How do I transcribe audio locally with Whisper?
Which Whisper to pick, the memory OpenAI documents and the commands its card prints.
Short answer
Run OpenAI's Whisper on your own computer: it turns an audio file into text, and the recording never leaves your machine.
- Pick: large-v3-turbo. About 6 GB of GPU memory; much faster, slightly less accurate.
- large-v3: About 10 GB; slightly more accurate, slower.
- Licence: Both allow commercial use.
Documented memory from OpenAI, as linked on each model page; licences from the Hub cards. Methodology
Which Whisper
Whisper is OpenAI's open speech-recognition model: it listens to an audio file and writes what was said, in the language spoken. It runs entirely on your computer, so recordings never leave it. The two current large checkpoints differ mainly in the decoder: turbo keeps large-v3's encoder and uses far fewer decoding layers, which makes it smaller and, by OpenAI's description, much faster, with a small loss of accuracy.
| Model | Parameters (B) | Documented GPU memory (GB) | RTX 3050 · 6 GB VRAM | Licence · commercial use |
|---|---|---|---|---|
| Whisper large-v3 | 1.55 | 10 | Needs more memory | Apache 2.0 · Allowed |
| Whisper large-v3-turbo | 0.809 | 6 | Fits | MIT · Allowed |
Run it
The ways Whisper large-v3-turbo's card documents, exactly as printed. Not run by this site.
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-turbo"
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-turbo"
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-turbo"
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"])Limits
OpenAI's memory figures are for NVIDIA (CUDA) graphics cards. No figure is documented for a Mac, so the site does not carry one over; community ports exist, but their memory and speed are their own.
Accuracy: on the Open ASR Leaderboard's English short-form results, large-v3 has an average word error rate of 7.44% and turbo 7.83% (lower is better), so turbo gives up about 0.4 points for its smaller decoder. These are the published checkpoints on the leaderboard's own GPU over eight English test sets, not runs on your hardware, and other languages are not measured here. (Open ASR Leaderboard, results file of 2026-05-22, read Oct 20, 2026)
Side by side: Whisper large-v3 vs large-v3-turbo for transcription