nshm

15 ms·

121.
by nshm 6y ago
One very important positive thing is that the whole readme is now displayed on mobile version and that is the one which is indexed by Google. Previously they only shown few top lines and thus all README was not visible in the search engine.
122.
by nshm 6y ago
Would be nice to test something opensource alongside. Like https://github.com/alphacep/vosk-api which runs on Android and iPhone offline.
123.
by nshm 7y ago
https://en.wikipedia.org/wiki/The_Eyes_of_Darkness definitely which tells the story of Wuhan-400 virus.
124.
by nshm 7y ago
You can check https://github.com/festvox/datasets-CMU_Wilderness , it has recordings of 700 languages created from New Testaments from http://www.bible.is/
125.
by nshm 7y ago
> the provided scripts are useless to anyone without access to the academic datasets they use Kaldi team shares all datasets when possible on http://openslr.org , a major collection of speech datasets. Librispeech and Tedlium
126.
by nshm 7y ago
Three is actually the only choice for harmony. > What makes two vases on a table less "balanced" than three? Two leads to dialectic competition between them. > What about a single vase? Static world > Four? Five? Chaos
127.
by nshm 7y ago
Worth a mention they are fighting a patent troll right now https://voicebot.ai/2020/03/04/mycroft-ais-legal-war-against...
128.
by nshm 7y ago
Even earlier there was Jasper https://github.com/jasperproject/jasper-client/commit/96a608... (2014)
129.
by nshm 7y ago
Those are popular science magazines. They had both scientific content and some science fiction too.
130.
by nshm 7y ago
You are welcome to try Vosk https://github.com/alphacep/vosk-api Advantages are: 1) Supports 7 languages - English, German, French, Spanish, Portuguese, Chinese, Russian 2) Works offline even on lightweight devices - R
131.
by nshm 7y ago
Ok, if you want to start with Kaldi it is probably easier to check kaldi-active-grammar mention above or https://github.com/alphacep/vosk-api
132.
by nshm 7y ago
Among all demos above Serenade seems the only product that does the right thing. Why the hell I should say aloud "colon" or "quote" where tool can put them automatically.
133.
by nshm 7y ago
Why would you work with google when there are much more accurate open source speech recognizers based on Kaldi? With that specific usecase it is very easy to beat Google on accuracy.
134.
by nshm 7y ago
Brain doesn't do backpropagation
135.
by nshm 7y ago
Thank you for the link, very exciting research. I definitely agree that there are separate activities to formulate the theorem and to prove already described one. The former requires some kind of creation and imagination and might not be au
136.
by nshm 7y ago
Till the point when computers will create thousands of proofs a man will have trouble to understand. We will join Lee Sedol team then. Thats kinda weird the article doesn't even mention the important work around the subject already pub
137.
by nshm 7y ago
The limits of a single man knowledge and understanding are perfectly real and discussion on how to overcome them is very important. I'm not a Steve Wolfram fan, but his answer in the article is much deeper than others. As for pure proo
138.
by nshm 7y ago
So, VS Express 2010 was 500Mb, then VS 2019 Express required 11 Gb for installation. When it has grown to 100Gb they decided to move it in the cloud. I said thank you and compiled everything with MINGW on Linux. Apt installed mingw in 30 se
139.
by nshm 7y ago
Being nice to others is not problem if you do not need anything from them. But if you want to force someone to do what he doesn't want to do, that is a problem. And being empathic doesn't always help.
140.
by nshm 7y ago
There is no single framework as you see, to dive fully you need to explore all of them. Each has some useful features.
141.
by nshm 7y ago
It more depends on features you already implemented. Check the arxiv file, if you do not have all those features already (lookahead lm, proper sentencepiece, label smoothing), consider this espresso.
142.
by nshm 7y ago
Everyone race to build an ASR toolkit these days, this one is a good candidate actually compared to others. A good modern technology with subword pieces, parallel training, fast decoding. Very competitive decoding accuracy. Disadvantages ar
143.
by nshm 7y ago
Exactly, also a frequent API changes for no reason. Tensorflow is a good example of this. Looks like every month they schedule a meeting on what they gonna break next. They seem to enjoy thousands of github issues about cryptic messages on
144.
by nshm 7y ago
I had that story long time ago. I've got A-DATA MyFlash FP1 stick with fingerprint unlock. Somehow it didn't work on Linux. I sniffed the USB and figured out that after fingerprint check it simply sends a fixed 3-byte command to t
145.
by nshm 8y ago
Abstractive summarization not yet working (like text generation in chatbots), extractive is ok and better done with RNN/BERT/something else neural network based, that's all you need to know.
146.
by nshm 8y ago
I think its on purpose to not give too much advantage to downloaders, there was a discussion long time ago, but they still continue with mp3. There are also other not so nice points: https://github.com/kaldi-asr/kaldi&#
147.
Microsoft’s New MT-DNN Outperforms Google BERT(medium.com)
121 points by nshm 8y ago|38 comments
148.
by nshm 8y ago
The journalist failed to explain "Laplace transform" part unfortunately. How something like complex variable operation happens in the brain. Most likely the original researchers meant something different.
149.
by nshm 8y ago
Just 4 frames per second sounds crazy slow.
150.
by nshm 8y ago
I like the way they politely skipped Kaldi WER on test-clean (4.31) in results table. Their best WER (4.91) will not look so attractive.

More ›