Link: https://opensubs.app/

I work on OpenSubs, a free, open source (AGPL-3.0) subtitle tool that runs entirely in the browser tab.

You drop in a video file, and Whisper transcribes it on your own machine, using transformers.js with WebGPU where available and WebAssembly otherwise. The model (40–250 MB) downloads once and is cached. There is no upload endpoint in the product, so the video has nowhere to go.

After that you can:

• fix lines by typing over them (click a timestamp to jump to that moment)
• translate into 20 languages with Chrome's built-in on-device translator
• pick one of 12 caption styles, including word-by-word highlighting
• export SRT / VTT / ASS, or burn the subtitles into an MP4 (libass compiled to WebAssembly, encoded with WebCodecs)

A few things I learned building it:

• Whisper hallucinates on silence and music ("Thanks for watching!", or the Japanese equivalent). A Silero VAD pass runs before Whisper, and a cleanup step drops the known stock phrases.
• Singing doesn't count as speech for the VAD, so a music video gets a "no speech found" warning. You can still force it.

Honest limits: it only takes video files, not audio-only files. Cue timings can't be edited yet. Builds are release candidates. Everything that runs locally is free with no account; the only paid part is optional cloud translation on our backend (US$5 for 1000 credits), and you can bring your own Claude / OpenAI / DeepL key instead.

Site: https://opensubs.app/
Code: https://github.com/open-subs/opensubs

Feedback welcome, especially on languages where the transcription goes wrong.

Open source, but made with a copyright infringing LLM.

Whisper is not an llm, just because it is trained on big datasets does not make it an LLM. Copyright infringement still happened during the training.

no, but claude is, and it's the main contributor to the repo.

Whisper is not an llm, just because it is trained on big datasets does not make it an LLM.

The training dataset consists of 680,000 hours of labeled audio-transcript pairs sourced from the internet using semi-supervised learning.

Brother, I don't think you know what a large language model is if you think it being trained on a human lifespan's worth of audio transcripts isn't enough.

Its job is NLP. It is an LLM, if not as general-use as something like ChatGPT.

Which software are we talking about? Unfortunately no one links to it, so I will guess. Whisper: website | source is an Ai model, specifically powered by OpenAi.

It sounds like a pretty cool project, thanks for sharing! As a browser-based project designed to run offline, you might want to consider shipping it as an electron app.

Why not just use the whisper plugin for Bazarr?

Whisper hallucinates on silence and music (“Thanks for watching!”, or the Japanese equivalent).

I've had Whisper ask me to donate to anime fansubbing groups. I wonder what it was trained on?

LLM is a subset of NLP. All LLMs are NLP but not all NLP are LLMs. To be more precise, LLM is a method of achieving NLP while NLP is the goal. Whisper is just a specialised speech-to-text engine. It doesn't have the text reasoning engine that is the core of LLMs. Yes, it has a dataset to inform it's encoder and decoder, but that doesn't make it an LLM. If you strip out the encoder, then you're getting closer.

tauri is better and more efficient for this.

I'm a lazy reader so don't know if this is implemented but it would be cool if it could be done live with the current audio out of the system. I know this was done for streamers etc using OBS/localvocal but it's a pain to get working in my experience - having to fiddle around with OBS filter settings.

i get ya, but also a lot of people don't have a dedicated self hosted setup. recently my hdd died and all my films were there, so now i'm downloading and deleting, which just feels like a drag to start bazarr for one film. this is handy for some people.

i love how anti ai people have become the wardens of copyright somehow.

What’s the max file size?

I’ve been looking for the fan edit of marvels Infinity saga.

It’s about 50 hours long divided into many many files ranging from 3-10 gb each. Would it be able to handle this? Or what’s the best way to split them up

Sometimes if I hit the voice typing on my phone, which I'm pretty sure is whisper, and then don't say anything, it just comes out with nonsense.

Just did this and also got...

Thanks for watching!

If you strip out the encoder, then you're getting closer.

That's flat-out not true. BERTs, for example, are LLMs and are encoder-only, just like GPTs are decoder-only. Whisper is trained on a large corpus of text and does NLP tasks. That's all it is to be an LLM.

