ldenoue

This thing is amazing can you add recording?

Perhaps some samples you or visitors create?

Then add a little sampler for beat and it’s a fantastic tool

This is the version of my open source library that many people liked https://news.ycombinator.com/item?id=42238890

This will give you 120 free minutes to try out this LLM powered YouTube correction.

LLM corrected transcript (using Gemini Flash 8B over the raw YouTube transcript) https://www.appblit.com/scribe?v=YD-9NG1Ke5Y#0

Did you see this using the API or the online product gemini?

It’s not uploaded anywhere: the client fetches Gemini servers directly from your browser.

But I understand it can be difficult to trust: that’s why the project is on GitHub so you can run it on your own machine and look at how the key is used.

I will try to offer a version that doesn’t require any key.

Although Gemini accepts very long input context, I found that sending more than 512 or so words at a time to the LLM for "cleaning up the text" yields hallucinations. That's why I chunk the raw transcript into 512-word chunks.

Are you saying it works with 70B models on Groq? Mixtral, Llama? Other?

Do you have something to read about your study, experiments? Genuinely interested. Perhaps the prompts can be made to tell the LLM it's specifically handling human speech, not written text?

I should do that, let me try.

The current Flash-8B model I use costs $1 per 500 hours of transcript.

Definitely: and just giving the LLM context before correcting (in this case the title and description of the video, often written by a person) creates much better transcripts.

My repo doesn't re process the audio track: instead it makes the raw ASR text transcript better by feeding it additional info (title and description) and asking the LLM to fix errors.

It is not perfect, it'd sometimes replace words with a synonym, but it is much faster and cheaper.

The low cost of Gemini 1.5 Flash-8B costs $1 per 500 hours of transcript.

I also created something web-based last year called ScreenRun https://screenrun.app

It really works best in Chromium-based browser which have WebCodecs.

I was playing with Gemini this weekend to solve that issue and your video transcript converted into malt.

Here’s the result: Clean YouTube Transcripts appblit.com Created by @ldenoue , this experiment uses the Google Gemini gemini-1.5-flash-8b model to clean up YouTube transcripts, which lack punctuation and paragraphs.

This whiskey news report highlights a variety of new releases and announcements from various distilleries. Notable releases include a 200th-anniversary Mallen time-space collection, a Glenmorangie ice cream-inspired whiskey available in the Netherlands and Belgium, a share purchase of a whiskey distillery group by Kari, a Big Peat 15th-anniversary edition, Kingsbar Distillery's first peated whiskey, a Johnny Walker bottle with environmental benefits and reduced weight, an Octo Series release, and a redesigned Killhumman 100% Islay. Other announcements include new wood finishes from Tomatin Distillery, a redesigned Shivers Regal bottle, a 35-year-old Glenallachie, an independent Highland Park bottling from Duncan Taylor, a new Redbreast 18-year-old, a Jimmy Russell bourbon, Angels Envy triple cask release, Jack Daniel's Single Barrel special release, and a 25th-anniversary SL Distillery release. The report also promotes upcoming live tastings on whiskey.com and whiskey.de.

Hello and welcome to the whiskey.com news on September 23, 2024.

We have several interesting news items.

First, Mallen is celebrating its 200th anniversary with the Time Space Collection. This collection, with two or three bottlings, includes "Space," a blend of two single malts. The outer ring is from a 1940 vintage (84 years old), and the core is from a 2018 distillation at the new distillery. Only 200 sets are available. Pricing inquiries should be directed to Mallen.

Next is the Time Space Mastery, a non-age-statement bottling with 14 different cask finishes.

We also have news I've already covered, concerning the Glenmorangie "Tale of Ice Cream," a sweet, creamy whiskey with honey and notes of lavender. It's available at whiskey.com for consumers in the Netherlands and Belgium.

A follow-up on the Distel Group news: The group, which includes Buachaille, Habanero, Deanston, and Toomore distilleries, has formed a subgroup. Kari recently purchased a 14.6% stake for $92 million USD. The group is now part of Cape Vin, a South African holding company.

Further news: Big Peat 15th Anniversary Edition, a predominantly red wine cask finish with 50% ABV, is available at whiskey.com in the Netherlands and Belgium.

Kingsbar Distillery is launching its first peated whiskey, the Kingsbar Coaltown. This new core range whiskey is matured 100% in ex-peated casks, resulting in a peated flavor profile. The non-age-statement bottling has 46% ABV.

Johnny Walker is introducing an environmentally friendly, lightweight whiskey bottle. Weighing only 180g, this bottle will drastically reduce the environmental impact of glass production. It holds a new variant of the Blue Label. Only 888 bottles will be available in 2025. The company plans to share the lightweight bottle technology in the future.

The Octo Series 15 (151, 152, 153) is also a highlight. These highly-peated whiskies are expensive and, according to the report, some editions may be quickly sold out. Check whiskey.com for availability.

Killman is also launching a 100% Islay 2024 bottling. This 9-year-old (at least), non-chill-filtered, uncolored bottling with 50% ABV is a redesign of the 100% Islay series and includes an age statement.

