directly using yt_dlp to chunk audio clips could cause unalignment issues

[haidog-yaqub]

Just found this download method could cause unalignment problem. For example, the start time in the meta is 5 second, but the actual downloaded audio starts from 10 second. This will also cause the length of some downloaded audio to be far less than 10 seconds. I suggest directly downloading the entire audio by yt_dlp and chunking desired clips locally.

[MorenoLaQuatra]

Thank you for opening the issue.

Are you referring to the following line? https://github.com/MorenoLaQuatra/audiocaps-download/blob/dbe83b56ef97fc82143e78d99614ec30cb0135e7/audiocaps_download/Downloader.py#L295

How do you discovered the bug?

[haidog-yaqub]

Thank you for opening the issue.

Are you referring to the following line?

https://github.com/MorenoLaQuatra/audiocaps-download/blob/dbe83b56ef97fc82143e78d99614ec30cb0135e7/audiocaps_download/Downloader.py#L295

How do you discovered the bug?

Yes, I checked the length of data downloaded by your repo and found some of them are very short, so I manually checked them and found the problem. I also tried yt-dlp in command line, but the issue still exists. It should be an yt-dlp bug.

[haidog-yaqub]

The way I solved it is to download entire audio and then chunk it by other tools.

[MorenoLaQuatra]

Thank you so much to raise the issue then. Can you suggest any other tool to use? Does downloading everything from yt-dlp and then manually cut (let's say, with torchaudio) solve the issue in your opinion?

[haidog-yaqub]

Thank you so much to raise the issue then. Can you suggest any other tool to use? Does downloading everything from yt-dlp and then manually cut (let's say, with torchaudio) solve the issue in your opinion?

Yes, downloading entire audio and cutting it by torchaudio works.

[MorenoLaQuatra]

Thank you again. Just to verify if I'm correctly identifying the problem, do you think this code will solve the issue?

        # Download the file using yt-dlp
        # os.system(f'yt-dlp -x --audio-format {self.format} --audio-quality {self.quality} --output "{target_file_path}" --postprocessor-args "-ss {start_seconds} -to {end_seconds}" https://www.youtube.com/watch?v={ytid}')
        # Download the ENTIRE audio file
        os.system(f'yt-dlp -x --audio-format {self.format} --audio-quality {self.quality} --output "{target_file_path}" https://www.youtube.com/watch?v={ytid}')
        # now manually cut the audio file
        try:
            waveform, sample_rate = torchaudio.load(target_file_path)
            waveform = waveform[:, int(start_seconds * sample_rate):int(end_seconds * sample_rate)]
            torchaudio.save(target_file_path, waveform, sample_rate)
        except Exception as e:
            print('Error loading audio file: ', target_file_path)
            print(e)
            # delete file if it exists
            if os.path.isfile(target_file_path):
                # delete file
                os.remove(target_file_path)

I did an automated check of a batch of 1000 files and the duration seems to be correct.

[haidog-yaqub]

Thank you again. Just to verify if I'm correctly identifying the problem, do you think this code will solve the issue?

        # Download the file using yt-dlp
        # os.system(f'yt-dlp -x --audio-format {self.format} --audio-quality {self.quality} --output "{target_file_path}" --postprocessor-args "-ss {start_seconds} -to {end_seconds}" https://www.youtube.com/watch?v={ytid}')
        # Download the ENTIRE audio file
        os.system(f'yt-dlp -x --audio-format {self.format} --audio-quality {self.quality} --output "{target_file_path}" https://www.youtube.com/watch?v={ytid}')
        # now manually cut the audio file
        try:
            waveform, sample_rate = torchaudio.load(target_file_path)
            waveform = waveform[:, int(start_seconds * sample_rate):int(end_seconds * sample_rate)]
            torchaudio.save(target_file_path, waveform, sample_rate)
        except Exception as e:
            print('Error loading audio file: ', target_file_path)
            print(e)
            # delete file if it exists
            if os.path.isfile(target_file_path):
                # delete file
                os.remove(target_file_path)

I did an automated check of a batch of 1000 files and the duration seems to be correct.

Yes, I think it should work. You can compare with previous data, especially those short ones.

[MorenoLaQuatra]

I will close the issue, open it again if something else is missing.