I might take another look into doing it in a different way that gradually builds up from successful calls, I just need to think of how to do this in a simple(ish) way for non-technical users and a way that still works well enough on low to mid tier laptops.

Ahh yea as for the models:

Speech to text - Nvidia Parakeet TDT 0.6b V3

Diarisation - Nvidia Marblenet for the speech detection, TitaNet-Large for the embeddings and then using NeMo multi-scale to do clustering around them

Two more improvements, audio quality improvements which is currently in the works and close to release and a new document generation model. I'm currently using a custom fine tuned Phi-4 (released December 2024!) model, that's _so old_ in the grand scheme of LLMs, I just haven't had time to benchmark and properly test some new models whilst this currently does a good job as it is. There has to be some gains here, but who knows!

It's a hard problem I've been working away on for a while now. It's far from perfect but every step brings it a bit closer.

Inbetween that I've been working on a "software house" (don't know how else to describe it). Pretty much codex/claude running on a VPS with a web interface and the ability to actually build and publish software / webapps etc.

It came from me sitting on the train on the way home and I had an idea pop into my head for a pebble watch app, I wanted to build it and test it there and then as I still had about 45 mins left. I know you can run codex against git repos in the cloud but I didn't want to faff with that (and last I remember the allowances for this were much less generous). It's been really fun, probably many things like this out there but it works exactly how I want it to. I plan to open source it once I've tidied things up a bit more.

You probably don't need the AI _they_ shoehorn in.

Thank you for the inspiration for a weekend project!

Have spent the last month giving the UI a bit of a modernisation refresh and simplifying/improving some elements based on early user feedback. There's also been a boat load of performance improvements in the dirarisation and document generation pipeline.

Feel free to download the prerelease version (its unsigned) here - https://downloads.blazingbanana.com/whistle-enterprise/unsta...

This looks very useful, will download and give it a shot later. It took me a few seconds to find it on your page, and only got to the screenshots in the "navigate" sections after clicking through a lot. I would suggest putting a screenshot or something on the landing page so people can see and understand what it is.

Thank you, it's nice to hear someone else has gone through similar pain (in a good way)!

It's been slow and steady, but it's hard. I've commented previously that whilst the cost to build software has plummeted compared to 2 or 3 years ago, the ability to sell it has got harder and I feel this will keep accelerating.

Linux - https://downloads.blazingbanana.com/whistle-subtitles/unstab...

Windows - https://downloads.blazingbanana.com/whistle-subtitles/unstab...

Mac - https://downloads.blazingbanana.com/whistle-subtitles/unstab...

This was built just for them so I've not spent too much time on the UI (ignore `unstable` in the name, it's just not on a proper release branch) but it's completely free so give it a go if you want. I'm working on the diarisation step so it can tag subtitles to people but that's not ready yet.

It utilises nvidia Parakeet as the ASR model, it is very much European language focused, the supported ones are:

Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Ukrainia

If these languages aren't what you're looking for let me know what you need and I'll see what I can do.

I use subtitles extensively for everything I watch, so if I can help someone make older movies more accessible with them then that would make me happy.

Maybe allow a new user to view previous data for their specific investment.

Also great video, it hooked me.

One of the hardest parts I've found is the diarisation (who said what) side of things. Trying to tune this and have it working in a way that doesn't absolutely grind the laptop to a halt or take forever to complete has been _hard_ but also extremely rewarding.

Another part has been the fine tuning side of the Phi-4 model, I'm on version 10 now, getting that pipeline down was a journey in itself, but I've got some great results. I wrote a bit about it in a comment here - https://news.ycombinator.com/item?id=48385906#48389625

I absolutely love working on this, I still wake up and the first thing I think about is voice transcription pipelines (sad I know), but I'm excited to see how much further performance and utility I can squeeze out.