Blisper

Transcribe audio files to text very rapidly

install with homebrew

brew install llimllib/whisper/blisper

I would love help packaging it for other package managers, but this is the only one I use so that's the only one I've made a package for at the moment.

building

Blisper currently only supports building on systems with homebrew.

To build, run brew install llimllib/whisper/libwhisper && make

usage

Usage: blisper [OPTIONS] <input-audio> <output-transcript>

Use whisper.cpp to transcribe the <input-audio> file into <output-transcript>

OPTIONS

  -config:       print the config for this app
  -format <fmt>: the output format to use. Defaults to "txt"
  -help, -h:     print this help
  -model, -m:    the name of the whisper model to use. Defaults to "small"
  -stream:       if passed, stream output to stdout
  -verbose, -v:  print verbose output

MODELS

  Valid models are: tiny.en, tiny, base.en, base, small.en, small, medium.en, medium, large-v1, large

  Blisper will automatically download a model if you do not already have it on your system

FORMATS

  Valid subtitle formats are srt, ssa, stl, ttml, txt, and vtt. The default format is txt

status

alpha. Basically functional but not yet easy to download and build

There are general discussions about the API in this thread

The main reason I want my own CLI for whsiper is that the binary built by make in the whisper.cpp repository expects you to manage your own models; I think that's cumbersome and user-unfriendly. Ideally I'd like to have a binary that can be brew installed, and this repository is a step towards it.

However, until the go binary can do parallel processing and access more of the functionality in whisper.cpp, I don't think this will reach a high enough level of quality to make it workable.

thanks

many thanks to @ggerganov for whisper.cpp

TODO

stream data from the WAV into the processing function, rather than doing it all in batch
- would save memory and increase speed
  - unless crossing the C boundary would be too costly
more configuration options
- smaller chunks
- here are the options the default binary supports

llimllib / blisper