stephensonsco

Thanks! I'm glad we got __something__ to play with. This is about a day of crawling + processing.

Glad you are having fun with it! That's why we put this version out. Have fun collecting flair ;P

What?! You don't like click bait-y titles that point to POCs?!! You are totally right, it's a POC. <<slink away>> Just a project we put out really, really early but hope it's fun/interesting.

Indexing the entire web's videos and presenting results well is a big computational+UI problem that we are still working on. We mostly help businesses get value out of their audio/video with our API. Full scale search engine is a ways away.

It's about the same in topology — process audio, search through that, get back timestamps — but we don't rely on transcribing the words correctly.

Bummer. The index size is pretty small right now but we're working on increasing it right now. What were you searching for?

The sexy part refers to the looks — sorry if our sense of humor is not translating well :(

Yep, we use deep learning to generate an index of audio features and an approximate search to look through it all. We've done lots of testing on both clear and average/noisy audio. With standard transcription matching you get ~alright results (50% retrieval) on clear audio but really terrible results on average/noisy audio (<20% retrieval). With deepgram you get great results on clear audio (90% retrieval) and still really good results (80% retrieval) on noisy/average quality content.

We're always trying to make the indexing and search better/faster. For Hoogley we had to pull some tricks to get the search to be really fast for a lot of users, so it isn't quite as high quality as it can be. Working on it though ;)

We started indexing gobs of Youtube videos and made the world's most luxurious, diamond encrusted video search engine. The audio is stripped and digested with Deepgram (that's us—hi!). When you search, you're brought right to when people are talking about your precious little search term.

The index is 4 million+ seconds of video and growing.

Apply now. Then immediately resume working on it. If you don't get in now, then apply with your awesome prototype that you've built over 6 months to the next batch. :)

The important difference is that DeepGram is a fuzzy search that can find matches even if a text transcript is wrong. This is what ups the accuracy very significantly.

I'm the other cofounder of DeepGram. I'm glad we're getting everybody's feedback!

Glad you like it. We're curious to see what contraptions people come up with.

This is really cool!

We made a python module to run complicated non-python code in python (like stuff that needs ridiculous environmental gymnastics). It's called sidomo (Simple Docker Module) and it's used to easily make python modules that are actually docker containers (well, docker images). These containers take an input, then hit it with the contained code, and send the output back to pure python.

The hello world:

  from sidomo import Container
  with Container('ubuntu') as c:
      for line in c.run('echo hello from the;echo other side;'):
          print(line)

Bummer about the alien-speak! The problem should be fixed now ;).

I think the point is the assembling of random photos/videos into a convincing representation of textures and animation (without devoted hardware, bringing the person in for a scan, etc), not that 'this is bleeding edge visuals'.

Yes! I have lazily been waiting for something like this to come along and hopefully bump out metapost. Maybe this has already happened and I am out of the loop on that though.

I love the idea -- whisperer, meal and all. Did you punish the disobedient GoPro?

You've got it. Deep neural networks are currently the best performers when doing automatic speech recognition.

You should be able to hit enter in the search field on your phone's soft keyboard (or the button might just say search). Which phone/operating system is this?

The first thing that comes to mind is podcasts. That are a slightly different direction, but a great fit for our service. They're long, have well spoken words, and are regularly released, so there is a mound of searchable audio data out there just waiting to be liberated.

Not yet but this is a pet project that I would love to get going (time strapped now, though). I used the Lexika API to get the times from query results (i.e. when the words were said), then used some hacked together python along with ffmpeg with a lot of auto-generated seeking command line arguments to splice together the video feed. Then the title screens were done in a normal editor.

I would love to make this an app though. Just search for phrases, choose clips, and output a cute video. You could just have a thousand people saying a "thousand", or something equally weird.

Thanks for the transcription feedback. You can look at this as a problem but we think it's a cute demonstration of our search ability. We didn't really set out to make transcriptions perfect (that's very, very hard), we wanted to make search very accurate. Having a better transcription model would be great though (we are working on that!).

We'll definitely work on the site. (We should have done that earlier.) Thanks again!

Yeah for sure! This site is basically a demo/test for our backend.

We built an API that processes audio and forms an index for that file. The API search function then goes into that index to look for queries. It doesn't just look for words that match the text you see, but also the way it sounds.

I am loving making mashup videos with it. Trump talking about walls: https://www.youtube.com/watch?v=F38EwKOmDdU

Yeah, good point. We knew about this but the accuracy wasn't all that great. Our algorithm is fuzzy and can find results even if they are transcribed incorrectly (even today, every speech transcription technology out there still makes plenty of mistakes).

Ah, that would be really nice to do. Our speech transcription only does English at the time but we definitely have our eye on other languages.

We do two things simultaneously. If there is no closed caption, then the video is submitted to our API to be processed (which takes longer on longer videos - a link shows up on the page to let you know the future URL). If there is a CC, then we parse that right away (which gives ok search results) but still submit to our API for processing, so later the search results will purely be from our indexing.

You are right that you need to input the particular video for now, but we are expanding to include multi-video search (constantly indexing in the background). Right now the use case is particularly nice for finding phrases in long videos (look for Donald Trump saying random stuff {wall, tremendous, many many, women}, it is pretty comical - e.g. http://www.lexika.io/?s=https%3A%2F%2Fwww.youtube.com%2Fwatc...).

This is really interesting. One my biggest problems is extracting cash out of stuff that I am too busy to sell. Printing a label and packing is a bit of a hassle still though. Make that easy and I'll make it rain.

Really nice and the speed is impressive. A useful feature for me would be a subtle drop down of close matches/other possibilities.