Skip to content

Latest commit

 

History

History
108 lines (67 loc) · 4 KB

README.md

File metadata and controls

108 lines (67 loc) · 4 KB
Raw

STREAM WHISPER

Well, I figured out distil-whisper would work better with stream capabilities, i.e ability to transcribe videos/audios directly from the internet, not just file paths. So, stream-whisper was born.

In case you still haven't figured, stream whisper is a fork of huggingface's distil whisper which is fundamentally a fast speech recognition and transcribing engine...(the devil is in the details, lol)

INSTALLATION

  1. Clone or download this repository

  2. install the requirements with pip:

     pip install requirements.txt

NB: If this fails, try pip3 install requirements.txt instead. 3. from the root directory, run:

    cd stream_whisper

USAGE

the only mandatory argument is argument (file url or link url).

python stream_whisper.py <url>

explicitly specify the output file:

	python run.py <url> --out '/output/path/of/audio'

specify the output the of the (transient) audio N.B: This is only required for a streamed (downloaded) audio. if the audio is a local audio, the out argument is totally unnecessary and might lead to errors

	python run.py <url>  --out '/output/path/of/audio'

specify the output path of the transcription with --text. specifically, --text 'cli' and not specifying --text at all, or --text 'both' will output to both cli and a default file 'output.txt' in working directory. meanwhile, any other value, e.g --text 'texter' or --text '/path/to/output/file' will output to the file path argument.

	python run.py <url> --out '/output/path/of/audio' --text '/transcription/output'

specify --long argument to specify audio length, for chunking while processing. otherwise, automatic length detection is used.

	python run.py <url> --out '/output/path/of/audio' --text '/transcription/output' --long

specify --short argument for audio lesser than 30 seconds length. otherwise, automatic length detection is used.

	python run.py <url> --out '/output/path/of/audio' --text '/transcription/output' --short

Additional Speed & Memory Improvements

Specify --spec to use Speculative Decoding

	python run.py <url> --out '/output/path/of/audio' --text '/transcription/output' --spec

specify --flash to use Flash Attention

	python run.py <url> --out '/output/path/of/audio' --text '/transcription/output' --flash

specify --bt to use better transformer

	python run.py <url> --out '/output/path/of/audio' --text '/transcription/output' --bt

specify --settings to use a custom settings file.

	python run.py <url> --out '/output/path/of/audio' --text '/transcription/output' --settings '/path/to/settings/file'

N.B: The original README.md of distil whisper is here.

TO-DO

  • add to pip
  • further tests
  • further features

Acknowledgements

Citation

If you use this model, please consider citing the Distil-Whisper paper:

@misc{gandhi2023distilwhisper,
      title={Distil-Whisper: Robust Knowledge Distillation via Large-Scale Pseudo Labelling}, 
      author={Sanchit Gandhi and Patrick von Platen and Alexander M. Rush},
      year={2023},
      eprint={2311.00430},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

And also the Whisper paper:

@misc{radford2022robust,
      title={Robust Speech Recognition via Large-Scale Weak Supervision}, 
      author={Alec Radford and Jong Wook Kim and Tao Xu and Greg Brockman and Christine McLeavey and Ilya Sutskever},
      year={2022},
      eprint={2212.04356},
      archivePrefix={arXiv},
      primaryClass={eess.AS}
}

EDIT: Stream-whisper will be taking a drastic turn in structure, with a rust cli tool, embedded flite, and other tools to ensure it becomes way powerful tool for the average and enlightened computer user