Whisper large-v3 vs large-v3-turbo for transcription
Local speech-to-text: the full model or the faster turbo
Facts from the reviewed catalog
| Specification | Whisper large-v3 | Whisper large-v3-turbo |
|---|---|---|
| Task | Audio & speech | Audio & speech |
| Parameters | 1.55B | 809M |
| License | Apache 2.0 | MIT |
| Commercial use | Allowedlicense text | Allowedlicense text |
| Quality · Open ASR Leaderboard (English short-form, 2026-05-22) | 7.44%average WER · lower is better · source | 7.83%average WER · lower is better · source |
| Memory | publisher's documented path from 10 GB (CUDA) | publisher's documented path from 6 GB (CUDA) |
| Checkpoint | official card | official card |
In the Open ASR Leaderboard (English short-form, 2026-05-22), Whisper large-v3 has an average WER of 7.44% and Whisper large-v3-turbo 7.83%: Whisper large-v3 gets 0.39 points fewer words wrong on the same test sets. The published checkpoints on the leaderboard's own GPU, same results file for both (commit 16743c6, collected 2026-10-20); not a run on your hardware. Scale: average word error rate (WER) over eight English test sets, %, lower is better.
Memory on three machines
Each machine's own reserve is added (graphics cards 2 GB; unified and system memory 25%, 2 to 8 GB). Planning estimates, not measured runs.
| Machine | Whisper large-v3 | Whisper large-v3-turbo |
|---|---|---|
| RTX 3050 · 6 GB VRAM | 10 GBBelow the documented minimum | 6 GBWithin a documented path |
| RTX 4060 · 8 GB VRAM | 10 GBBelow the documented minimum | 6 GBWithin a documented path |
| MacBook Air 13-inch M4 · 16 GB unified | No documented path for this platform (a CUDA figure is not carried over to a Mac) | No documented path for this platform (a CUDA figure is not carried over to a Mac) |
Whisper large-v3: publisher's documented path from 10 GB (CUDA). Whisper large-v3-turbo: publisher's documented path from 6 GB (CUDA). Formulas
Everything that runs on each machine: RTX 3050 · 6 GB VRAM · RTX 4060 · 8 GB VRAM · MacBook Air 13-inch M4 · 16 GB unified
Which to choose
OpenAI's card describes turbo as large-v3 with fewer decoding layers: 0.809 billion parameters against 1.55, and OpenAI documents about 6 GB of GPU memory for turbo against 10 GB for large-v3. That is why turbo fits a 6 or 8 GB graphics card. OpenAI's figures are for CUDA GPUs; no path is documented for a Mac, so that row is a dash. OpenAI says turbo is much faster with a small loss of accuracy, and describes it as optimized for transcription. On the Open ASR Leaderboard's English short-form results (file of 2026-05-22, collected 2026-10-20), large-v3 averages 7.44% word error rate and turbo 7.83%, so turbo gives up about 0.4 points of accuracy. Beyond that, the choice rests on memory, speed and licence: large-v3 is Apache 2.0 on its Hub card and turbo MIT, and both allow commercial use.
Files and run commands
The default file of the most downloaded collected source, or the commands each card documents. Not run by this site.
Whisper large-v3
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"])Whisper large-v3-turbo
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"])Numbers each publisher reports
Each card's own results with its own settings. Publishers choose different tests and settings, so these values are never compared across publishers here.
Whisper large-v3 · reported by OpenAI
Whisper large-v3-turbo · reported by OpenAI
Limits and dates
The published quality benchmark does not test a local quantized build. Memory estimates are not speed measurements or guaranteed hardware fits. Repository size, when available, is not VRAM. Formulas
Page updated · Catalog reviewed · Open ASR Leaderboard (English short-form, 2026-05-22), collected