Hacker News Reader

adrianbgon Speech Recognition Is Not Solved

Here are some benchmarks on telephone speech, including both APIs and human transcription services:

https://remeeting.com/app/benchmarks

Google actually did pretty badly for us on extended telephone speech. Not sure why.

adrianbgon Speech Recognition Is Not Solved

I agree. I think a big part of the reason that people use WER is that it's relatively unambiguous and easy to measure.

adrianbgon Speech Recognition Is Not Solved

Totally. You may want to take a look at the papers from CHiME4 for more along those lines:

http://spandh.dcs.shef.ac.uk/chime_challenge/chime2016/resul...

I'm really fascinated by the whole idea of blind source separation and the fact that speech signals are "sparse" in frequency space.

We've had a similar experience looking for hardware / open source beamforming. There's a package called beamformit, but I think it's pretty old.

adrianbgon Speech Recognition Is Not Solved

The 8.5% in this file is what you'd compare to Microsoft and IBM's recent ~5% results.

https://github.com/kaldi-asr/kaldi/blob/master/egs/fisher_sw...

Kaldi hasn't been in first place on that dataset recently, but it was a few years ago.

On other more researchy datasets (eg. for distant speakers or languages other than English), the best system is often based on Kaldi.

adrianbgon Speech Recognition Is Not Solved

Unfortunately that Tensorflow integration didn't include acoustic modeling, so you still need to use Kaldi's neural net toolkit for that.

adrianbgon Speech Recognition Is Not Solved

Exactly. This is also why Google sponsors the CHiME challenge, the existence of which is more proof that ASR is pretty far from solved.

http://spandh.dcs.shef.ac.uk/chime_challenge/

adrianbgon Speech Recognition Is Not Solved

The problem wrt open source / free solutions is data. Kaldi is open source and gets state of the art results -- but the data costs a lot of money. Training the models is doable on a commodity GPU although it takes quite a while.

adrianbgon Speech Recognition Is Not Solved

Hah sure. By "hard," I meant that it's the largest hurdle. And probably even the common research datasets aren't enough to give you results competitive with Google, etc. AFAIK Google uses its own hand-transcribed data.

adrianbgon Speech Recognition Is Not Solved

Although some models are end-to-end neural nets, most of the ones in production (and all of the ones that get state of the art results) only use a neural net for one part of the process. Lots of people are as surprised as you, but that's the way it is.

Edit: I should say that in state of the art results there tend to be multiple components, including multiple neural nets and the tricky "decode graph" that gok and I are talking about. These are trained separately then get stuck together, as opposed to being trained in an end-to-end fashion.

adrianbgon Speech Recognition Is Not Solved

You are correct that the problem is data. Kaldi is hard to use, but making it easier to use isn't as hard as getting good training data. Mozilla's project is a good start for some purposes. One flaw with it is that they're having people read sentences. When people read, they tend to speak more clearly than when they're figuring out what to say on the fly. This means models trained with Mozilla's data will tend to need people to speak extra clearly than eg. a model trained on conversational data.

(edits: spelling/grammar)

adrianbgon Speech Recognition Is Not Solved

Speech to text is exactly speech recognition. Speech to text isn't solved.

adrianbgon Speech Recognition Is Not Solved

The low WER numbers you've probably seen are for conversational telephone speech with constrained subjects. ASR is much harder when the audio source is farther away from the microphone and when the topic isn't constrained.

adrianbgon Speech Recognition Is Not Solved

If you think that's bad, check out the Geordie accent from Newcastle: https://www.youtube.com/watch?v=ZY4TT3VtR8o

According to Wikipedia it's "a direct continuation and development of the language spoken by the Anglo-Saxon settlers" of the region. https://en.wikipedia.org/wiki/Geordie

adrianbgon Speech Recognition Is Not Solved

CMUSphinx is really old. Kaldi is hard to use, but it's much better.

adrianbgon Show HN: Serverless speech recognition with WebAssembly

Alexa's recognition is made more difficult by the fact that you talk to it from a distance. Still, though, you're right that only handling predefined commands should make it much easier.

adrianbgon Nonlinear Computation in Deep Linear Networks

I wonder if this / gradient descent would work with integers, and with mod 2^n as the nonlinearity.

