Faster Whisper Large v3 API

Send a recording and read the transcript back, in any of the 100 languages the model card lists, from $0.024 per hour of audio.

Get API key Read the docs

Audio to text
100 languages
1.55B parameters
MIT licence

from openai import OpenAI

client = OpenAI(
    base_url="https://api.modelsheep.com",
    api_key=API_KEY_MODELSHEEP,
)

response = client.audio.transcriptions.create(
    model="Systran/faster-whisper-large-v3",
    file=open("audio.mp3", "rb"),
)

print(response.text)

How to use the Faster Whisper Large v3 API

1
Create a ModelSheep account and copy an API key from the dashboard.
2
Point your client at the base url https://api.modelsheep.com and set the model to Systran/faster-whisper-large-v3.
3
Post the audio file to /audio/transcriptions as a multipart form rather than a JSON body, and read the transcript from the text field of the reply.

What Faster Whisper Large v3 is

Faster Whisper Large v3 is OpenAI's Whisper large-v3 speech recognition model, converted by SYSTRAN to the CTranslate2 format. The weights are the ones OpenAI trained on 1 million hours of weakly labelled audio and 4 million hours of pseudo-labelled audio. OpenAI reports 10 to 20 percent fewer errors than large-v2 across a wide range of languages. SYSTRAN reports that the CTranslate2 build runs up to four times faster than the reference implementation at the same accuracy and uses less memory, so we serve this conversion rather than the original repo.

What it does well

Converts calls, meetings, or podcasts to text and detects spoken language
Translates speech in another language into English text in one pass.
Returns timestamps with the text so a transcript aligns with the recording

What it is not for

Even accuracy across all 100 languages. OpenAI reports strong results in
A transcript no one reads. Weak supervision makes the model write text that

Calling it

Post the audio as a multipart form to /audio/transcriptions, with the file in file and Systran/faster-whisper-large-v3 in model. Set language to skip detection, or leave it out and let the model read the language off the first window. The 0K token figure is the decoder limit for one 30 second window, not a document window, so it puts no cap on how long a recording can be.

Systran/faster-whisper-large-v3 /audio/transcriptions

SDKs and languages

Faster Whisper Large v3 answers on an HTTP endpoint, so any language that can post JSON reaches it. These are the ones the box above writes out.

OpenAI SDK: PythonTypeScript
HTTP: cURLPythonTypeScript

Specifications

Author	Systran
Category	audio
Accepts	audio
Returns	text
Total parameters	1.55B
Active parameters	dense, all of them
Score	not rated
Languages	100Counted from the language codes the model ships with. OpenAI documents strong results on about ten of them, and lower accuracy with more invented text on the rest.
Precision served	FP16
Context	0K tokens
Vector	n/a
Pulls	973.5K in the last 30 daysWeights pulled from Hugging Face, read on 2026-09-24. It counts the repository, not this API.
License	MIT

Price

Input

$0.024

per hour of audio

Output

n/a

this model returns a vector

Against OpenRouter

11%

below the cheapest of 1 provider

The cheapest OpenRouter price for this model was $0.027 in and n/a out, with a median of $0.027 and n/a across the comparable providers. Read on 2026-09-17.

License

SYSTRAN publishes the converted weights under MIT, and OpenAI publishes the original weights under Apache 2.0. Both permit commercial use. No vendor terms sit on top of either.

Parameters: 24.3B
Context: 33K
Input per 1M: $0.090
Output per 1M: $0.270
Released: Jul 2025
Pulls, 30 days: 24.4K
License: Apache 2.0