OpenAI: Whisper large-v3
whisper-large-v3- Price
- From $0.000031
- Modalities
- Audio
About Whisper large-v3
Whisper large-v3 is the open-weight checkpoint of OpenAI's Whisper speech recognition model, with about 1.55 billion parameters. It is trained on roughly 1 million hours of weakly labeled and 4 million hours of pseudo-labeled audio, supports 99 languages, and uses 128 mel bins plus a Cantonese language token. OpenAI reports 10 to 20 percent fewer errors than large-v2, and the model card lists an Apache 2.0 license.
Where it works well
- Open weights that you can host yourself, keeping audio inside your own infrastructure.
- Reported 10 to 20 percent error reduction over large-v2 across a wide range of languages.
- Covers 99 languages, including a dedicated Cantonese token added in this version.
- Can produce word-level and sentence-level timestamps for alignment.
When to choose another model
- It is a full-size model, so it decodes slower and needs more memory than the turbo variant.
- It is a plain transcription checkpoint, with no speaker labels and no streaming interface.
- OpenAI's GPT-4o transcription models report better word error rates on their own evaluations.
Getting started
Create API key
Create a key in Console, then use it with every model on the platform.
Send your first request
Copy the example for your language and run it against the endpoint.
Speech to textTokenLab endpointPOST/v1/audio/transcriptionscurl -X POST "https://api.tokenlab.sh/v1/audio/transcriptions" \ -H "Authorization: Bearer sk-xxx" \ -F file="@audio.mp3" \ -F operation="stt" \ -F model="whisper-large-v3" \ -F language="en" \ -F response_format="json"
Parameters
Speech to text
| Parameter | Values and limits | Default | Notes |
|---|---|---|---|
file | Inputs: ≥ 1, ≤ 1 | Not provided | Send one audio file in multipart/form-data. Maximum file size is 25 MiB. |
language | Not provided | Not provided | This field is accepted. The current public contract does not specify a complete value range or list. No additional values are promised. The omitted-field default is not verified. |
Pricing
The Verified price applies to Verified, which costs less on most models. The Official price is the model maker's published price and applies to the more reliable Official route. Auto bills the route that completes the request.
Speech to text
per second- Official price
- $0.0000308
- Verified price
- —
- Discount
- —
| Official priceper second | Verified priceper second | Discount | |
|---|---|---|---|
| Speech to text | $0.0000308 | — | — |
Usage & activity
Success rate is the share of requests that completed. Latency is how long a full response takes; P95 means 95% of requests finished within that time.
Usage & availability
Last 24 hours- Requests
- Success rate
- P95 latency
- Total tokens
Data is based on aggregate user requests, excluding status checks.
Open in Console
Open Whisper large-v3 in Console with a prompt ready to edit or send.
Help me try whisper-large-v3 with a short audio request at /v1/audio/transcriptions. Show the result and cost.
Use cases
Self-hosted transcription
Run speech-to-text inside a private network for audio that cannot leave your environment.
Multilingual media
Transcribe broadcasts, lectures, or videos in many languages with a single checkpoint.
Cantonese content
Use the dedicated language token for Hong Kong and Guangdong audio instead of treating it as generic Chinese.
Subtitle alignment
Produce word-level or sentence-level timestamps for matching transcript text to specific video segments.
Prompt examples
Transcribe this Cantonese radio clip and keep the output in written Chinese.
Transcribe this German conference talk with sentence-level timestamps.
Transcribe this field-recording interview in Portuguese and return plain text.
FAQ
What is Whisper large-v3?
It is the open-weight, 1.55 billion-parameter Whisper checkpoint for speech recognition and speech translation across 99 languages. It was trained on 1 million hours of weakly labeled and 4 million hours of pseudo-labeled audio.
What changed from large-v2?
The input spectrogram uses 128 mel bins instead of 80, a Cantonese language token was added, and the model card reports a 10 to 20 percent error reduction across many languages.
How is it different from whisper-large-v3-turbo?
The turbo checkpoint cuts the decoder from 32 layers to 4, decoding faster with a minor quality loss. Whisper large-v3 keeps the full decoder for the best accuracy of the two.
What license does it use?
The model card on Hugging Face lists Apache 2.0, which allows commercial use and self-hosting. Check the card yourself before shipping, since licenses can differ between Whisper checkpoints.
Is it the same as whisper-1?
No. Whisper-1 is OpenAI's managed API model, while large-v3 is the open checkpoint released for download. They share the Whisper family but are separate offerings.
How much does Whisper large-v3 cost?
On TokenLab, Whisper large-v3 costs $0.0000308 per second. The pricing table above shows the full breakdown. Rates depend on the billing unit, specification, and usage. Compare matching conditions in the model's detailed pricing; a single rate does not determine the total cost.
Which endpoint should Whisper large-v3 use?
Use https://api.tokenlab.sh/v1/audio/transcriptions for Whisper large-v3. The request example below shows the matching code shape.
Which operations does Whisper large-v3 support?
Whisper large-v3 supports Speech to text. Select an operation above to see its endpoint and request example.
Compare Whisper large-v3
Sources
Reviewed Oct 2, 2026
