ElevenLabs Scribe v3 Live Migration Guide: Turn Prediction, Keyterms and Live Diarization

The risk in a Scribe v3 Live migration is how easy it looks. You change one model ID, the list price stays put, and live calls start running through a new speech model without a code review flagging anything. If the new model handles your accents, product names or turn-taking worse than the old one, nothing fails loudly. The transcripts just drift, and an agent starts mishearing account numbers or cutting callers off. Below is what changes, how the new features behave, and a scorecard for deciding on your own evidence.

What Scribe v3 and Scribe v3 Live Change for Scribe v2 Users

On the FLEURS benchmark across 102 languages, ElevenLabs reports that Scribe v3 lowers word error rate (WER) from 16.94% to 12.38%, which it describes as 27% fewer word errors than Scribe v2. For realtime, the launch blog says that across 102 languages Scribe v3 Live reduces semantic WER by 43% and standard WER by 31% compared with Scribe v2 Realtime. Before comparing semantic WER with other vendors' figures, we suggest checking how each vendor defines the metric.

Pricing is unchanged. Scribe v3 is listed at $0.20 per hour of audio, the same as Scribe v2, and Scribe v3 Live is listed at $0.22 per hour, the same as Scribe v2 Realtime. Base cost gives you no reason to move and no reason to wait. Paid add-ons are a separate question.

How to Migrate: Swapping the Model ID

A developer on the timbal-ai project opened a GitHub issue to adopt the new models. They reported that both models were already available on their API key and worked through their existing voice STT adapter with no code changes. They were also candid about how little that proved. They described it as a smoke test on a single clean synthetic clip, with latency figures that included the WebSocket connection time, and said plainly that it was not a benchmark.

So a clean integration probably needs a configuration change, not a rewrite. Whether accuracy improved is a different question.

We suggest treating the official Realtime API reference as the authority on how to pass the model ID and session options, rather than third-party snippets, because parameter names can change between releases.

Run v2 and v3 side by side

Don't hard-code the new ID. Put the model choice behind a feature flag or per-tenant setting, so some sessions use Scribe v3 Live while the rest stay on Scribe v2 Realtime. That gives you a live comparison and a rollback measured in seconds. Log the model ID with every stored transcript. If you already trace LLM and agent calls, as in our OpenTelemetry GenAI tracing playbook, add the speech model ID as an attribute so a regression points back to the model behind it.

Post-swap checklist

• [ ] Session settings carry over: audio format, sample rate, language and optional features.
• [ ] You know which commit strategy is active.
• [ ] Agent logic and stored records use committed text, not partials.
• [ ] Error handling covers dropped connections, authentication failures and a fallback to v2.
• [ ] Dashboards split latency and error rates by model.

Using commit_strategy=turn_prediction in Voice Agents

In a realtime session the server streams partial transcripts that keep changing as audio arrives. A commit finalizes the current segment. If your agent waits for committed text before responding, commit timing shapes how the whole conversation feels. Late commits make the agent seem slow. Early ones make it talk over people who haven't finished.

Scribe v3 Live supports commit_strategy=turn_prediction. According to the commit strategies guide, with this strategy the server commits the current segment once a partial transcript reaches a turn_probability of 0.9 or higher. Put simply, the server predicts whether the speaker is done and commits when it's confident.

We suggest comparing turn prediction against whichever commit approach you use today, on the same recordings, rather than assuming it will be better for your callers.

While testing, log each turn_probability value and commit event with timestamps, next to what the speaker actually said. Look for commits that fired mid-thought and commits that arrived long after the speaker stopped. Start with hesitant speakers and people reading out long numbers, since they expose problems fastest. For a structured way to score this, see our same-clock map for turn-taking evaluation.

The guide describes the threshold as server behavior, so we suggest not planning around tuning it. Spend the effort on how your agent recovers when a user keeps talking after a commit.

Keyterms and Keyterm Libraries for Domain Vocabulary

Generic speech models tend to miss the words a business cares about most: product names, street names, postcodes, policy numbers, drug names and internal jargon. Teams may find this is where domain vocabulary features earn their keep, provided they test them on their own errors.

ElevenLabs has published an example from its customer Eneco. According to ElevenLabs, in real customer calls the agent recognized Dutch postcodes correctly on the first try 40% more often, and house numbers 15% more often. Read that as a vendor illustration, not a forecast for your own vocabulary.

For larger vocabularies there's keyterm_library. Data Studios notes that the documentation says a library can hold up to 10,000 terms, while another line on the same page mentions 100,000. If your design depends on a very large library, confirm the real limit with ElevenLabs first.

Keyterm prompting is a paid add-on, listed at $0.05 per hour on the API pricing page, so curate the library rather than adding every term you have. A simple loop works well:

1. Pull real call transcripts and mark the terms the model got wrong.
2. Group the errors by type, such as names, alphanumeric codes, addresses and product terms.
3. Add the terms that fail often and affect the business outcome.
4. Re-run the same audio with and without the library.
5. Review the library on a schedule and drop terms that no longer help.

If callers switch languages mid-conversation, we suggest testing on recordings that actually contain code-switching, and checking the models documentation and the API reference for how any multilingual options treat your language pairs rather than guessing.

Live Diarization: What It Enables and How Far to Trust It

Live speaker labels make it possible to build running meeting notes with names attached, agent-assist tools that react only to the customer, and call QA that separates agent from customer while the call is still happening.

