whisper-large-v3-turbo

whisper-large-v3-turbo

OpenAI's faster Whisper model for speech recognition and speech translation, with 809M parameters and only 4 decoding layers. For developers and anyone who transcribes audio.

Made by openai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: whisper-large-v3-turbo is a speech recognition model from OpenAI. It transcribes spoken audio into text and can also translate speech into English. It has 809 million parameters.

It is a fine-tuned, pruned version of Whisper large-v3, with the decoding layers cut from 32 to 4. The card says this makes it much faster, with a small loss in quality.

Who it suits: Developers and teams who need to turn recordings into text, and anyone who wants Whisper quality at a faster speed on their own machine.

Industry | Who uses it | How they use it | Example
Media | Journalists and editors | Transcribe interviews and recordings | Turn an audio file into text
Software | Developers | Add speech to text to an app | Run it through a Transformers pipeline
Education | Teachers and students | Caption lectures and talks | Get timestamps for each sentence

Skip it if you need the best accuracy, since the card says the speed gain costs a small drop in quality.

Where to use it:

• Python programs, through the Hugging Face Transformers pipeline
• A computer with a CUDA GPU, or just a CPU
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Transcription | Anyone with recordings | The spoken words as text
Translation to English | Multilingual teams | English text from speech in another language
Timestamps | Caption makers | Time marks for sentences or for single words

How to set it up:

• Python and pip
• The Transformers library, plus Datasets and Accelerate for the example
• A GPU is optional, since the example falls back to the CPU

1. Install the libraries.
2. Load the model and its processor.
3. Build a pipeline and pass it an audio file.

pip install --upgrade pip
pip install --upgrade transformers datasets[audio] accelerate

How to use it: Create a speech recognition pipeline, then call it with the path of your audio file. A list of files with a batch size lets it transcribe several at once.

import torch

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "openai/whisper-large-v3-turbo"

    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
)

result = pipe("audio.mp3")
print(result["text"])

Put your own value in its key setting first.

Option | What it does
pipe("audio.mp3") | Transcribes one local file
pipe(list, batch_size=2) | Transcribes several files in parallel
language in generate_kwargs | Tells the model the language when you already know it
task set to translate | Gives English text from speech in another language
return_timestamps=True | Adds sentence-level timestamps
return_timestamps="word" | Adds word-level timestamps

Title: Decoding layers
Whisper large-v3: 32 layers
Whisper large-v3-turbo: 4 layers

Example: You pass the path of an interview recording to the pipeline and print the text. Add the word-level timestamp option and you also get a time mark for each word.

For newcomers and experts: If you are new to speech recognition, run the file example first and check the text against what you hear. Then try timestamps.

If you know the field, the card says Transformers supports the Whisper decoding strategies, including temperature fallback and conditioning on previous tokens. Whisper predicts the language itself unless you pass one.

Install Transformers and friends
Load the model and processor
Build a speech recognition pipeline
Pass an audio file and read the text

Good to know:

• The model was trained on more than 5M hours of labeled audio, and the card says it generalizes across many datasets
• By default it transcribes, and a translate task turns the speech into English text
• Part of the card was written by the Hugging Face team and part comes from the original card
• Hugging Face lists 265 quantized and 683 fine-tuned versions made by other people
• License: MIT
• Cost to run: none named
• Data it reaches: the files you download and the audio you give it
• Not for: work that needs the top accuracy of the full large-v3 model

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Parameters: 809M
• Weights: about 1.6 GB in F16; about 404 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 4,969,987
• Likes: 3,524
• License: mit
• Languages: en, zh, de, es, ru, ko, fr, ja
• Based on: openai/whisper-large-v3
• Quantized versions: 267
• Fine-tunes: 686
• Tags: transformers, safetensors, whisper, automatic-speech-recognition, audio

Links

• Model: https://huggingface.co/openai/whisper-large-v3-turbo
• Files: https://huggingface.co/openai/whisper-large-v3-turbo/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download openai/whisper-large-v3-turbo

Then follow the usage in its model card. Nothing in this kit runs by itself.