woodson

21 ms·

271.
by woodson 7y ago
As another practitioner, I find it odd that the author mentions frameworks such as OpenNMT but not Kaldi when comparing SST toolkits/frameworks. Overall, I get the feeling that the author hasn't worked with speech data for too lon
272.
by woodson 7y ago
Twitter’s web client is a PWA
273.
by woodson 7y ago
It means you constantly have to re-login on a PWA (e.g. Twitters web client PWA).
274.
by woodson 7y ago
> There are no huge libraries full of C code like pandas or scipy. Isn't Julia's LinearAlgebra a wrapper around BLAS/LAPACK implementations (OpenBLAS/MKL etc.)?
275.
by woodson 7y ago
Already happening in Austria: (link in German) https://futurezone.at/netzpolitik/ausgangsbeschraenkung-a1-l...
276.
by woodson 7y ago
Are you going to release the speech data you collect at https://speech.talonvoice.com/ or is it proprietary?
277.
by woodson 7y ago
You mean like currently BLAS/LAPACK? Written in FORTRAN, wrapped by any scripting language, e.g. Python's NumPy/SciPy, Julia's LinearAlgebra (part of stdlib, https://docs.julialang.org/en/v1/std
278.
by woodson 7y ago
They provide useful packages that aren't (easily) available via pip. For example, numpy with MKL or Tensorflow GPU builds that don't require/presuppose CPU support for new vector instructions. Moreover, their package installs
279.
by woodson 7y ago
If you have access to the corresponding audio recordings, would you be able to share them?
280.
by woodson 7y ago
Including the audio? I downloaded the transcripts from their bucket, but couldn’t find any information on how to obtain the corresponding recordings. At Interspeech 2019, the author basically told me that they couldn’t share it.
281.
by woodson 7y ago
They can’t, because they are trained on more-or-less high-quality recordings of people reading books out loud. Phone conversations are very different, not just in audio quality but in the way people speak.
282.
by woodson 7y ago
Sadly they don’t (well, can’t) release the audio+transcripts as dataset, as they clearly don’t own the rights.
283.
by woodson 7y ago
Are you working on VC (input: speech of one speaker, output: the same spoken content, but sounds like another speaker) or speaker-adaptive speech synthesis (input: text, output: speech)? Also check out ParallelWaveGAN, another high-quality
284.
by woodson 7y ago
Kaldi is much much better, but when used for low latency recognition on the device, the accuracy will be lower than it could be because one would use much smaller models adapted to the constrained proccessing power of the device. It would s
285.
by woodson 7y ago
Thanks for sharing, that’s very interesting! Are Lisp and/or Prolog used in a lot of projects in the public sector in Österreich? From what I’ve heard, many (most?) projects at BRZ use Java or C# stacks.
286.
by woodson 7y ago
From your linked site: “A person who has obtained a U.S. Master’s degree or higher qualifies for this cap quota.” A foreign advanced degree doesn’t qualify.
287.
by woodson 7y ago
Wrong: There is no advanced degree pool. There is a "has a degree from some U.S. tertiary institution" pool. Doesn't matter how many Ph.D.s awarded by top universities in the rest of the world one has.
288.
by woodson 7y ago
> During the same time AMD was still surprisingly popular in Germany. Perhaps partly because AMD (and then later GlobalFoundries) has fabs in Dresden, Germany.
289.
by woodson 7y ago
As the name implies, LibriTTS is meant for people (researchers, really) to develop text-to-speech (TTS) systems. The point isn't that LibriSpeech isn't clean enough. Rather, it's that conversational speech is very different f
290.
by woodson 7y ago
True. But no ATM fees! (Or did that change?)
291.
by woodson 7y ago
Two years ago someone tweaked Kaldi to build to wasm ( https://github.com/adrianbg/kaldi.js ). AFAIR it ran at decent speed in a browser (with small models), but it hasn't been maintained since.
292.
by woodson 7y ago
Unfortunately, it’s not the solution – the discriminator is only going to work well on deepfakes produced by that particular GAN model, not the plethora of other possible models.
293.
by woodson 7y ago
The 20K are not for advanced degrees but for graduates of U.S. universities.
294.
by woodson 8y ago
That must have been some time ago, since Slovenia joined the Schengen area in 2003 (implemented 2007).
295.
by woodson 8y ago
Wish there was an ebook/PDF version of the book.
296.
by woodson 8y ago
That’s not quite true. When it came out, a 6s was about $649 and an SE was $399.
297.
by woodson 8y ago
I'm not sure if those models would work with the version in the current master (wav2letter++). The old version of the master branch (wav2letter, written in lua/torch: https://github.com/facebookresearch/wav2le
298.
by woodson 8y ago
Even more so if they compared it to Kaldi TDNN-LSTM with RNNLM lattice rescoring (test-clean 3.22%, apparently): https://github.com/kaldi-asr/kaldi/blob/master/egs/librispee...
299.
by woodson 8y ago
It has dependencies on CUDA and/or Intel MKL, so not really suitable as-is for embedded/phones.
300.
by woodson 8y ago
IMHO, the TIMIT corpus should no longer be used in most application-driven speech recogniton research, as it’s small and completely unrealistic for any real world application. Furthermore, nobody cares about phone error rates, as recognizin

More ›