A transformer model being encoder-only, encoder-decoder, or decoder-only has fuck-all to do with whether it's an LLM. In fact, it being a transformer model at all isn't required for it to be an LLM.

Did whisper improve since last year?

Last time I tried to generate subtitles for a German TV show, the subtitles were more inacurrate than what I would have written down (German is my third language and I struggle to understand anything that isn’t voiceovered) and timing was so bad I didn’t even bother to generate subtitles for the second episode.

I've used Whisper on my private peertube instance recently and the results were pretty decent

Nice, another blindly generated app by Claude Code. 37 commits and already announcing it tells it all.

BERT is a special case. The academic consensus is rightly considering it an LLM since it operates on text but the industry doesn't consider it an LLM since LLM in its colloquial meaning has drifted to mean a model generating text (which BERT doesn't). Whisper is not an LLM since it's working on audio, so it is an Automatic Speech Recognition model (ASR).

Text in = LLM in academic circles (BERT and GPT)

Text in, text out = LLM in all circles (GPT only)

Audio in, text out ≠ LLM (Whisper)

Now, we can argue whether we want to accept the colloquial meaning of LLM, but the fact is that Whisper is not an LLM. And neither ASR nor BERT causes even a fraction of the damage GPT is doing, but that's another discussion.

Statement of fact. Claude was engineered with copyright material.

whats wrong is not copyright theft. whats wrong is labour that was stolen.

Yes and no. As an author of open source projects, they were made available for people to fork, improve and conrribute under permissive licenses. Then come LLMs that uses code like mine to generate code or slop. My choice about license was made before LLMs. Now they use derivatives of my work without attribution as required by the license.

Even if I change the license, it won't be respected, and nothing else will change. To me, that's copyright infringement on me, ignoring the hours I put in.

On a separate note, LLMs only produce more of the same, ignoring the environmental and economic impacts. Often, more of the same is not what we need. It leads to the same mistakes being made. I've never used LLM to code, never will.

Which is absolutely legit. I only criticize the people releasing a “production ready” product, which is nothing but a PoC and then also claim that it is AGPL-3, while abandoning it after 2 months. This hurts the open source community imho.

oh no, how awful of them to release the code back into the wild for free with a copyleft license. truly an awful thing to do.

as a pirate, i shit on copyright, culture should be free. the issue with llm is not that they used copyrighted shit, if the llm would be open source and not owned by a corpo, i'd be contributing to knowledge. here, you're just contributing to some gringo westoid billionaire's pockets. everything is a remix, i dont have an issue with that. so again, the issue is labour theft in marxists terms.

This hurts the open source community imho.

Nah. At worst the open source community is indifferent. If someone can pick up where this left off then it benefits the community.

All "generated code" hurts open source, actually.

while abandoning it after 2 months. This hurts the open source community imho.

I dont expect them to maintain it for eternity. The problem is people expect too much from the authors as if it's the author's duty to fix bugs and add features. And tbh, this attitude hurts open source community more than anything. Ffs, the code is open source, do what you like. The author is not responsible for everything. Be grateful that the code is open-source.

How? Genuinely curious.

Badly written code hurts open source, but humans are just as guilty of that as machines.

I am not talking about the many projects, where you have a single developer just enjoying making software, while providing it to the open source community. I have zero expectations to those projects.

I am talking about people promoting their open source products. Making a fancy website. Act professionally. All this behavior causes people to have expectations. So why invest time into promoting it in the first place? There is literally no need to do this, expect for the owner to attract more and more people.

Don’t promise a sky scraper if you can only maintain a bungalow.

I just made an open-shoe! This is the first FOSS shoe ever. Everything is free but if you want both shoes, you'll need a subscription. OpenShoe is right foot only. Any foot, any size, durable and incredibly versatile. Left shoe only available in sizes M5, M7, M11.5, and M16.

Most videos I watch are in English, Spanish or German, I don't need an translation for it, less one by AI.

These are not mutually exclusive.

For starters, maintainers are getting bombarded with "contributions" that are just generated code, because now unskilled coders that can't even write a "hello world" project are just dumping what they think are great ideas without doing any kind of quality control at all.