ElevenLabs reports that on the English portion of its 16-language real-world dataset, Scribe v3 Live has a diarization error rate of 3.1%, compared with 5.9% to 7.3% for the other realtime models it tested. For batch, the company says Scribe v3 makes 17% fewer diarization errors than Scribe v2 on recordings with three or more speakers, a common difficulty for diarization.

Hold those numbers loosely. AlphaSignal notes that the diarization results come from a separate internal dataset covering meetings, agent calls and voice notes in 16 languages, and that these internal results have not been independently verified. ElevenLabs says it plans to open-source that dataset.

Test the hard cases: overlapping speech, voices that sound alike, people who join late and long stretches where one person dominates. Our speaker-to-transcript handoff checks for overlapping speech apply here too.

Scribe v3 Live Pricing, Add-ons and Predictable Bills

*Table note: add-on rates are the listed API rates. Confirm the realtime add-on rates that apply on your plan.*

Add-ons are where the bill grows. Data Studios estimates that add-ons can raise the batch bill by 60% and the live bill by about 91%, and it advises enabling them only where custom vocabulary or automated formatting removes manual work.

Two habits help. First, enable add-ons per workflow, not across the whole account. A claims line that hears policy numbers all day may justify keyterms, while a general meeting recorder may not. Second, track cost per useful hour: audio that produced a transcript usable without manual correction. A cheaper hour that still needs human cleanup isn't really cheaper.

Compliance and Data Retention Options

For Enterprise customers, ElevenLabs offers Zero Retention Mode. The documentation says that in this mode, most data in requests and responses is deleted immediately once the request completes.

Note the word 'most.' Read the page for what's excluded and record it in your data protection assessment. If you handle health information, ask ElevenLabs directly whether a HIPAA Business Associate Agreement is available on your plan and whether it covers the realtime features you plan to use.

Check your own systems as well. Vendor deletion doesn't help much if your application, logs or analytics warehouse keep raw audio and transcripts indefinitely, and that includes the debug logs your evaluation will create.

Before You Test: Why Vendor Numbers Aren't Enough

Every accuracy figure in the launch materials is vendor-reported, and ElevenLabs' own pages don't fully agree. The blog cites a 43% reduction in semantic WER across 102 languages, while the models documentation says that on FLEURS across 73 languages, Scribe v3 Live has a 44% lower semantic WER than Scribe v2 Realtime. Quote one source precisely instead of mixing them. Scores are only comparable when you know the protocol behind them, as we argue in why benchmark scores need protocols.

The benchmark has limits too. As AlphaSignal points out, because FLEURS contains clean, read speech, its WER results don't establish performance on noisy calls or heavily overlapping conversations.

Averages also hide wide variation between languages. ElevenLabs reports Live gains of 77% for Greek, 54% for Danish, 45% for Croatian and 36% for Hindi. For batch, it reports 42% for German, 35% for Spanish and 23% for English. If your traffic is mostly English, your gain may land well below the multilingual average.

Independent signals exist, but they're snapshots. DEV Community reported that on the Artificial Analysis leaderboard for recorded audio, read on November 12, Scribe v3 ranked second among 64 models at 1.9% word error, behind Alibaba Cloud Fun-Realtime-ASR-preview at 1.7%. A recorded-audio ranking says nothing direct about realtime behavior on your traffic. That's why the test below runs on your own audio.

The Swap-Safe Scorecard: Testing Scribe v3 Live on Your Own Audio

The Swap-Safe Scorecard is a seven-step plan that ends in one table: the same audio, both models, six rows. Fill in the two result columns, and agree the decision rule with stakeholders before anyone sees results.

Decision rule: v3 Live wins only if it improves the rows tied to your business outcome without making interruptions or cost per useful hour worse.

Step 1: Build a representative evaluation set. Sample real traffic, including noisy calls, varied accents, crosstalk, domain terms and every language you serve. Get proper consent and apply your retention rules to the set itself.

Step 2: Write references and lock the rows. Create reference transcripts and speaker labels for a subset, and fix the metrics before testing starts.

Step 3: Run both models side by side. Send the same audio through both with identical settings. Then repeat with keyterms on, so you can separate the model's gain from the add-on's.

Step 4: Translate WER into business impact. ElevenLabs offers a useful frame: a one-hour meeting at 150 words per minute is about 9,000 words, so the gap between a 4% and a 2% WER is 180 wrong words, three for every minute someone was talking. Run the same arithmetic for your own call lengths, and weight errors in the entities that drive outcomes.

Step 5: Measure end-to-end agent behavior. With turn prediction on, track interruptions, response lag and recovery when users keep talking. A cleaner transcript that makes the conversation feel worse is not a win.

Step 6: Treat public benchmarks as a sanity check. ElevenLabs reports that Scribe v3 lowers WER on Earnings-22, which is accented English financial calls, from 5.99% to 3.91%, and on VoxPopuli, which is European Parliament speech, from 2.13% to 1.71%. Check the launch blog to confirm whether these describe the batch or the Live model before using them as a realtime baseline.

Step 7: Roll out gradually. Move traffic by tenant or percentage, watch the scorecard rows on live dashboards, and keep a one-switch rollback to v2 until the new model has held up for a full business cycle.