Tomatin Distillery is releasing three new wood-finished bottlings: Madera, Amarone, and Tempano cask finishes. These 10 bottlings are available at 40% ABV and are also available at whiskey.com for consumers in the Netherlands and Belgium.

Shivers Regal 18-year-old is receiving a new, slimmer, and taller bottle design with reduced glass weight and intensified coloring. The release date is scheduled for the end of September.

Glenallachie 35-year-old, Billy Walker's oldest Glenallachie whisky, is a blend of three PX sherry hogsheads, one oloroso puncheon, and two virgin oak barrels with 48% ABV. No coloring or chill filtration. Available at whiskey.com.

Independent bottler Duncan Taylor is releasing the Accolade Highland Park 1970. This 54-year-old Highland Park is available in 158 bottles. The cask has been in Duncan Taylor's possession since its inception. It boasts 42.9% ABV and sherry cask maturation.

From Ireland, Red Breast 18-year-old, an Irish single pot still whiskey, is now a permanent addition to the core range, expanding from the core range four years prior. It's available at 46% ABV and matured in Cream Sherry casks.

Turning to the United States, Wild Turkey is partnering with bourbon legend Jimmy Russell. A special 8-year-old bourbon, the Jimmy Russell, will be released internationally this fall. It has a 50.5% ABV.

Angels Envy is expanding its Signature Series with a new triple cask finish (Hungarian Oak, Spanish Campin Oak, and French Oak). The release is scheduled for October. This 92 proof (46% ABV) bottling will be available in the US and, hopefully, internationally.

Jack Daniels is releasing a Single Barrel special release called Koi Hill. Bottled at cask strength, the whisky is between 61 and 68.75% ABV depending on the warehouse level. It's approximately 11 years old.

Finally, SL Distillery in Bavaria, Germany, is celebrating its 25th anniversary with a Distillers' Choice Edition. Using five undisclosed casks, this bottling boasts 48.8% ABV with no coloring or chill filtration. Available at whiskey.de and whiskey.com.

Upcoming Live Tastings: Live tastings will be held on whiskey.com and whiskey.de. The first, featuring Patrick, will be on October 31st at 7:00 CEST (whiskey.com) and 8:00 CEST (whiskey.de). Live tasting kits are available for purchase on whiskey.com.

Thank you for watching.

I use firebase auth with Google, Facebook and email (magic link)

This is live at https://screenrun.app/

Working on restoring punctuation marks in raw YouTube transcripts. Bert model trained as a token classifier works but it doesn’t scale well with more languages. See https://www.appblit.com/scribe for the current solution

Now looking into fine tuning a small LLM like Gemma 2b-it to this task.

Any advice on other LLMs that would work appreciated.

Also bonus if this LLM can run in a browser

Same here: I work on speech for a big company and working from the office is terrible. I had to squat the restroom 3 times yesterday so I could work (talk to my phone!)

At home I am so much more productive and zero commute.

Because of the badging policy in place, I end up scheduling non productive days at the office (doing email, reading other docs, meetings which are always with remote folks anyway but at least they see a genuine meeting room or phone booth behind my pretty face, so I guess that counts? ;)

You can also try Scribe (free chrome extension and iOS app) https://www.appblit.com/scribe

The new ai voices feature (text box below the stage) lets you generate real sounding audio from any text you type. But you can also record your mic and have speech to text recognize the text.

In both cases the subtitles appear over your video.

Looks like you’ve used ScreenRun for your video demo, thanks!

So that would explain why people are interested in tools that help manipulate pdf files?

My issue with PDFs was the fact they are hard to read on smaller screens because the format isn’t natively easy to reflow.

I guess another issue today is RAG for AI: how can a pdf be chunked in meaningful pieces. Funny to see old methods such as document layout analysis being used for RAG.

Actually we can run these networks today on any WebGPU enabled browser. If no WebGPU then transformersjs will still work (but slower). I use for example a DistilBert token classifier to restore punctuation in YouTube transcripts in Scribe: everything happens locally.

Demo https://www.appblit.com/scribe

I did something like this with Scribe where the formatting actually happens locally in your browser using a small token classifier based off MobileBert https://www.appblit.com/scribe

Yep the title is misleading. Here’s the Gemini summary https://g.co/gemini/share/efe3dc1870aa

I also built ScreenRun, a video editor with zooms and fake clicks running 100% client side with web codecs. You can try for free https://screenrun.app/

Readable transcript https://www.appblit.com/scribe?v=5t1vTLU7s40

Now you can also watch the video, and drag it anywhere you want.

Demo in my tweet https://twitter.com/ldenoue/status/1788231159140196804

Just updated the app so export to PDF now shows your highlights colored.

Quick hack: just replace youtube.com with appblit.com to see the transcript of any YouTube video.

Example: https://www.youtube.com/watch?v=0AaNT7XO41I https://www.appblit.com/watch?v=0AaNT7XO41I

The site embeds the YouTube video so we can hear the audio, so I could show the video as well, just not sure where :)

Perhaps split in 2 columns on wide displays (video left, transcript right), and portrait (e.g. mobile phone), video on top and transcript below it?