notes.md

Notes

for windows:

1. http://www.lfd.uci.edu/~gohlke/pythonlibs/#numpy

• get the scipy whl and numpy blas whl whl

Spectrograms

TODO:

NOW

• make init script nicer for someone else to use
• tests for learning.py - v important

LATER

• produce a graph at the end for the test data
• Dockerise so others can easily develop (no real way of testing with scientific python packages with travis at the moment, unless docker is used)
• Turn TODO list into github issues

Would be Cool

• make the fpython stuff nicer to use (currently in the bin folder of env, setup needs to add it to the folder)
• think of ways of making the experiments even more lightweight

The Non-vocal baseline:

• 2 methods - spectrograms and raw waveform.

Papers and misc. Information

• http://cs231n.stanford.edu/reports/Cs_231n_paper.pdf
• https://papers.nips.cc/paper/3674-unsupervised-feature-learning-for-audio-classification-using-convolutional-deep-belief-networks.pdf
  • 20ms window size, 10ms overlap
  • PCA whitening, (80 components)
• http://research.microsoft.com/pubs/230136/IS140441.PDF
• http://benanne.github.io/2014/08/05/spotify-cnns.html
  • used mel scale for frequency and logarithmic amplitude scaling

https://www.youtube.com/watch?v=qv6UVOQ0F44

Datasets That Look Useful:

• RML emotion database: promising, although maybe a bit small, could go well with another corpus.

• Montreal affected voices corpus: emotional vocalizations (laughing, crying etc.)

• Italian affected voices corpus

• Toronto Emotional Speech Set:

• Berlin Emotional Speech Database:

• RAVDESS: Currently in use

• Surrey Audio-Visual Expressed Emotion (SAVEE) Database

• Vera am Mittag German Audio-Visual Spontaneous Speech Database.

• Estonian Emotional Speech Corpus http://peeter.eki.ee:5000/

The ones below I looked at, but were not useful

• buckeye corpus: Long conversations of american english - not much metadata

• voxforge: Lots of recordings, none of it about the users emotions/stress levels, therefore likely not to have much range.

• Santa Barbara Corpus of Spoken American English: detailed but with no stress/emotion.

• Spoken Language Corpora at the Research Center on Multilingualism: same as the UCSB

• The Spoken Turkish Corpus at METU Ankara: Also normal corpus

• Spoken Corpus Klient with the Corp-Oral Corpus at ILTEC Lisbon

• OLAC: Open Language Archives Community

• BAS Bavarian Archive for Speech Signals

• Simmortel Speech Recognition Corpus for Indian English and Hindi

• ELRA: the European Language Resources Association

• The PELCRA Conversational Corpus of Polish