nshm

22 ms·

181.
by nshm 10y ago
https://github.com/kaldi-asr/kaldi/blob/master/egs/fisher_sw... In Microsoft paper http://arxiv.org/pdf/1609.03528v1.pdf Table 5 it's a line "Povey et al. [19] LSTM&q
182.
by nshm 10y ago
Open source Kaldi gives you 7.8%, Microsoft didn't went too far. Also, major issue with this kind of research is that they combined several systems in order to get best results. Most practical systems don't use combinations, they
183.
by nshm 10y ago
Single microphone won't do that. You need a microphone array.
184.
by nshm 10y ago
It is not the case. Evernote was created by https://en.wikipedia.org/wiki/Stepan_Pachikov who also founded Parascript company http://www.parascript.com , which was doing OCR long before Abbyy and is mostly s
185.
by nshm 10y ago
Arduino probably won't stand radiation for a long. In case people interested, in Russia there is a crowdfunded project running to send a space probe to the moon.
186.
by nshm 10y ago
LIME is a software to analyze any machine learning classifier https://github.com/marcotcr/lime , they have very good paper in KDD too http://www.kdd.org/kdd2016/papers/files/rfp0573-ribeir
187.
by nshm 10y ago
As a developer I'm not quite happy with a single-letter names. Of course it's ok for minor local things to be named 'x' but for more important values and functions there is no problem these days to have a readable names
188.
by nshm 10y ago
You can use Speechmatics directly
189.
by nshm 10y ago
It is easy to introduce different measure which will make deep learning fail and which will improve the state of the art and humanness significantly. They are: 1) Learning from small amount of examples, make training database very small wit
190.
by nshm 10y ago
You need to try CMUSphinx http://cmusphinx.sourceforge.net . In keyword spotting mode it is very lightweight. Python example is here: https://github.com/cmusphinx/pocketsphinx/blob/master/swig&
191.
by nshm 10y ago
Read the book: Automatic Speech Recognition A Deep Learning Approach http://rd.springer.com/book/10.1007%2F978-1-4471-5779-3 . Also read another book http://www.amazon.com/Spoken-Language-Processing-Algo
192.
by nshm 10y ago
Congratulations to the team. Intel does not have a track record for excellent software unfortunately, but who knows, they seem to be quite serious about IOT.
193.
by nshm 10y ago
Exactly, this simple example demonstrates how hard is to learn even a simple function. Playing with your code, 10 hidden units give error 0.6 in 1000 iterations, 1000 hidden units give error 0.02 in 10000 iterations so it takes way longer t
194.
by nshm 10y ago
Oh yeah, our deep learning is a super cool technology, no need to study math, believe us, just buy more GPUs and they will help you, throw the data on the fan. Or, even better, give us your data, we'll decide what to do with it for you
195.
by nshm 10y ago
Strictly speaking if you split the parameter set on batches and iterate over batches optimizing each set of parameters with a gradient, it is not strictly a gradient decent, it is more a combination of coordinate decent (because you select
196.
by nshm 11y ago
Looks like a research in machine learning will only be done in huge corporations. You'll need an amount of funding comparable to LHC. Time to use better models like kernel ensembles, maybe they are not that accurate, but they are easie
197.
by nshm 11y ago
Absolutely, and even more, you can not do modern research unless you are part of Google. Only big corporations have computing power and data access to do large scale data research with neural networks. This basically rules out a lot of te
198.
by nshm 11y ago
Any of them have this information available in some form - openmary, festival. For example in Festival, you can access synthesized utt markup with utt.something functions like this: (utt.save.segs (utt.synth (Utterance Text "Hello worl
199.
by nshm 11y ago
I participated in Gnome and Linux kernel development in 2002-2008, that was an amazing time. Those days open source systems were successfully competing with commercial software, Firefox was on edge, Linux desktop was really promising, Linux
200.
The Continuing Silence of Bourbaki (1997)(docs.google.com)
32 points by nshm 11y ago|6 comments
201.
by nshm 11y ago
Worth to note that Librivox also gives a significant boost to text-to-speech and speech recognition research, currently librispeech corpus based on librivox is the largest and quite useful corpus of spoken English. http://www.dan
202.
Tiny, blurry pictures find the limits of computer image recognition(arstechnica.com)
2 points by nshm 11y ago|0 comments
203.
by nshm 11y ago
The effort definitely not minimal. You have to collect 10000 transcribed hours first. This is not easy in many languages.
204.
by nshm 11y ago
Sorry, nothing definite yet. Any sort of tree-based computation which makes early pruning decision is more efficient than full matrix multiplication, even if matrix is sparse. Then, high-order computation like power is again more efficient
205.
by nshm 11y ago
Those interested in optimal neural network compression might consider the paper "Bitwise Neural Networks" by Kim and Smaragdis http://paris.cs.illinois.edu/pubs/minje-icmlw2015.pdf which enables much better c
206.
by nshm 11y ago
Sure, but they aren't necessary the most efficient model possible
207.
by nshm 11y ago
Mark my words, DNNs are not really the most efficient structure for predicting model due to the "distributed" representation that makes them good predictors, but also makes them hard to train and resource-consuming to apply. In fe
208.
by nshm 11y ago
It is not necessary to use phonemes in the model, however, it greatly restricts model's ability to learn, in particular to learn many special cases like rare foreign words. You need much more data to train the model and basically you l
209.
by nshm 12y ago
You can use CMUSphinx http://cmusphinx.sourceforge.net for keyphrase verification. For example you can find Android demo for keyphrase spotting at http://cmusphinx.sourceforge.net/wiki/tutorialandroid
210.
by nshm 12y ago
CMUSphinx - open source speech recognition system http://cmusphinx.sourceforge.net

More ›