Dear Colleagues,
I have a few basic questions concerning audio/speech/voice features and their extraction. I will be grateful if you can send me detailed answers plus references to the most relevant books and papers:
(1) What file formats are the most recommend to work with from the viewpoints of: quality of the voices, quality and number of extracted features, and available extraction programs?
(2) What are the most recommend available (free/commercial) programs that help to extract audio features and the web-links to these programs?
(3) What are the audio features that we can extract from such program(s)?
(4) What are the classification(s) for these audio features?
(5) Do we have low-level and high-level audio features? If yes, What are they?
Thanks in advance and best regards,
Yaakov
Dear Dr. HaCohen-Kerner,
Matlab is probably an all purpose swiss-army-knife software for speech analysis. Another software is Praat http://www.fon.hum.uva.nl/praat/.
Praat can open/edit several sound files such as WAV files, AIFF files, AIFC files, FLAC files, and MP3 files. However, to my knowledge WAV is the most common format.
Praat can be used for spectral analysis (spectrograms), pitch analysis, formant analysis, intensity analysis, jitter, shimmer, and voice breaks. In addition, the software incorporates speech synthesis facilities; It allows labelling and segmentation of speech sounds.
More advanced features of this software include the change of pitch and duration contours and filtering. Because Praat has its own scripting language, it also highly programmable,
References
Paul Boersma & David Weenink (2013). Praat: doing phonetics by computer [Computer program]. Version 5.3.51, retrieved 2 June 2013 from http://www.praat.org/
See also Paul Boersma's page for additional references: http://www.fon.hum.uva.nl/paul/praat.html
Kind Regards,
Charalambos Themistocleous
you can use this sotware
link : http://www.fon.hum.uva.nl/praat/
you can find ;
Speech analysis:
spectral analysis (spectrograms)
pitch analysis
formant analysis
intensity analysis
jitter, shimmer, voice breaks
cochleagram
excitation pattern
you can also use goldwve to record your audio file.
the audio format depends on what you want to do.
I recommend to use known database like in : NIST
you will find more information in
Daniel Ellis homepage
you have to define what you want to do : I mean the applications:
Speech recognition
Speaker recognition
Audio segmentation
Langage recognition
etc,
for example, you use classifier like GMM , neural network , Ivector in speaker recognition.
But every field has his own best classifier , we can't generalize.
yes , we have low-level and high-level audio features
I'll limit my comments to features, as I can't recommend specific programs (although pretty much everything has open source code available nowadays).
The basic features that are used for speech synthesis, having been inspired by a simple model of biological speech generation) are the LPC coefficients. They have many forms, and can be converted via linear transformations into other feature sets such as PARCORs and LSFs. These features model the state of the vocal tract. There are modified versions, such as mel or bark transformed LPCs, and various mechanisms to make the features more robust.
Obviously spectral information, such as can be obtained by pitch and format tracking, can be useful. Unfortunately, it is harder to accurately obtain, and often harder to use. Once again, it is often used on a mel or bark scale.
The cepstral coefficients are a great tool for speech analysis. A particular variety - the LPC cepstrum, is known to have a useful property, namely that the Euclidean distance in LPC cepstrum space is highly correlated with the Itakura-Saito distance. LPC ceptral coefficients are frequently used in both speech and speaker recognition.
There are acousto-phonetic features that are very useful in research (and perhaps out-perform other feature sets), but tend to be rather voluminous for practical application.
Finally, a newer approach is not to extract arbitrary features at all, but to learn which features are optimal for a given task (e.g., using deep neural networks).
Y(J)S
Most of fellow researchers have already given the suggestions. Can you explain the purpose of feature extraction that will help everyone to suggest you better.
Hello thanks for thes given suggestions and links.
But I don'y understand still now what are the audio features that we can extract from such program(s)?
When I take a file.wav, what'll I have in output after feature extraction. is it a matrix with attributes(which) and lines(which)? How?
Jaudio will provide 72 feature values in csv or arff file format. Praat can provide pitch or intensity listing in text file and so on..
Thanks Makarand Velankar do you have a kind of script or paper relating how to extract feature in PRAAT?
praat documentation and examples are self explanatory.
Many videos on youtube and links as mentioned below are available.
http://www.fon.hum.uva.nl/praat/
See http://www.AudioContentAnalysis.org for low level feature extraction scripts for Matlab.
I have came across Sonic Visualiser software, can I use this software to extract features from singing voice (without music)?
Please refer to the tutorial on Voice Features.
https://www.youtube.com/watch?v=TGcXyn_-IZo
- Badges
- Science topic