model.transcribe() passes unknown progress argument to DecodingOptions

[rabbitxyz2026]

Description

I'm trying to receive transcription progress through a callback passed to model.transcribe(). However, the progress argument is not recognized and is forwarded to DecodingOptions, which raises a TypeError.

Environment

OS: Windows 10
Python: (please fill in your version)
PyTorch: (please fill in your version)
Whisper: (please fill in your installed version)
CUDA: Yes

Upgraded latest whisper.

pip install --upgrade --no-deps --force-reinstall git+https://github.com/openai/whisper.git

Minimal Reproducible Example

import whisper
from pathlib import Path
import torch

def progress_transcribe(segments, segment_size, left_segment, rate):
    print(segments, segment_size, left_segment, rate)

device = "cuda" if torch.cuda.is_available() else "cpu"

model_dir = Path(r"whisper_models")
model_dir.mkdir(parents=True, exist_ok=True)

model = whisper.load_model(
    "large",
    device=device,
    download_root=model_dir,
)

result = model.transcribe(
    r"audio.mp3",
    progress=progress_transcribe,
)

print(result["text"])

Output

Traceback (most recent call last):
  File "D:\Desktop\Speech\test.py", line 16, in <module>
    result = model.transcribe(..., progress=progress_transcribe)
  File "...site-packages\whisper\transcribe.py", line 295, in transcribe
    result: DecodingResult = decode_with_fallback(mel_segment)
  File "...site-packages\whisper\transcribe.py", line 200, in decode_with_fallback
    options = DecodingOptions(**kwargs, temperature=t)
TypeError: DecodingOptions.__init__() got an unexpected keyword argument 'progress'

Expected Behavior

I expected the progress callback to be invoked during transcription, or at least for model.transcribe() to recognize the progress argument.

Actual Behavior

Instead, the progress argument appears to be forwarded to DecodingOptions, which does not accept it, resulting in the following exception:

TypeError: DecodingOptions.init() got an unexpected keyword argument 'progress'

Question

Is the progress callback still supported?

If it is, am I using it incorrectly?

If it is no longer supported, would it be possible to document this or provide an alternative way to receive transcription progress?

[Samielakkad]

progress is not part of the current upstream openai/whisper transcribe() API, so your call is behaving exactly as the implementation is written.

The function accepts known transcription parameters plus **decode_options. Any unknown keyword therefore becomes a decoding option. Later, Whisper constructs DecodingOptions(**kwargs, temperature=t), and DecodingOptions has no progress field, which produces your traceback.

You can confirm both paths in the current whisper/transcribe.py:

• the function signature ends with **decode_options
• the built-in progress implementation is an internal tqdm bar, enabled only when verbose is False

If you only need visible progress, use:

result = model.transcribe(
    r"audio.mp3",
    verbose=False,
)

The current behavior is:

• verbose=True: print decoded segments
• verbose=False: show the tqdm progress bar
• verbose=None: show neither

There is no supported programmatic progress-callback parameter in upstream Whisper at the moment. For an application callback, the reliable options are to maintain a small wrapper/fork of transcribe.py that calls your function immediately after the existing pbar.update(...), or open a feature request for an explicit progress_callback parameter. Monkey-patching tqdm can work, but it depends on an internal implementation detail and is brittle across releases.

So remove progress=progress_transcribe; you are not using an old callback incorrectly. The callback API simply is not present in the current upstream function.

[Samielakkad]

Update: I closed PR #2811 after finding several earlier progress-callback implementations, including the newer tested PR #2797. The callback API is still not part of main; tracking the existing proposal is clearer than maintaining another competing implementation.

[sylvesterkaczmarek]

Upstream Whisper doesn't support a progress= callback in transcribe(). Unknown kwargs get forwarded to DecodingOptions, which causes this error. Use verbose=False for the built-in progress bar; a callback currently needs a small wrapper/fork.