ElevenLabs Scribe Speech-to-Text for Voice Agents on Phone Calls

ElevenLabs Scribe v2 Realtime is the streaming version of ElevenLabs' speech-to-text, built for live calls. It accepts 8 kHz μ-law phone audio and covers 90+ languages. It costs $0.39 an hour, and ElevenLabs quotes about 150 ms median latency. Dograh, the open-source AI voice agent platform we build, runs it on your key, self-hosted or on Dograh Cloud.

Key Takeaways

• Scribe v2 Realtime takes 8 kHz μ-law phone audio as it arrives.
• The 2.2% accuracy score belongs to batch Scribe v2, not the realtime model.
• One ElevenLabs key in Dograh covers both speech-to-text and the voice.

Speech-to-text decides what your agent thinks the caller said, so it is worth knowing well. We explain what Scribe is, how its realtime model treats phone audio, its latency and price, how it compares on accuracy with Deepgram and other options, and how it runs in Dograh.

Two Scribe models, one for live calls

ElevenLabs sells two Scribe speech-to-text models, and only one fits a live call. Scribe v2 transcribes recorded files, with speaker labels for up to 32 speakers and keyterm prompting for up to 1,000 terms, at $0.22 an hour. Scribe v2 Realtime is the streaming version. It takes audio in small chunks while the caller is still talking and sends text back as it goes, which is what a voice agent needs.

The realtime model comes with word-level timestamps, entity detection across 65 entity types with redaction, and transcript editing from plain-language instructions. Speaker labels and keyterms belong to the batch model, though ElevenLabs also sells keyterm prompting as a realtime add-on. Both models cover 90+ languages and can detect the language on their own.

How Scribe v2 Realtime handles phone audio

Phone audio is where a speech-to-text model earns its place. Most phone calls still reach the agent as 8 kHz audio, while a browser demo runs at 16 kHz, and the narrower band blurs the sounds that tell "fifteen" from "fifty". Scribe v2 Realtime accepts PCM audio from 8 kHz to 48 kHz and μ-law, the encoding Twilio and most North American phone lines send, so a phone stream can go in as it arrives.

Accepting the format is a different thing from hearing your callers well. How much accuracy drops on phone audio depends on the model and on the phone path, so test it on recordings of your own calls, with your own names and numbers. Our explainer on how the voice AI pipeline works shows what one misheard word does to the rest of the turn.

Scribe also decides when a piece of speech is final. Its voice activity detection commits a segment when the caller goes quiet, or your code can commit segments by hand. That committed transcript is what the language model answers, so the commit timing feeds straight into how fast the agent replies.

Latency, the vendor figure and your own

ElevenLabs quotes a median of about 150 ms from speech to a committed transcript for Scribe v2 Realtime. ElevenLabs' marketing gives 30 to 80 ms in its launch post for agents and under 100 ms for the API. The three figures measure different moments under conditions you cannot see, so treat them as a starting point.

Speech-to-text is usually the slowest stage of a voice agent's turn, slower than the language model, and the whole turn needs to stay under about 800 ms to feel natural. So measure Scribe inside your own agent, on real calls from your region. In Dograh, each call's trace shows the speech-to-text, AI and voice time for every turn, so comparing Scribe with another provider comes down to reading two traces.

Accuracy and languages on calls

The best independent score belongs to the batch model. On the Artificial Analysis Word Error Rate Index, version 2, which runs on recorded files, Scribe v2 scores 2.2%, against 4.0% for OpenAI's GPT-4o Transcribe and 5.2% for Deepgram Nova-3. The index weights conversational agent speech at half and fills the rest with parliament speeches and earnings calls.

Scribe v2 Realtime, the model an agent runs, is not on that board. A streaming model hears less of the sentence before it commits, so treat the batch score as a hint for your shortlist and let your own recordings decide.

For languages, Scribe v2 Realtime covers 90+ and can detect the language itself. Concurrency depends on your ElevenLabs plan. Free allows 6 realtime streams, Starter 9, Creator 15, Pro 30, and Scale and Business 45, with more on Enterprise. Each live call holds one stream, so your plan caps how many calls can be transcribed at once.

A fair test takes an afternoon. Pick 20 to 30 recorded calls from your own line, write out what was really said once by hand, and run each provider over the same files. Count the errors that change what the agent does, such as names, dates, amounts and account numbers, and ignore the filler words. That count tells you more than any leaderboard row.

Open Source Alternative to Vapi / Retell

Self-hosted voice agent platform — no per-minute fees

dograh-hq/dograh

Star on GitHub

What Scribe v2 Realtime costs

ElevenLabs' API pricing puts Scribe v2 Realtime at $0.39 an hour, about $0.0065 a minute. Keyterm prompting adds $0.08 an hour, and transcript editing adds $0.12 an hour. The base rate is the same on every plan, and the plans differ in how many hours they include, from about 2.5 hours on the smallest to about 2,538 on the largest.

Here is one example month. A line takes 1,000 calls of 5 minutes, so 5,000 minutes, or 83.3 hours, of speech-to-text. At $0.39 an hour that is $32.50. With keyterm prompting on, the rate is $0.47 an hour and the month is $39.17.

Speech-to-text is a small part of the bill next to the voice and the language model. Our guide to ElevenLabs pricing covers the voice side, and the same key pays for both.

Scribe vs Deepgram and other options

Deepgram is the natural comparison for Scribe on phone calls, and the two land close on price. The table prices the same 5,000 minutes, with a self-hosted open model for teams that want the audio kept in-house.

Deepgram's pricing gives more streams on pay as you go, while Scribe v2 leads the batch accuracy index and lists more languages. On price per minute the gap is small, so the deciding test is your own recordings. Our ElevenLabs vs Deepgram comparison covers both vendors across speech-to-text and voices.

For a wider field, including AssemblyAI, Speechmatics and Soniox, see our roundup of the best speech-to-text APIs for voice agent calls.

Running Scribe and ElevenLabs voices in Dograh

ElevenLabs works in Dograh for both jobs, speech-to-text with Scribe v2 Realtime and the voice, on your own key. Dograh is the orchestration layer that runs the call flow, tools and telephony, and the speech and language models are a separate layer you choose, here both from ElevenLabs. Add your ElevenLabs key under Providers, then pick ElevenLabs for speech-to-text and for the voice in a bring-your-own-key configuration on the Models page.

In the speech-to-text settings, the model is scribe_v2_realtime and the language defaults to English, or "auto" to let ElevenLabs detect it. Dograh streams each call at its own sample rate and lets Scribe's voice activity detection commit every segment. If your ElevenLabs account uses a data residency endpoint, such as the EU one, set it as the base URL.

For the voice, the models run from Flash v2.5, the default, up to Eleven v4 Turbo, and any voice ID works, cloned voices included. Pre-recorded clips can play for the lines the agent says most, with the ElevenLabs voice as the fallback for everything else.

Two settings work with other providers. The Dictionary passes hard words to Deepgram, Dograh's own models, AssemblyAI, Soniox and Speechmatics, so with Scribe, check brand and product names on recorded calls. Language Hints work with those Dograh models and Deepgram Flux. Your own key also means the caller's audio goes to ElevenLabs, so calls stay in-house only with open-weight models on your own servers.

Self-hosted Dograh has no per-minute platform fee, but servers, model keys and phone minutes still cost money. Dograh Cloud charges a 1¢-a-minute platform fee. Either way, ElevenLabs bills Scribe and the voice to your own account at its API rates.

Start with twenty recorded calls from your own line. Run them through Scribe v2 Realtime and one other provider, read the traces, and keep the one that gets your callers' names right.