and, i-vectors do not really work for short utterances.
Recently a set of papers were published about "memorization" in neural networks. For example:
Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer https://openreview.net/forum?id=B1ckMDqlg
Understanding deep learning requires rethinking generalization https://openreview.net/forum?id=Sy8gdB9xx
It seems that large memory system has a point, you don’t need millions of computing cores in CPU and, it is too power-expensive, you could just go ahead with very large memory and reasonable amount of cores to access memory with hashing (think of Shazam or randlm, or G2P by analogy). You probably do not need heavy tying either.
Advantages are: you can quickly incorporate new knowledge, just put new values in memory, you can model corner cases since they are all still accessible, and, again, you are much more energy-efficient.
Maybe we will see mobile phones with 1Tb of memory sometimes.
Very relevant research to build the speech recognition in a language from public resources from the internet.
The cool thing is there is a leaderboard, so it is still possible to submit results:
Neural Speech Recognizer: Acoustic-to-Word LSTM Model for Large Vocabulary Speech Recognition
Hagen Soltau, Hank Liao, Hasim Sak
Unsupervised Feature Selection on Data Streams / Streaming Anomaly Detec...
Today, we see the use of streaming algorithms to figure out anomaly detection or unsupervised feature selection. Streaming Anomaly Detection
Solving the Cocktail Party Problem with a 3D Printed Metamaterial Disc
If you have ever tried to give vocal directions to your smartphone while amidst a group of voices, you know how hard it is for a computer to
Neural Word Embeddings as Implicit Matrix Factorization
Recently at the Paris Machine Learning meetup there was a brief presentation on Word2Vec by Charles Ollion. Well, I was wondering about the
How Crowdsourcing Will Help Startups Build Their Own Versions of Siri | ...
Speech recognition is hard, even for the world’s largest tech companies. Apple and Google draw on massive collections of recordings of real
Theory of Convex Optimization for Machine Learning / Estimation in high ...
Sebastien Bubeck just came out with a monograph on the Theory of Convex Optimization for Machine Learning while Roman Vershynin just release
Intel Pays Up To $30M For A Personal Assistant Platform From Ginger Soft...
Apple has Siri, and now Intel has Ginger. The chipmaker has made one more acquisition to bolster its advanced computing and artificial intel
With A Voice Interface API For Any App, Wit.ai Wants To Be The Twilio Fo...
Last year, voice technology giant Nuance quietly acquired VirtuOz, a developer of virtual assistants for online sales, marketing and support
Hot. Cool. Yours. Fin! Sochi Olympics close with breathtaking show
After two weeks of cheering, daring and record breaking, the Sochi 2014 Olympic Games finally bids farewell as athletes and fans gather one
Saturday Morning Video: Unraveling dolphin communication complexity: Pas...
For some odd reason, this video from last saturday on the Analyzing Animal Vocal Sequences Investigative Workshop could not play correctly.
Chrome hack lets websites keep listening after you close the tab
Toying around with voice-recognition apps, developer Tal Ater noticed something strange. Because of a quirk in Chrome's microphone settings,
Sunday Morning Insight: Randomization is not a dirty word
From  The recent announcement of Yann LeCun's appointment as a director of the new Artificial Intelligence Lab at Facebook and Geoff Hint
Machined Learnings: ICML 2013: Sparse, Deep, and Random
ICML 2013 was a great conference this year, kudos to the organizers. It's too big for an individual to get a comprehensive view of everythin
OpenEars 1.3.0 out now with Pocketsphinx and Sphinxbase .8 | Politepix
Sign up for the Politepix OpenEars frameworks mailing list here in order to receive infrequent notifications of when OpenEars frameworks (su
Speech recognition engine PocketSphinx landed in Ubuntu 13.10 by default...
Months ago, the developers announced and explained Ubuntu's converged vision, where a singular OS is to power phones, tablets, desktops, TVs