Pretty much SOA; most end-to-end systems use these spectrograms on short time slices. The alternative is mel-frequency cepstral coefficients, which are used more in GMM-HMM speech recognition than for DNN.

For other illiterates like me:

* SOA = state of the art

* GMM-HMM = Gaussian mixture model - Hidden markov model

* DNN = deep neural networks (more than 1 hidden layer)

* mel-frequency cepstral = MFC ;-)

Could sum1 write a GM script to replace all abbr.s with their ffs? The over use of abbr.s is one of the most XABs on Hacker News.

Could someone write a Grease Monkey script to replace all abbreviations with their full forms? The over use of abbreviations is one of the most extremely annoying behaviours on Hacker News.

(Thank you https://www.allacronyms.com/aa-search?q=annoying&cx=01082138...)

How does this work when the spectrogram is finite time slice?!

Lots of overlapping. It's a sliding window function. Ballpark for most algorithms: 10 ms of new audio, 90 ms of old audio.

Why not just use time-domain convolutions?

What's the difference between the false-color frequency spectrum image and the frequency spectrum itself? Assuming that there's one horizontal pixel per sample, one vertical pixel per frequency bin, and that the color spectrum encodes the intensity to sufficient resolution, I'm not sure there is any difference.

The bitmap is just a convenient representation of a two-dimensional array, and PNG is a convenient compression library!

CNN on spectral images can work well. But most SOTA uses LSTM for speech recognition

Thank you! LSTM: http://colah.github.io/posts/2015-08-Understanding-LSTMs/

That's right. This is not state of the art. LSTM is good for sequences and sequential data (like audio). But with the approach hereby described, prediction can be done in parallel and asynchronously. Which is something ;)