DeepGram (YC W16) Is Building a Google for Audio | Hacker News Reader

DeepGram (YC W16) Is Building a Google for Audio

blog.ycombinator.com

41 pointsascertain28 comments

Many speech tech startups are sued endlessly over patents [0], primarily by Nuance, and then when things look hopeless, Nuance offers to buy and make all the lawsuits go away [1]. How do you view this situation, and how will you fight the lawsuits?

[0] http://techcrunch.com/2011/06/09/nuance-sues-vlingo-again-ov... [1] http://techcrunch.com/2011/12/20/after-years-of-patent-litig...

We're not familiar with the behind-the-scenes details of those cases, but we're not competing with Nuance in the transcription game. We are doing really killer audio search and view competition through that lens.

I'm pretty sure it's in their best (legal) interest to not comment at all, doing so could potentially imply that they are aware that they are infringing patents, which would be legal seppuku.

EDIT: IANAL and the above is not legal advice but I have some familiarity with patent law

Also, it's questionable whether Nuance has any grounds in many of its suits (as is the case many other patent trolls, pun not intended), but a small startup won't have the resources to fight it regardless (or construct an ice rink in front of a certain courthouse in a humid subtropical climate). Unethical legal bullying.

That seems like an insane strategy. Everyone how gets acquired there must be rooting for the company (or at least the executives and management) to fail.

Insane?

http://www.bizjournals.com/boston/blog/mass-high-tech/2008/0...

List of all acquisitions: https://en.wikipedia.org/wiki/Nuance_Communications#Nuance_a...

In the Boston tech scene, it is pretty well-known what Nuance does and why, and is one of the primary reasons I am very against software patents, despite having several to my name.

The patents identified there seem like they're on very loose footing. Firstly they're all from halfway the previous decade ... about 50 years after voice recognition was invented (and cute detail: even their client-server voice patent doesn't quite work because the first IBM speech recognizers ran on a cluster out of necessity, therefore they had "remote voice endpoints").

Everything else is hopelessly out of date. Today voice recognition happens with DNNs or LSTMs. None of those patents would apply to such methods. They seem to have attempted to patent obvious applications of voice recognition.

http://www.google.com/patents/US6839669

Obviously identifies their old product (and I'm pretty sure I can find prior art for this one) where you could report your own voice and have it, say, open notepad when you say "open notepad". It's about making a computer respond to direct commands, and audio prompts. Seems to me like this would (should ?) never apply to a system that transcribes audio. Any modern interaction system would not be about responding to commands, except in an extremely broad sense.

A page filled with a list of prior art: https://books.google.com.au/books?id=f3IV90zLmaEC&pg=PA209&l...

https://www.google.com/patents/US6785653

Is about interactions with users based on prompts. Essentially a scripted interaction with people

Again this was filed in 2000. There has got to be prior art for this one. I remember using a system that did this before I entered high school, in 1994. And that was a pretty mature program, I bet when it comes to demos and papers you should be able to find systems doing this in the 70s.

Obvious prior art, as the computer is obviously using prompts (e.g. the purchase confirmation) : https://www.youtube.com/watch?v=I7q1cE9_AaQ1 (date: 1989)

https://www.google.com/patents/US7058573

Is about methods to adjust past recognized words on semantic meaning. Seems to be talking about this sort of thing "I was about two" => "I was about to". A neural net recognizer doesn't do this at all.

Also obviously prior art : this IBM engineer explains that this is exactly what he's doing. Judging by that computer this is early 80s: https://www.youtube.com/watch?v=cpdm1O-ob48

http://www.google.ch/patents/US7127393

Is about limiting voice recognition results based on a set of rules. Why would anyone work things this way in 2016 ? It is far easier to just transcribe and respond conversationally to the transcription (or just train a network to respond directly, negating the need for actually having any rules at all. You'd just show the system what to do).

Interesting video: https://www.youtube.com/watch?v=uw8XbmzOG5s

I posted this comment five days ago. It seems like audio indexing and searching might be the next big thing.

> I've been wanting to make something like this for a while. Really nice product you have! I think this kind of service will really take off in the future. Imagine having an app that constantly records everything and allows you to search it later. Questions like "What did James tell me last November about traveling to Europe?". It would also eliminate hearsay, since you would no longer have to trust one person's word against another — you could simply search the transcript of that moment in the past. In the very long run, I almost wonder if such a tool would make lying obsolete.

I'm Noah, one of the founders. I also think lifelogging will be common in the future because it's becoming easier to record everything and the benefits of having that dataset are also increasing (that's where we come in) You're correct that lying could become obsolete. It could be sort of like the changes brought on by the Information Age, where now it's pointless to argue over truths that are the public domain (since we can just look up the fact whenever we need). It would extend that to private life.

Have you seen "The Entire History of You" episode of Black Mirror?

It's one of our favorites :) I don't think it would be so bad to record our lives. We would be more efficient scientists, journalists--explorers of information. We could focus on the here and now instead of taking notes.

Or we would be stuck focusing on the best [and the worst] of past because it's persistent, like in the episode.

Reliving memories seems like a much more powerful drug when streamed directly to your retina just as you experienced it.

I'm pretty sure it requires a whole other level of nostalgia to be immersed and get lost in transcribed texts of the audio of an event.

Not that I'm saying the latter is what the tech will remain like, but at least at first that's what it will probably look like.

This story by Ted Chiang [1] is an examination of exactly the future you describe. Forget lying to others, this could prevent our leaky, revisionist, memories from lying to ourselves.

[1] https://subterraneanpress.com/magazine/fall_2013/the_truth_o...

Relevent discussion: https://news.ycombinator.com/item?id=10163666

This is awesome

That's terrifying. Making lying obsolete is not worth having everything I do recorded.

Good way to waste your battery.

I have always wondered how patents and copyright apply to deep learning networks. It is pretty obvious that the NMIST deep learning convolution network uses the same algorithm patented in letter separation patents by OCR companies. And yet ...

I wonder if you were to go further. Suppose you let a sufficiently deep neural network see all simpsons episodes for instance. And it manages to make new simpsons episodes. This will at some point become possible. At that point, does it count as a new work ? Does it fall under simpsons copyright ? Where is the line ?

I would love to see a computer-generated simpsons episode. I bet you could do a basic version using DeepDream applied to the audio and video of a different video / tv show. I imagine the voices and images would morph to sound more simpson-like.

Looks great! I created a simple NodeJS wrapper for it: http://bit.ly/1YYUbzX

Awesome, I'll take a look!

Awesome! Thank you. :-)

Thanks!

How does this compare to http://clarify.io?

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!

The idea is similar but the accuracy is much better.

This is incredible. BIG ups.

Reply on news.ycombinator.com