speech_to_txt_to_LLM_to_cmd_sh

This is very much a prototype of a concept .. but something I know could be improved to be something pretty powerful.

Future versions will be published here: http://github.com/tombrown86/

Speech to text to local LLM to shell cmd.

Records audio in console allowing user to vocally describe a desired shell command / operation.

This is converted to an LLM prompt which is run on a local LLM (like gpt-2 or llama), and converted to (ideally) a desired linux command.

Warning: This takes a while to run on your average machine and some of the suggested commands are a bit awful but I'm convinced with the right model this could be awesome :)
At the moment it may be better to use a super decent online model via it's API :)

E.g.
"Computer, show me the files in the current directory"
"find me a list of files called blah that contain blahblah and sort them by blahblahblah..."

Setup:

# If using youre gcloud to get audio transcribed
gcloud auth application-default login

brew install ffmpeg
brew install sox
chmod +x s2cmd.sh

# update GCLOUD_PROJ_NAME, MODEL and BIN vars in sh before running

./s2cmd.sh

Speech to text step

This currently relies on google clouds transcription service for which you need a project + API etc.

Also uses sox and ffmpeg to create / clean a wav recording.

Alternatively, skip speech to text step and just have the sh accept the prompt text :)

prompt=""
for arg in "$@"
do
  prompt="$prompt $arg"
done