How much does the ElevenLabs speech-to-text API cost, and how do you use it?

The ElevenLabs speech-to-text API costs $0.22 per hour of audio for its Scribe v2 model at list price ¹, which works out to about $0.11 for a 30-minute recording ². The real-time version is $0.39 per hour ³. Keyterm prompting and transcript editing are billed on top of the base rate ⁴. To use the API, you create a key in the ElevenLabs dashboard and post your audio to the speech-to-text endpoint with model_id set to scribe_v2. The rest of this page covers the prices, the request and a Python example. For a comparison with other providers, see the guide to speech-to-text APIs.

The base price is billed by the hour of audio you send

ElevenLabs bills its speech-to-text service by how long the audio is that you send for transcription ⁵. The rate is per hour, not per minute, so divide a minute count by 60 before you compare it with a per-minute price from another vendor. The comparison of list prices puts the Scribe v2 rate next to the other providers on one scale.

The vendor's pricing table is the source for the rate. It lists Scribe v2 and Scribe v2 Medical at the same price, a base rate of $0.22 per hour, and the real-time Scribe v2 Realtime model at $0.39 per hour ^{1 3}. The Medical model is priced the same as the standard model in that table ⁶.

The same table has an Hours included column, so check the plan you would buy, not only the per-hour rate ⁶.

Keyterm prompting and transcript editing cost extra

The base rate is not the whole bill. ElevenLabs bills two options separately from transcription. Keyterm prompting, which biases the model toward your names and terms, adds $0.050 per hour for Scribe v2 ⁶. Transcript editing adds $0.070 per hour ⁷.

These are list prices, so enterprise terms can differ. The speech to text overview says keyterm prompting accepts up to 1000 terms with Scribe v2 ⁸. Check that limit on the current page before you send a long list of names.

Call the endpoint with your key and a model ID

To use the API, follow these steps. The examples use the same endpoint that ElevenLabs documents in its API reference.

1. Create an API key. ElevenLabs says you create the key in its dashboard ⁹. Store it in an environment variable, not in your source code.
2. Choose the model. Set model_id to scribe_v2. The API reference lists the field as the ID of the model to use for transcription ¹⁰.
3. Post the file. Send a multipart form to https://api.elevenlabs.io/v1/speech-to-text with the audio in the file field and the key in the xi-api-key header, as the ElevenLabs API reference shows.
4. Read the JSON. The reply includes language_code and a text field with the transcript, plus word-level timing.

A Python example with the requests library looks like this:

import os
import requests

url = "https://api.elevenlabs.io/v1/speech-to-text"
headers = {"xi-api-key": os.environ ["ELEVENLABS_API_KEY"]}

with open("interview.mp3", "rb") as audio:
    response = requests.post(
        url,
        headers=headers,
        files={"file": audio},
        data={"model_id": "scribe_v2"},
    )

response.raise_for_status()
print(response.json() ["text"])

We wrote this example from the documented fields and did not run it against a live key, so run it on a short file first. The two ElevenLabs pages give different file limits. The speech to text overview says files up to 3 GB are supported ¹¹, while the API reference says the file must be less than 5.0 GB ¹². Check the current limit before you send a long recording in one request.

The vendor's accuracy claims need your own test

ElevenLabs says Scribe v2 has over 98 percent transcription accuracy ¹³. The pricing page gives no test set or method for that figure, so treat it as a claim. Independent results are more useful. In the Artificial Analysis AA-WER v2.0 benchmark, Scribe v2 had the lowest error rate, at 2.3 percent ¹⁴. That is one benchmark with its own datasets and text rules, and the guide to word error rate explains why scores from different tests do not compare directly.

Test the API on your own recordings, with the names and terms you use. The guide to speaker labels explains how to check the speaker turns that the API returns when you turn diarization on, because a speaker label is a separate feature with its own errors.

Data handling depends on the model and the plan

ElevenLabs says its Scribe v2 Medical model is HIPAA-eligible, with business associate agreements and a Zero Retention Mode ¹⁵. That is a statement about one model. For the standard model, read the data terms for your plan before you send recordings that contain personal information. The guide to whether voice to text is safe lists the questions to ask.

It depends on volume, the real-time need and the accuracy you need

• Batch files suit the standard Scribe v2 model at $0.22 per hour. Real-time captions need the real-time model at $0.39 per hour, so check whether you need streaming before you pay the higher rate.
• Heavy keyterm or editing use adds $0.050 and $0.070 per hour on top of the base rate ^{6 7}, so the add-ons can matter at large volumes.
• Local use avoids the bill. Whisper runs on your own computer, which we cover in the question on Whisper and ElevenLabs.

What to do

1. Create an API key in the dashboard and store it outside your code.
2. Run the Python example on a five-minute recording with model_id set to scribe_v2.
3. Estimate your monthly hours, and add keyterm and editing costs only if you use them.
4. Compare the transcript with a hand-checked passage, and count the word errors before you choose.

Related answers: how to use the Azure Speech-to-Text API, what OpenAI's speech-to-text API costs and whether there is a free speech-to-text API. The overview of speech recognition explains the technology.