026)
| LecSync Realtime Speech API | ElevenLabs Scribe v2 Realtime | |
|---|---|---|
| Session setup | POST /v1/realtime/connect, then one WebSocket | Single-use token, then one WebSocket |
| List price | $0.006/min ($0.36/hr), every feature included | $0.39/hr, transcription only |
| Lower published price | Same rate at any volume | "$0.28 per hour & lower" on annual Business plans |
| Billing unit | Accepted audio, to fractions of a second | Per audio minute |
| Languages | 60 | 90+ |
| Latency (vendor figures) | ~0.28 s to first word on screen; translation +~88 ms | ~150 ms |
| Live translation | Included; any of 60 into any other, one pair per stream | Not offered on the realtime endpoint |
| Speaker labels during the session | Yes, on every final segment, with timestamps | Not listed; vendor FAQ says it "isn't a priority" for realtime |
| Custom terms | Glossary terms, forced translation pairs and domain context, included | Up to 50 keyterms, +$0.08/hr |
| Minutes / summary | digest: true, sections pushed during the session | Not offered |
| Entity detection (PII, PHI, PCI) | Not offered | Yes, +30% on the base rate |
| Audio input | 16 kHz mono PCM (v1) | PCM 8–48 kHz or μ-law 8 kHz; dual channel "not planned" |
| Concurrency | Not published | 6 (Free), 9 (Starter), 15 (Creator), 30 (Pro), 45 (Scale/Business) |
| Max session length | Not published | Limit exists (session_time_limit_exceeded), value not published |
| Endpoints / regions | 6 gateway regions | Default plus US, EU, India and Singapore residency endpoints |
| Data retention | Nothing stored unless store: true; then 30 days | Zero retention mode for Enterprise only |
| Free tier | None | Free plan available |
| SLA / compliance (public) | Best-effort, no SLA today | SLAs and SOC 2 on Enterprise; HIPAA BAA via sales |
The translation path with Scribe
ElevenLabs does translate speech, but through Dubbing, which works on uploaded audio and video. Its own FAQ says "realtime or live dubbing is not currently available". The realtime endpoint also offers transcript editing, an experimental natural-language rewrite of each committed segment. The docs pitch it for formatting dates and expanding acronyms, not as a translation feature.
For translated live captions on Scribe, you need a second service. A typical setup:
- Take each
committed_transcriptfrom Scribe. - Send it to a machine translation API or an LLM.
- Match the reply back to the right caption line, and handle timeouts and retries for both services.
Waiting for committed text keeps translations stable, but it also means they trail the speaker by a full segment. LecSync streams the translation next to the transcript and pairs the two by segment ID, so the client never has to match them. If a translation node fails, LecSync falls back automatically and the stream keeps going. The docs also say a provider failure can leave a single segment without a translation.
More on this trade-off in built-in translation vs a DIY STT + MT pipeline.
What 1,000 hours cost
ElevenLabs rates are list prices from the API pricing page. For the translation step we use Google Cloud Translation's NMT rate ($20 per million characters after the first 500,000 each month) as one example. We assume about 50,000 characters of transcript per hour of speech. An LLM or another provider will price differently.
| 1,000 hours of live captions | LecSync | Scribe v2 Realtime |
|---|---|---|
| Transcription only | $360 | $390 at list; $280 or less on annual Business |
| + live translation | $360 | ~$1,380 ($390 + ~$990 Google NMT) |
| + speaker labels (Scribe v2 batch rerun at $0.22/hr) | $360, labels arrive live | ~$1,600, labels arrive after the session |
| + custom terms (keyterms at $0.08/hr) | $360 | ~$1,680 |
The second column is calculated from public rates, not a quote. If you only need same-language captions and can commit to an annual Business plan, ElevenLabs costs less. Once translation is in scope, LecSync is cheaper by a wide margin. For more vendors, see what 1,000 hours cost with translation, speakers and summaries.
Where ElevenLabs is the better pick
- Same-language captions at volume. $0.28 per hour or less on annual Business plans, below LecSync's $0.36.
- Lower quoted latency. About 150 ms against LecSync's 0.28 s to first word. Vendors measure this differently, so test with your own audio.
- Language coverage. 90+ languages, with word error rate bands published per language for the Scribe v2 family.
- Audio options. Sample rates from 8 to 48 kHz and μ-law for phone audio. LecSync v1 takes 16 kHz mono PCM only.
- Redaction. Live entity detection for PII, PHI and PCI categories.
- Voice agents. Scribe v2 Realtime is available as an option inside ElevenLabs Agents, next to its text-to-speech models.
- Getting started. A free plan with 6 concurrent realtime sessions, a published concurrency table, data residency endpoints, and Enterprise SLAs with SOC 2.
Where LecSync fits better
- Captions in a second language. Translation streams with the transcript at about +88 ms and comes back paired by segment ID, with nothing extra to integrate or pay for.
- Who said what, while it's being said. Every final segment carries a speaker label with
startMsandendMs. See building a meeting client with speakers and timestamps. - Notes during the meeting.
digest: truepushes sections of minutes over the same socket, each with a time range in the recording. - Terms that hold in both languages. The glossary biases recognition and pins how names and product terms are translated, at no extra cost.
- Latency that doesn't drift. After a network stall, the server caps backlog at 2 seconds instead of queueing it. More in why we cap backlog at 2 seconds.
- One flat rate. $0.36 per hour for everything, with no plan to buy first and nothing charged for empty connections.
LecSync's limits are real too: 60 languages, one language pair per stream, 16 kHz mono PCM only in v1, no published concurrency quota, no free tier, no entity redaction, no compliance certifications listed and no SLA today.
FAQ
How much does ElevenLabs Scribe v2 Realtime cost?
$0.39 per hour of audio at list price on the ElevenLabs API pricing page, checked October 14, 2026. The Scribe Realtime product page says $0.28 per hour or lower on annual Business plans. Keyterm prompting adds $0.08 per hour and transcript editing $0.12 per hour. Usage is metered per audio minute.
Does Scribe v2 Realtime translate live?
No. The realtime WebSocket has no translation parameter. ElevenLabs translates speech through Dubbing, which works on uploaded files, and its FAQ says live dubbing is not available. For translated live captions you pair Scribe with a separate machine translation API or LLM.
Does Scribe v2 Realtime label speakers?
Not according to its public docs. Speaker diarization (up to 32 speakers) is listed for the batch Scribe v2 model, not the realtime one. The realtime FAQ says diarization "isn't a priority at the moment", and dual-channel input is not planned.
What are the Scribe v2 Realtime concurrency limits?
Realtime sessions per plan: Free 6, Starter 9, Creator 15, Pro 30, Scale and Business 45, with higher limits on Enterprise. The server closes a connection after 15 seconds without client messages. A maximum session time exists, but its value isn't published.
What does LecSync include in $0.006 per minute?
Realtime transcription in 60 languages, realtime translation between any two of them, audio-aligned timestamps, speaker labels, glossary and domain context, AI translation polish, live structured minutes, and optional 30-day storage of the recording and transcript. Turning features on or off doesn't change the rate.
Bottom line
Pick Scribe v2 Realtime if you need same-language transcription in many languages with very low latency, phone audio, or live redaction, especially at annual Business volume. Pick LecSync if your live captions have to be translated, attributed to speakers, or summarised as the meeting runs, all at a flat $0.36 per hour. Create a key on the LecSync developers page and open your first stream with the connect docs. Still comparing? See LecSync vs Gladia and LecSync vs Deepgram.