Introducing Longhand - a call recording transcription app for Graphene OS

christianwu

Everything that can be Kotlin is. The C++ is only for the Qwen 3.5 4B summary model, which runs through llama.cpp. It's just a small wrapper that lets Kotlin load the model:)

Johnnyloans

Wow, that'd be amazing if it allocated words to the correct speaker 100% of the time like your screenshots show.. I'm used to transcripting board meetings with many speakers so maybe that is the issue.

I will often get:

Bob: hello, what is
Tracy: your name? Tracy. Hi Tracy.

Does it remember each speaker across transcriptions and labels them automatically? Perhaps with a "best guess" or something.

christianwu

It's been pretty accurate in my experience, maybe not 100%, but definitely over 90%. Feel free to try it out, any useful feedback is appreciated. I'm going to continue improving it over the next days and weeks.

To answer your question, yes it does recognize each speaker by voice fingerprinting, and labels them automatically.

graphleo

Thank you, that is what I need I'll check it out.

Johnnyloans

Very cool! I like it

2 thoughts

the ability to edit a transcription
the ability to add common corrections. E.g. change UV -> youvee in future transcriptions

Novaliss

Just tested on a swedish recording. It came out like a dadaistic poem..☺️

I'd guess it works better in english though.

Johnnyloans

Novaliss It came out like a dadaistic poem

Hahaha

😭

GazelleInitial

Cool idea thank you!

edent

This is brilliant! Worked well on my 8 Pro.

christianwu

Thank you everyone for the feedback. Working on adding more features at the moment.

Novaliss sorry about that, I haven't tested it in all languages yet. Finding an accurate model small enough to run on the phone for some of the European languages is challenging, but I'm working on it.

t1me

Brilliant project
Please if possible add Hindi language too

graphleo

Yes, worked pretty well.

graphleo

For the most useful feature would be global search on scripts including filters by contact.

christianwu

t1me Thank you! Working on it now!

christianwu

graphleo Current version already has global search of all transcripts, including contact names. I'm working on a feature that shows the matching line in the call where the search term appears, and tapping it jumps directly to that section of the transcript.

GazelleInitial

christianwu
Is there anyway that the compressed files could be transcribed better? I make a lot of calls and the WAV files are pretty big.

christianwu

It's currently quite functional on the compressed files, but not as good as the WAV files obviously. The solution to this would be to add an option to download a bigger voice recognition model, which would consume more storage and battery, but probably yield a ~ 10-20% increase in accuracy. I may add this option if people are wanting it.

Novaliss

christianwu totally get it. I reccon it would perform a lot different if the model could concentrate on one specific language at a time. The output indicated a model that impressively nailed private, not found in a swedish dictionary, just special composite words spoken between friends. Add the package with 25 different languages for the model from which to interpret the next sentence as. The computational demand goes, like very high.

I really like the concept though. And apart from the output when feeding it swedish speech, the app did perform in a rational fashion. I'll watch this space🫡

GazelleInitial

christianwu
This might be out the scope of your project. But could the app re-encode it to a compressed file after it's done it's analysis/transcription.

christianwu

GazelleInitial Will look into this. Appreciate the suggestion!