Distil-Whisper distil-large-v3

Distil-Whisper distil-large-v3

Turns English speech into text and runs about 6.3 times faster than Whisper large-v3 while staying close in accuracy on long audio. For developers and teams who transcribe recordings.

Made by distil-whisper 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: distil-large-v3 is an English speech recognition model. It is a distilled, smaller version of OpenAI's Whisper large-v3, with 756 million parameters against 1550 million. It was tuned for accurate transcription of long recordings.

The card says it stays within 1% word error rate of large-v3 on long audio and runs 6.3 times faster. It works with Transformers and with the popular Whisper libraries.

Who it suits: Developers and teams who need fast English transcription on their own hardware, for clips and for long recordings alike.

Industry | Who uses it | How they use it | Example
Media | Podcast and video teams | Transcribe recordings to make captions and notes | Turning an interview recording into text
Software | App developers | Add speech to text inside a product | A voice note feature in an app
Education | Lecturers and students | Transcribe lectures for search and review | Writing up a recorded class
Research | Speech researchers | Compare distilled and full Whisper models | Measuring error rates on their own audio

Anyone who needs transcription in languages other than English can skip it, since the card lists English only.

Where to use it:

• Python with the Hugging Face Transformers library, version 4.39 or newer.
• Whisper cpp, Faster-Whisper and OpenAI Whisper, where the card says it is compatible.
• Transformers.js and Candle, which the card also covers.
• Windows, macOS or Linux with Python, for downloading the files.

Use case | Who | What they get back
Short clip transcription | App developers | Text for audio under 30 seconds
Long recording transcription | Media teams | Text for long audio using sequential or chunked methods
Faster pipelines | Engineers | A smaller model that replaces large-v3 with a speed gain

How to set it up: You need Python and a recent Transformers library.

• Python
• Transformers 4.39 or newer, plus accelerate and datasets
• A GPU is optional, since the example falls back to the CPU
• Audio files to transcribe

1. Upgrade pip.
2. Install or upgrade Transformers, accelerate and datasets with audio support.
3. Run the short-form example below.

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

How to use it: Load the model into a speech recognition pipeline, then pass it audio. This example transcribes a clip shorter than 30 seconds from a small test dataset.

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 = "distil-whisper/distil-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,
    max_new_tokens=128,
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
sample = dataset[0]["audio"]

result = pipe(sample)
print(result["text"])

Put your own value in its key setting first.

To transcribe your own file, pass its path to the pipeline in place of the sample:

result = pipe("audio.mp3")

Tool | What it does
pipeline | Wraps the model so you can pass audio and get text back
Short-form | Handles clips shorter than 30 seconds
Sequential long-form | Transcribes longer audio with OpenAI's sequential algorithm
Chunked long-form | Splits long audio into chunks and joins the text

Example: Pass the path of a recording called audio.mp3 to the pipeline. The result holds the transcribed text under its text key, which you can print or save.

For newcomers and experts: If you are new, a speech recognition model listens to audio and writes down the words. Install Transformers, run the short example once on the test clip, then swap in one of your own audio files.

For experts, the model was distilled so that the sequential long-form algorithm, which Whisper libraries use by default, gives accurate output. The card reports long-form error rates of 10.8 sequential and 10.9 chunked against 10.0 and 11.0 for large-v3. It beats distil-large-v2 on sequential long-form, at 10.8 against 15.6.

Install Transformers and its helpers
Load the model and processor
Build a speech recognition pipeline
Pass an audio clip or file path
Read the transcribed text from the result

Title: Speed compared with large-v3, higher is faster
large-v3: 1.0 times
distil-large-v2: 5.8 times
distil-large-v3: 6.3 times

Good to know: The points below cover limits and costs.

• Language: English only.
• License: MIT.
• Cost to run: no API key or paid service is named. A GPU speeds it up but is optional.
• Data it reaches: it reads the audio you give it, and the example also downloads a test dataset from the Hugging Face Hub.
• Not for: other languages, or long audio fed in one go without the long-form methods.

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Parameters: 756M
• Weights: about 1.5 GB in F16; about 378 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 754,155
• Likes: 386
• License: mit
• Languages: en
• Quantized versions: 11
• Fine-tunes: 24
• Tags: transformers, tensorboard, onnx, safetensors, whisper, automatic-speech-recognition, audio, transformers.js

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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