Legion Health (YC S21) Is Hiring Founding Engineers to Fix Mental Health with AI (workatastartup.com)

YTNT is a simple web-app for manually typing notes while watching a YouTube video. Simultaneously type and control playback with keyboard shortcuts. Export to Word (.docx) when you're done.

I'm seeking general feedback.

NOTES:

- The only way to "save" (for now) is to export to a Word .docx

- Speech-to-text only works in Chrome

GitHub: https://github.com/kavinaidoo/ytnt

Blog Post: http://archive.today/FRqwp

Thanks, Kavi

Comments (6)

Leftium · 15h ago

I forked oTranscribe and added some features to solve a similar problem.

- demo: https://otranscribe.netlify.app/?vsl=definedefine

- source code: https://github.com/Leftium/oTranscribe

- CLI tool to generate OTR (oTranscribe) files from (YouTube) SBV/TTML files: https://github.com/Leftium/otrgen

In my case, I wanted to start with the (auto-generated) YouTube transcript and get clickable timestamps. This makes it much faster to search through the content of a video: I can read/search much faster than watching a video, even on 2X speed.

I could also add my own notes to the transcript.

If you add support for loading transcripts like this, it could work cross-browser without the microphone/speech-to-text.

kavinaidoo · 5h ago

Aha, starting with the YouTube transcript is a great idea. I wanted to start with it but I couldn't find a way to get it from the YouTube iFrame API so I went with the Web Speech API to listen and convert.

Forgive me for the confusion, the demo link implies that the transcript is loaded from the video but I see in the code that there's a pre-existing "/txt/definedefine.md" that is loaded. How are these SBV/TTML files downloaded from YouTube in the first place? I assume that it is a separate process? I see you are then using otrgen to presumably convert these so they can be used by oTranscribe.

If I could load the transcripts dynamically when loading the YouTube video that would be a great feature.

Leftium · 4h ago

The demo simply demonstrates how I used my tool. It requires some manual set up:

- TTML files are downloaded via CLI: `yt-dlp.exe --skip-download --write-auto-sub --sub-format ttml`

- TTML files are converted to the OTR .MD format via my CLI tool

- The MD file can be dragged & dropped onto the web app.

---

I think it is possible to download SBV/TTML files, but the download must be done from the server due to CORS restrictions.

My app didn't go this far due to limitations of the (serverless) platform it is hosted on. Also it was faster to just do the steps manually vs. developing a server that does it.

I have seen many services that download the transcript. Here are a few:

- https://youtubetranscript.com

- https://kagi.com/summarizer

- https://www.tubepen.com

However, note YouTube may block your server if you download too many transcripts: https://kagifeedback.org/d/4451-universal-summarizer-cant-fi...

Leftium · 15h ago

I'm planning a beat-aware YouTube player. Unfortunately, it is not possible to access the audio stream data across the YouTube embed. (For beat-detection analysis.)

I considered using the microphone like this. It's nice to see it works. Although there seems to be a time limit to how long the microphone can record?

My plan was to download video (youtube-dlp), then make a CLI tool that analyzes and uploads the beat-detection data. An advantage of the CLI tool is it can complete the analysis faster than playing the video at 1X.

kavinaidoo · 5h ago

Yeah, I also wanted to "pipe" the audio directly to the Web Speech API but had to resort to using the mic. Another feature I wanted to add was to have a keyboard shortcut insert a screenshot of the video into the notes. Handy for diagrams etc but I hit some roadblocks there too and the workarounds were getting crazy.

Regarding the time-limit, I'm entirely accessing the audio stream using the Web Speech API and it "decides" (usually when there's a sufficient pause) when to finalize a recognition result. It was also firing a "recognition.onend" event after a certain amount of time (some old details here: https://stackoverflow.com/questions/38213580/chromes-webkits...) so I have a workaround where if the user did not turn off speech-to-text, it is immediately restarted. You'll see console warnings "userStopped == false - trying to restart" when this happens.

I assume you'll have no such issues because you'll be handling the "raw" audio data and not working through this API. Also, I've noticed that when using the Web Speech API on Safari, it does not "hear" what the tab is playing. It only hears external audio. I'm not sure whether this will be an issue for you but for me it means my app is Chrome only (suboptimal). Forgive me for my naïveté in this arena, It's the first time I'm using any browser audio API.

Would your tool run on a server and work with a frontend for YT link ingestion or will you just use it yourself from the cli?

