Ask HN: Is there demand for a local-first live transcription tool? | Hacker News Reader

Ask HN: Is there demand for a local-first live transcription tool?

14 pointsuohzxela15 comments

As a hearing impaired software engineer, I have built my own local-first solution to transcribe entirely locally in real time word by word. It's my daily driver for transcribing meetings, interviews, etc. Because of its local-first capability, I do not have to worry about privacy concerns when transcribing meetings at work as all data stays on my machine. It's about as fast as Otter.ai although there's definitely room for improvements in terms of UX and speed. Caveat is that it only works on MacBooks with Apple silicon, fronted by a very simple TUI.

I am thinking of putting it up as an open source project on GitHub to garner interest. If interest is high, commercializing it as a product would be a next step. Before I go too deep in that direction, I'm curious about the market demand. How does the HN community think about the demand for a local-first live transcription tool?

View on HN

> If interest is high, commercializing it as a product would be a next step

Why? OpenAI's Whisper is a free local model with native GUIs on most operating systems. I wrote one myself in a weekend, it's one of hundreds of similar GUIs that exist.

What functionality would you be offering that people can't otherwise get for free? This is a pretty competitive space, trying to sell a TUI is going to leave you gravely disappointed by the sales (unless it's extremely impressive).

Most of the free Whisper GUI offerings do not provide real-time transcription; even if they do, it's not performant enough for my needs.

TUI is just a starting point. Differentiators would be a better GUI, and better real-time performance.

The concern I have for commercialising this product is that there's a very high chance Apple will roll this out as a feature in macOS 17 or 18. It seems like a natural evolution of Apple Intelligence.

Edit: Just noticed that Apple Notes in Sequoia already has "Live audio transcription"

It’s a good point. I briefly looked at their GUI and it’s very tied to the Apple Notes surface. It’s a good way to provide real time transcription to the masses, but their UX and speed leaves a lot to be desired.

Why do we have so many custom notes app when we can just use Apple Notes?

I built one and use it daily myself, it uses Whisper for local transcription and Llama 3.2 for local AI conversation, it has good UX as well, support live record or importing any audio/video files.

https://apps.apple.com/us/app/notes-ai/id6477414161

There's the MacWhisper app which seems to be used a lot, so perhaps you could integrate, or do something similar? I guess if you open-source it the MacWhisper dev can also integrate it

What model are you using? Whisper Turbo can do real time transcription and there are implementations that are pretty fast.

I’m using Whisper medium.en. I’ll check out Turbo, thanks! Looks like it’s released last month. My solution is pretty model agnostic. If Turbo is indeed as fast as they advertised with lower WER than small.en, then it will supercharge my solution as well (it’s already pretty fast and accurate).

Is it stronger/better than Real-time Whisper WebGPU ?

https://huggingface.co/spaces/Xenova/realtime-whisper-webgpu

I've tried this demo before. My solution is better in terms of latency and accuracy. You can place Otter.ai and Real-time Whisper WebGPU side by side and see the difference. That's the difference I observed with my solution as well.

This reminded me of https://github.com/usefulsensors/moonshine

Yeah I checked that out. My current Whisper model is small.en which has a lower WER and is fast enough for real time transcription.

I would love a local transcription app that works on Android and Linux.

Yes, my partner is hearing-impaired and I’ve been looking for a tool like this for them

Good to know. Do the existing real time transcription apps work? If not why?

Reply on news.ycombinator.com