Skip to main content

​
Function Signature

def load_model(
    name: str,
    device: Optional[Union[str, torch.device]] = None,
    download_root: str = None,
    in_memory: bool = False,
) -> Whisper

​
Parameters

​
name
str
required
One of the official model names listed by whisper.available_models(), or path to a model checkpoint containing the model dimensions and the model state_dict.Available official models:
  • tiny.en, tiny
  • base.en, base
  • small.en, small
  • medium.en, medium
  • large-v1, large-v2, large-v3, large
  • large-v3-turbo, turbo
Models with .en suffix are English-only variants.
​
device
Union[str, torch.device]
default:"None"
The PyTorch device to put the model into. If not specified, automatically selects "cuda" if available, otherwise "cpu".Common values: "cuda", "cpu", "cuda:0", torch.device("cuda")
​
download_root
str
default:"None"
Path to download the model files. By default, uses "~/.cache/whisper" (or $XDG_CACHE_HOME/whisper if the environment variable is set).The function creates this directory if it doesn’t exist.
​
in_memory
bool
default:"False"
Whether to preload the model weights into host memory. When True, the model checkpoint is kept in memory as bytes rather than being read from disk.This can be useful for deployment scenarios where filesystem access is limited.

​
Returns

​
model
Whisper
The Whisper ASR model instance, ready for inference. The model is already loaded onto the specified device and has its state_dict loaded.The returned model has alignment heads set if it’s an official model (used for word-level timestamps).

​
Example

import whisper

# Load the base model (auto-detects CUDA)
model = whisper.load_model("base")

# Load the turbo model on CPU
model = whisper.load_model("turbo", device="cpu")

# Load from a custom checkpoint file
model = whisper.load_model("/path/to/custom-model.pt")

# Load with custom cache directory
model = whisper.load_model("small", download_root="/custom/cache/path")

# Load in-memory (useful for serverless deployments)
model = whisper.load_model("base", in_memory=True)

​
Notes

​
Model Download and Caching

  • On first use, models are downloaded from Azure CDN to the cache directory
  • Downloaded models are verified using SHA256 checksums
  • Subsequent calls reuse the cached model files
  • If checksum verification fails, the model is re-downloaded automatically

​
Model Selection

Choose a model based on your requirements:
  • Speed: tiny (fastest) → turbo → base → small → medium → large (slowest)
  • Accuracy: tiny (lowest) → base → small → medium → large → turbo (highest)
  • English-only: Use .en variants for better English performance
  • Multilingual: Use non-.en models for 99+ languages

​
Device Compatibility

  • FP16 (half precision) is only supported on CUDA devices
  • CPU inference automatically uses FP32 even if FP16 is requested
  • The model returns a warning if performing inference on CPU when CUDA is available

​
Error Handling

try:
    model = whisper.load_model("nonexistent-model")
except RuntimeError as e:
    # Raises: Model nonexistent-model not found; available models = [...]
    print(e)

​
Alignment Heads

Official models automatically have alignment heads configured, which are used for:
  • Word-level timestamp extraction
  • Cross-attention pattern analysis
  • Dynamic time warping for precise timing
Custom checkpoints loaded from file paths will not have alignment heads set unless they were included in the checkpoint.

Build docs developers (and LLMs) love

Get started for freeTalk to us