Leftium · 4h ago

Ideally, my tool would run on the server. Especially if I wanted to monetize the service. (However I think there may be legal issues that are larger than the technical ones...)

However it's just a small hobby project. So realistically I think this is how it will work:

- If user tries to load a video without beat data yet, instructions are shown for how to add the beat data. The instructions will be running CLI command like `npx upload-beat-data [YOUTUBE-URL]`

- The CLI tool will download the audio, then upload the beat data to my site.

- The site will also log which videos got a lot of requests, but are missing beat data. So I can manually add the beat data myself.

I hacked a dating app (and how not to treat a security researcher) (alexschapiro.com)

Embeddings Are Underrated (technicalwriting.dev)

The Barbican (arslan.io)

RIP Usenix ATC (bcantrill.dtrace.org)

Launch HN: ParaQuery (YC X25) – GPU Accelerated Spark/SQL

A community-led fork of Organic Maps (comaps.app)

University of Texas-led team solves a big problem for fusion energy (news.utexas.edu)

Legion Health (YC S21) Is Hiring Founding Engineers to Fix Mental Health with AI (workatastartup.com)

Show HN: Airweave – Let agents search any app (github.com)

Demonstrably Secure Software Supply Chains with Nix (nixcademy.com)

Byte Latent Transformer: Patches Scale Better Than Tokens (arxiv.org)

Why GADTs matter for performance (2015) (blog.janestreet.com)

Reviving a Modular Cargo Bike Design from the 1930s (core77.com)

Ruby 3.5 Feature: Namespace on read (bugs.ruby-lang.org)

Tailscale 4via6 – Connect Edge Deployments at Scale (tailscale.com)

OpenEoX to Standardize End-of-Life (EOL) and End-of-Support (EOS) Information (openeox.org)

Universe expected to decay in 10⁷⁸ years, much sooner than previously thought (phys.org)

Spade Hardware Description Language (spade-lang.org)

Show HN: CLI that spots fake GitHub stars, risky dependencies and licence traps (github.com)

I ruined my vacation by reverse engineering WSC (blog.es3n1n.eu)

The FTC puts off enforcing its 'click-to-cancel' rule (theverge.com)

Optimizing My Hacker News Experience (reorientinglife.substack.com)

A Typical Workday at a Japanese Hardware Tool Store [video] (youtube.com)

Continuous glucose monitors reveal variable glucose responses to the same meals (examine.com)

Ash (Almquist Shell) Variants (in-ulm.de)

Continuous Thought Machines (pub.sakana.ai)

Traffic Fatalities Are a Choice (asteriskmag.com)

Intellect-2 Release: The First 32B Model Trained Through Globally Distributed RL (primeintellect.ai)

Armbian Updates: OMV support, boot improvents, Rockchip optimizations (armbian.com)

Making PyPI's test suite faster (blog.trailofbits.com)

Implicit UVs: Real-time semi-global parameterization of implicit surfaces [pdf] (baptiste-genest.github.io)

Plain Vanilla Web (plainvanillaweb.com)

Why Bell Labs Worked (1517.substack.com)

CrowdStrike CEO cuts his voting power by 92% with unexplained gifts (bloomberg.com)

Car companies are in a billion-dollar software war (insideevs.com)

Ex-UK Special Forces break silence on 'war crimes' by colleagues (bbc.com)

Absolute Zero Reasoner (andrewzh112.github.io)

High-school shop students attract skilled-trades job offers (wsj.com)

Scraperr – A Self Hosted Webscraper (github.com)

Writing an LLM from scratch, part 13 – attention heads are dumb (gilesthomas.com)

The Audio Stack Is a Crime Scene (fireborn.mataroa.blog)

Title of work deciphered in sealed Herculaneum scroll via digital unwrapping (finebooksmagazine.com)

How friction is being redistributed in today's economy (kyla.substack.com)

I hacked my clock to control my focus (paepper.com)

US Copyright Office found AI companies breach copyright. Its boss was fired (theregister.com)

Gig Companies Violate Workers Rights (hrw.org)

A simple 16x16 dot animation from simple math rules (tixy.land)

Ask HN: Cursor or Windsurf?

ToyDB rewritten: a distributed SQL database in Rust, for education (github.com)

LSP client in Clojure in 200 lines of code (vlaaad.github.io)

Show HN: YT NoteTaker – Simple Manual Note Taking on YouTube Videos

Comments (6)