converting audio files to markdown must be a pretty recent feature
Quite curious… does it actually do that and if so how? Because STT to get a plaintext file or subtitle (so with timing) has been available via e.g. Whisper quite efficiently for a while now. If this though does do more, e.g. structure (differentiating a title, list, etc) I’d like to learn how.
Might open up a GDPR related issue there. I don’t think people using such a library assume they need connectivity nor that their data would be send to a 3rd party.
Quite curious… does it actually do that and if so how? Because STT to get a plaintext file or subtitle (so with timing) has been available via e.g. Whisper quite efficiently for a while now. If this though does do more, e.g. structure (differentiating a title, list, etc) I’d like to learn how.
There is nothing special going on. This whole project is just a bunch of python libraries coupled together to a cli tool. It uses the package SpeechRecognition to connect to the google speech recognition api: https://github.com/microsoft/markitdown/blob/main/src/markitdown/_markitdown.py#L691
Pretty uninteresting and a bit disappointing. Pandoc is a lot more interesting.
Thanks for the clarification. I checked the code you linked and noticed
recognize_googleand seems it’s relying on https://github.com/Uberi/speech_recognition which then seems to rely on https://github.com/Uberi/speech_recognition/blob/master/speech_recognition/recognizers/google.py so basically are they using an API, sending all the audio data to Google servers?Yes, this is how I read it as well. The library would support to use a local model, but they decided to just send the audio data to Google.
Might open up a GDPR related issue there. I don’t think people using such a library assume they need connectivity nor that their data would be send to a 3rd party.