whisper-large-v3

whisper-large-v3

OpenAI's speech-to-text model that transcribes audio in many languages, can translate speech into English text, and adds sentence or word timestamps. For anyone turning recordings into text.

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 is OpenAI's model for turning speech into text. It transcribes audio in many languages, and it can also translate speech into English text. The Whisper family learned from over 5 million hours of labeled audio, so it copes with many kinds of speech without extra tuning. Large-v3 keeps the architecture of the earlier large models, with two small changes. It reads spectrograms with 128 Mel frequency bins instead of 80, and it adds a language token for Cantonese.

Who it suits: It suits anyone who needs text from recordings, such as editors, subtitle makers, support teams and researchers, and developers who add speech recognition to an app.

Industry | Who uses it | How they use it | Example
Media | Podcasters and editors | Turn recordings into text | A transcript of an interview
Video | Subtitle makers | Make captions with timestamps | Sentence-level times for subtitles
Customer support | Support teams | Transcribe recorded calls | Searchable text from a call archive
Research | Linguists | Transcribe or translate speech in many languages | Spanish speech turned into English text

If you need live captions with minimal delay, you can skip it, since the card's examples work on recorded audio.

Where to use it:

• Python with the transformers library and its pipeline class
• Windows, macOS or Linux, on a GPU or a CPU, since the card's code picks whichever is available
• Apps that need a local speech recognizer

Use case | Who | What they get back
Transcription | Editors | Text from an audio file of any length
Translation | Researchers | English text from speech in another language
Captions | Subtitle makers | Text with sentence-level or word-level timestamps

How to set it up: You need these first:

• Python and pip
• The transformers, datasets and accelerate libraries
• Disk space for the weights, about 3.1 GB in F16
• A GPU helps with speed, but the card's code falls back to the CPU

1. Upgrade pip and install the libraries.
2. Download the model files, or let transformers fetch them on first load.
3. Run the Python example, which loads the model and transcribes a sample.

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

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

How to use it: Load the model into a pipeline, then pass it an audio file or a list of files. You can set the language, ask for English translation, or ask for timestamps.

import torch
from datasets import load_dataset
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"
    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,
)
dataset = load_dataset("distil-whisper/librispeech_long", "clean", split="validation")
sample = dataset[0]["audio"]
result = pipe(sample)
print(result["text"])

Put your own value in its key setting first.

result = pipe("audio.mp3")
result = pipe(sample, generate_kwargs={"task": "translate"})
result = pipe(sample, return_timestamps=True)
print(result["chunks"])

Tool | What it does
pipeline | Transcribes audio of any length
batch_size | Transcribes several files in parallel
return_timestamps | Adds sentence-level or word-level times
task translate | Gives English text from speech in another language

Example: You point the pipeline at a recorded interview saved as an mp3 file. The model works out the spoken language and returns the text. If you ask for timestamps, each chunk of text comes with its start and end time.

For newcomers and experts: Automatic speech recognition means speech in, text out. A good first step is the card's sample code, then your own file.

The training mix has 1 million hours of weakly labeled audio, plus 4 million hours of pseudo-labeled audio that large-v2 helped to gather. Transformers supports Whisper's decoding strategies, such as temperature fallback and conditioning on previous tokens, and the card shows how to switch them on.

Load the model into a pipeline
Pass an audio file
The model detects the language and transcribes
Read the text or the timestamps

Good to know: The card reports fewer errors than the previous large-v2 model.

Title: Training audio for large-v3 in millions of hours
Weakly labeled: 1 million hours
Pseudo-labeled: 4 million hours

• Accuracy: errors drop by 10% to 20% against large-v2 across many languages, by the card
• Languages: Hugging Face lists English, Chinese, German, Spanish, Russian, Korean, French and Japanese
• Other versions: 56 quantized and 1,107 fine-tuned versions made by other people are listed
• License: Apache 2.0
• Cost to run: none named, you supply the hardware
• Data it reaches: the audio you pass in, and the sample dataset that the card's example downloads

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Parameters: 1.5B
• Weights: about 3.1 GB in F16; about 772 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 3,489,398
• Likes: 6,802
• License: apache-2.0
• Languages: en, zh, de, es, ru, ko, fr, ja
• Quantized versions: 56
• Fine-tunes: 1,108
• Tags: transformers, pytorch, safetensors, whisper, automatic-speech-recognition, audio, hf-asr-leaderboard

Links

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

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