Just having to be educated enough to even build a project was a very important barrier to the amount of submissions.

Now any Tom, Dick, or Harry can submit endless amounts of shit to a huge amount of different projects.

Maintainers can't keep up with the review process for this level of garbage. This has gotta be killing huge amounts of projects, or at least slamming the brakes on them.

Personally I'm totally in favor of some kind of foundational killswitch within compilers that just detects AI output and refuses to compile. Now that just kicks the can down the road because other compilers exist, but you get what I'm saying.

I can't be arsed to trust any AI code AT ALL. So every single character of it is poison IMHO, but given what it's currently doing to the industry already (and the fact that we're still a single digit amount of years into its existence at all), nothing good causes this many problems.

We're a lot closer to solving capitalism than we are to making any of the effort out towards AI being useful. Until then, any attempt at making it good will just be helping fascists have more tools at their disposal.

It goes way deeper than "but it helps me do stuff". That opinion honestly reads as if you're gen alpha.

And even in a perfect world where capitalism is barely remembered as a bump in the road to civilization, I will never be able to trust any code not written by humans. The kids will call that whatever the millennial insult version of "boomer" is, but I hope that attitude never goes away.

If you're doing your own thing, and you wanna add in some slop code to finish up a little change to something, you go right ahead. But don't you dare submit that to any actual open source project.

I mean you have to have a better reason to ban AI from open source projects than just bad code.

Limits on submission size and frequency could be placed on Human and AI equally.

Humans don't typically submit massive amounts of code frequently. Now any script kiddie-level idiot is out there feeling like they're a goddamn wizard. Meaning there a LOT more submitters vibe "coding", hence the submission problem.

Every single character should be handcrafted, and the author should know what every piece is doing (and how it's doing that) at any point you can point at. Skipping the hard part helps NO ONE.

Even WANTING to use AI to produce code just reeks of the same energy that high schoolers give when relying on Cliff Notes to write a book report.

Not system-wide audio, but per-tab: there's a browser extension that puts subtitles over whatever is playing in the tab, transcribed as it plays. It's in the Chrome and Firefox stores now, free and AGPL like the rest. It won't replace OBS + localvocal for anything outside the browser, but there's no filter chain to set up — install it, open the tab, hit Start.

Two things to set expectations. It runs behind the audio: it transcribes in windows, and on integrated graphics that's tens of seconds, so the overlay prints how far behind it currently is rather than pretending it's live. And it only works where the page itself plays the media — a site embedding someone else's player in an iframe, or a file served without CORS headers, can't be captured at all, and it tells you so instead of sitting there.

On permissions: it asks for no site access when you install it. When you hit Start it requests that one site, and nothing else.

Honest answer: for files that size, use the desktop app rather than the web page. The browser build holds the decoded audio in memory, so multi-GB files are where it starts to struggle, and I haven't tested anything close to 10 GB in a tab.

The desktop app (macOS, Windows, Linux) exists for exactly this — long files, no browser in the way — and it can use large-v3-turbo (1.6 GB, 99 languages) instead of the small models the web version is limited to. Two things worth knowing before you download: the builds are unsigned, and burning subtitles into the picture needs an ffmpeg built with the ass filter.

Files are processed one at a time, so a 50-hour fan edit split across many files is a batch job rather than one run.

There is a desktop app already — macOS, Windows and Linux. It's Tauri rather than Electron, so it's a much smaller download, and it's meant for the cases a page is bad at: long files and batches, with large-v3-turbo instead of the small models the browser is limited to.

Worth knowing before you try it: the builds are unsigned, and burning subtitles into the picture needs an ffmpeg built with the ass filter (Homebrew's plain ffmpeg formula doesn't have it; the Windows build offers to install one for you).

That depends much more on the model size than on the year. In the browser we're limited to small models — tiny (English only), base and small, 40 to 250 MB — and on a German TV show those are exactly where Whisper gets shaky.

The desktop app and the CLI can run large-v3-turbo (1.6 GB, 99 languages), which is a different experience on non-English speech. If your last try was a small model, that's the part worth changing before judging it.