Whisper large-v3
OpenAI · openai/whisper-large-v3
Weights license: Apache 2.0 · code and components may use other terms
Quality
Word error rate · Open ASR
0–8% · lower is better · this model outlined
Source: huggingface/open_asr_leaderboard scripts/data/en_shortform.csv @ 16743c6 (2026-05-22; results as of 2026-03-26) · collected 2026-10-20 · Apache-2.0 · models without a result are not drawn (no result ≠ 0)
Word error rate · Open ASR
Published result · collected 2026-10-20 · huggingface/open_asr_leaderboard scripts/data/en_shortform.csv @ 16743c6 (2026-05-22; results as of 2026-03-26)
- WER · AMI
- 15.95 %
- WER · Earnings-22
- 11.29 %
- WER · GigaSpeech
- 10.02 %
- WER · LibriSpeech clean
- 2.01 %
- WER · LibriSpeech other
- 3.91 %
- WER · SPGISpeech
- 2.94 %
- WER · TED-LIUM
- 3.86 %
- WER · VoxPopuli
- 9.54 %
- RTFx (speed on the leaderboard's GPU)
- 146
Benchmark conditions
Open ASR Leaderboard run of the published checkpoint on the leaderboard's GPU; transcripts and references normalised (Whisper-style) before scoring. Not a quantized variant or a local hardware measurement.
English short-form, macro-average of 8 test sets: AMI, Earnings-22, GigaSpeech, LibriSpeech clean, LibriSpeech other, SPGISpeech, TED-LIUM, VoxPopuli · huggingface/open_asr_leaderboard scripts/data/en_shortform.csv @ 16743c6 (2026-05-22; results as of 2026-03-26)
Reported by the publisher
Quality × memory
No comparable quality × memory pair
Image, video and audio models have no shared quality scale, and their memory comes from the publisher's documented minimum path, so quality is not charted against memory.
Consider Whisper large-v3 from OpenAI for the speech recognition (speech to text) task (1.55B parameters).
The catalog lists the publisher's documented memory path under Runs on and has not reviewed other configurations. Actual memory depends on precision, resolution, duration and offload; consult the official source.
Runs on
Enough memory 46
A publisher documents a path within this memory, on this platform (CUDA or Apple MPS). Conditional paths need offload or quantization.
Show all 46
One memory tier short 21
The same machines with the next memory size up would have enough.
Show all 21
Hardware profiles grouped by the publisher's documented minimum memory path and platform. Not a measured run.
Measured speed
No published measurement for this checkpoint
Speed (real-time factor, or seconds per clip) is shown only when a source publishes it with hardware, build and setup. It is never estimated.
How to run
Publisher-documented memory path: Official README table, large row: ~10 GB VRAM (the table has no separate large-v3 row). Actual memory depends on precision, resolution, duration and offload; consult the official source. Official hardware evidence
Documented ways to run it
Copied as printed in the linked source (read 2026-10-18); not run by this site.
pip install --upgrade pip
pip install --upgrade transformers datasets[audio] accelerateimport torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
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 = AutoModelForSpeechSeq2Seq.from_pretrained(
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"])pip install --upgrade pip
pip install --upgrade transformers datasets[audio] accelerateimport torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
from datasets import Audio, 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 = AutoModelForSpeechSeq2Seq.from_pretrained(
model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True
)
model.to(device)
processor = AutoProcessor.from_pretrained(model_id)
dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
dataset = dataset.cast_column("audio", Audio(processor.feature_extractor.sampling_rate))
sample = dataset[0]["audio"]
inputs = processor(
sample["array"],
sampling_rate=sample["sampling_rate"],
return_tensors="pt",
truncation=False,
padding="longest",
return_attention_mask=True,
)
inputs = inputs.to(device, dtype=torch_dtype)
gen_kwargs = {
"max_new_tokens": 448,
"num_beams": 1,
"condition_on_prev_tokens": False,
"compression_ratio_threshold": 1.35, # zlib compression ratio threshold (in token space)
"temperature": (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
"logprob_threshold": -1.0,
"no_speech_threshold": 0.6,
"return_timestamps": True,
}
pred_ids = model.generate(**inputs, **gen_kwargs)
pred_text = processor.batch_decode(pred_ids, skip_special_tokens=True, decode_with_timestamps=False)
print(pred_text)pip install --upgrade pip
pip install --upgrade transformers datasets[audio] accelerateimport torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
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 = AutoModelForSpeechSeq2Seq.from_pretrained(
model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=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,
chunk_length_s=30,
batch_size=16, # batch size for inference - set based on your device
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"])Evidence level: Publisher documentation linked; no execution reproduced by this site.
Before downloading
- Check the exact checkpoint and license linked here.
- Confirm a compatible runtime and precision or quantization for your platform.
- Allow for context, batch size and any extra media components beyond the weights.
Technical details
What the source and the catalog store for this checkpoint. Missing fields are left out, never zero.
- Checkpoint / version
- openai/whisper-large-v3
- Task and modality
- automatic-speech-recognition · audio
- Count and basis
- 1.55B · manufacturer model card
- Released
- openai/whisper commit history (GitHub API), commit "large-v3 (#1761)" · read 2026-10-09source
- Published repository size
- 24.7 GBNot runtime memory or a minimum download.files
- Published components
- Whisper ASR checkpoint
- Runtime and precision
- No local execution verified.
- Commercial use
- Allowed Apache License 2.0 text (SPDX) · read 2026-10-18license text
Every stored field
- Revision checked
- 06f233fe06e710322aca913c1bc4249a0d71fce1
- Architecture
- WhisperForConditionalGeneration
- Library
- transformers
- Hub task tag
- automatic-speech-recognition
- Formats
- safetensors, pytorch
- Repository created
- 2023-11-07creation is not the model's release
- Repository updated
- 2024-08-12a commit is not a release date
- Hub downloads (30 days) · likes
- 3,499,023 · 6,692this repository only · read 2026-10-25
- Read from the Hub page
- task, transformers class, task tags, file formatsHub overview and file listing · read 2026-10-18
- Catalog state
- Hub config or weights found; no inventory benchmark link
The official model card identifies automatic speech recognition and 99 languages. No local speed or memory requirement is published here.
Sources
Identity and published specifications come from the checkpoint page. Benchmark scores belong to the named release and do not measure speed on your computer.
- Publisher checkpoint and model card
- Weight license and use terms
- Quality benchmark source · Word error rate · Open ASR
- Publisher hardware documentation
Catalog reviewed: 2026-09-30. Benchmark reviewed: 2026-10-20.















































