How to get the progress bar while transcribing?

[benjieperez]

Here is my code.

import whisper

Load Whisper model Large

MODEL = whisper.load_model("model/medium.pt", in_memory=True)

Set options

translate_options = dict(task="translate", **dict(language=translate_language, beam_size=5, best_of=5))
result_transcribe = MODEL.transcribe(audio, **transcribe_options, fp16=False, verbose=False)

When i enable verbose=False, this is the terminal output while transcribing,

Detected language: Turkish
60%|███████████████████████████████████████████████████████████████████████████ | 4058/6058 [00:21<00:00, 277.58frames/s]
100%|███████████████████████████████████████████████████████████████████████████ | 6058/6058 [00:21<00:00, 277.58frames/s]

I want to get the progress bar while its transcribing is it possible?

Thanks,

[mayeaux]

This is how I do it in JavaScript: https://github.com/mayeaux/generate-subtitles/blob/master/helpers/formatStdErr.js#L9

[benjieperez]

How to set PYTHONUNBUFFERED? No idea on this.

[glangford]

@benjieperez

• Example Dockerfile should document to run python unbuffered docker-library/python#604 (comment)
• https://docs.python.org/3/library/sys.html#sys.stdout
• https://stackoverflow.com/questions/2231227/python-subprocess-popen-with-a-modified-environment

[glangford]

@benjieperez ...something along these lines is worth a try

import subprocess, os
whisper_env = os.environ.copy()
whisper_env["PYTHONUNBUFFERED"] = "1"
subprocess.Popen( ... , env=whisper_env)

[benjieperez]

@benjieperez ...something along these lines is worth a try

import subprocess, os
whisper_env = os.environ.copy()
whisper_env["PYTHONUNBUFFERED"] = "1"
subprocess.Popen( ... , env=whisper_env)

Still no luck, the progress bar shows after the process not while processing the audio.

[benjieperez]

---

Date Run: 2023-01-23 10:44
Running Command: whisper files_uploaded/63c8cb48abc63045d80fd5d2.wav --model=medium --model_dir=model/ --verbose=False --fp16=False --task=transcribe --output_dir=output/63c8cb48abc63045d80fd5d2
STDOUT: Detected language: Turkish

STDOUT:
STERR:

None
(This is blank, after the process finished, this will be the result)
STERR: 0%| | 0/6060 [00:00<?, ?frames/s]

{'percentDone': '0%', 'timeElapsed': '00:00', 'speed': 0, 'percentDoneAsNumber': 0.0, 'timeRemaining': {'string': '00:00', 'hoursRemaining': 0, 'minutesRemaining': 0, 'secondsRemaining': 0}}
STERR: 36%|███▋ | 2200/6060 [00:10<00:17, 215.65frames/s]

{'percentDone': '36%', 'timeElapsed': '00:10', 'speed': '215.65', 'percentDoneAsNumber': 36.0, 'timeRemaining': {'string': '00:17', 'hoursRemaining': 0, 'minutesRemaining': 0, 'secondsRemaining': 17}}
STERR: 74%|███████▍ | 4500/6060 [00:18<00:06, 256.08frames/s]

{'percentDone': '74%', 'timeElapsed': '00:18', 'speed': '256.08', 'percentDoneAsNumber': 74.0, 'timeRemaining': {'string': '00:06', 'hoursRemaining': 0, 'minutesRemaining': 0, 'secondsRemaining': 6}}
STERR: 100%|██████████| 6060/6060 [00:21<00:00, 295.05frames/s]

{'percentDone': '100%', 'timeElapsed': '00:21', 'speed': '295.05', 'percentDoneAsNumber': 100.0, 'timeRemaining': {'string': '00:00', 'hoursRemaining': 0, 'minutesRemaining': 0, 'secondsRemaining': 0}}
STERR: 100%|██████████| 6060/6060 [00:21<00:00, 276.00frames/s]

{'percentDone': '100%', 'timeElapsed': '00:21', 'speed': '276.00', 'percentDoneAsNumber': 100.0, 'timeRemaining': {'string': '00:00', 'hoursRemaining': 0, 'minutesRemaining': 0, 'secondsRemaining': 0}}
STERR:
Total Time Processed: --- 43.48728823661804 seconds ---

[zefr0x]

I think it would be great to have a function like whisper.transcribe(), that yields results rather then waiting to the end then returning every thing at ones.

I think it's not complicated, since whisper.transcribe() can print result in realtime when we set verbose=True.

[aadnk]

Another option is to override the tqdm.tqdm progress bar in Whisper itself:

import os
import sys
import tqdm
import urllib.request 

class _CustomProgressBar(tqdm.tqdm):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self._current = self.n  # Set the initial value
        
    def update(self, n):
        super().update(n)
        self._current += n
        
        # Handle progress here
        print("Progress: " + str(self._current) + "/" + str(self.total))

# Inject into tqdm.tqdm of Whisper, so we can see progress
import whisper.transcribe 
transcribe_module = sys.modules['whisper.transcribe']
transcribe_module.tqdm.tqdm = _CustomProgressBar

import whisper
model = whisper.load_model("medium")

if not os.path.exists("sample1.wav"):
    urllib.request.urlretrieve("https://github.com/itsupera/audiobook_alignment/raw/main/samples/sample1.wav", "sample1.wav")

result = model.transcribe("sample1.wav", language="Japanese", fp16=False, verbose=None)
print(result['text'])

That way, you don't have launch Whisper in another process, or parse the output. I have a complete example of how this can be done here:

• https://huggingface.co/spaces/aadnk/whisper-webui/blob/main/src/hooks/whisperProgressHook.py#L95

I use it to pass progress from Whisper into Gradio. But yeah, hopefully there will be a proper API soon so we don't have to rely on method hooks or parsing the console output.

[bees4ever]

love to see it as a standard callback of the transcribe function

[J-Klinke]

I hope I'll get a reply on this old discussion, but I having trouble with your example. I tried with this code snippet and the code from the linked repo, but the output is still buffered / appears as transcribe() terminates. Can you think of anything I'm doing wrong?

For my test i copied your code from whisperProgressHook.py and progressListener.py into one python script and only adapted the audio file and the language.
Thanks a lot!

[jbeltran73-2]

#!/usr/bin/env python3

import whisper

model = whisper.load_model("large-v3")
result = model.transcribe("scale.m4a")

with open("transcription.txt", "w") as f:
        f.write(result["text"])

I don't see the progress bar with this simple code. I've searched and transcribe.py has tqdm already.
What I could be doing wrong?

Thank you

[dkleptsov]

Many functions have "verbose" boolean parameter. It controls how much information function will print to terminal.

Transcribe also has this parameter:
result = model.transcribe("scale.m4a") - will print nothing
result = model.transcribe("scale.m4a", verbose=False) - will print only progress bar
result = model.transcribe("scale.m4a", verbose=True) - will print detailed information

[jbeltran73-2]

Thank you trying it.