adrianbgon AllenNLP – An open-source NLP research library, built on PyTorch

The "Who did What" dataset seems much better in this respect:

Passage: Britain’s decision on Thursday to drop extradition proceedings against Gen. Augusto Pinochet and allow him to return to Chile is understandably frustrating ... Jack Straw, the home secretary, said the 84-year-old former dictator’s ability to understand the charges against him and to direct his defense had been seriously impaired by a series of strokes. ... Chile’s president-elect, Ricardo Lagos, has wisely pledged to let justice run its course. But the outgoing government of President Eduardo Frei is pushing a constitutional reform that would allow Pinochet to step down from the Senate and retain parliamentary immunity from prosecution. ...

Question: Sources close to the presidential palace said that Fujimori declined at the last moment to leave the country and instead he will send a high level delegation to the ceremony, at which Chilean President Eduardo Frei will pass the mandate to XXX.

Choices: (1) Augusto Pinochet (2) Jack Straw (3) Ricardo Lagos

https://arxiv.org/abs/1608.05457

adrianbgon The DAO close to $100m in funding. Soon to be the largest crowdfunding ever

Is this a cult?

adrianbgon Fraternal Order of Police Data Dump

Unions originated in response to maltreatment of workers. How would you solve that problem without them?

adrianbgon Designing Schemaless, Uber Engineering’s Scalable Datastore Using MySQL

Is it me or could they have done this way more easily by building some indexing and triggering functionality on top of Cassandra? Even two years ago when they started. Instead they built sharding, indexing, triggering and a Cassandra-like data model on top of MySQL.

adrianbgon My simplified response to Paul Graham's simplified essay

Really interesting. I wonder how politicians would rate.

adrianbgon Google, D-Wave, and the case of the factor-10^8 speedup

Not efficient at the start is a little different than "we have no reason to think it will ever be more efficient".

adrianbgon I am Peter Roberts, an immigration lawyer who does work for YC and startups. AMA

E2's have to be financed by citizens of your country of origin. http://startupimmigrationattorney.com/e2-visa-for-startups-p...

adrianbgon I am Peter Roberts, an immigration lawyer who does work for YC and startups. AMA

Yeah. Their investment comes in the form of a salary.

I don't know the details for H1b's specifically but I can tell you that they haven't had trouble meeting visa requirements for people in the program.

adrianbgon I am Peter Roberts, an immigration lawyer who does work for YC and startups. AMA

http://unshackled.co/

You can transfer your H1b to them and work on your thing.

adrianbgon Google, D-Wave, and the case of the factor-10^8 speedup

D-Wave's computers work as intended and exhibit "computationally relevant quantum effects" but we still don't know whether the strategy they are using will ever yield practical advantages. Other groups (Google, University of Maryland) seem likely to achieve genuine quantum speedups over best-known classical algorithms within the next few years.

adrianbgon Gamma rays from the Milky Way's core look ever more like signs of dark matter

Could it be that dark matter is stretched out along an axis that's invisible to us and gravity? Maybe the particles don't interact with each other or anything else because they're not actually that close to each other except when projected onto the dimensions familiar to us, and on which gravity acts.

adrianbgon Comcast injects JavaScript into webpages to show copyright notices to customers

None. The fact that everything goes through an extra hop causes a slowdown.

adrianbgon Jeff Dean explains TensorFlow [video]

Well, it looks way more scalable than Theano or Torch while being as easy to use as Theano. I'd say that's pretty exciting considering the number of groups working on way lower-level scalable neural nets.

This is "not a game-changer" in the same way map-reduce isn't a game-changer wrt for loops.

Also check out TensorBoard, their visualization tool (animation halfway down the page):

http://googleresearch.blogspot.com/2015/11/tensorflow-google...

adrianbgon Google to Fold Chrome Operating System into Android

Android and ChromeOS embody pragmatism and idealism. Android does everything everyone needs it to do but it's rotten on the inside because the focus is always on shipping. ChromeOS is a great vision for the future of detached computing but no one uses it because the vision hasn't been fully realized. I've been wondering whether this was going to happen ever since Chrome tabs on Android became equal citizens with the other apps. Seems like another good move from Google.