Hello,
I have been developing an algorithm using DSP techniques to select a dominant speaker in a conference in which a cluster of smartphones have been used each with microphone and speaker enabled.
These smartphones can be placed in any orientation with a minimum distance between them be 1 foot.
I have researched a lot in order to find something like that, but mostly I get microphone arrays etc, but not smartphones (iOS and Android) specific. Also, this cluster of phones is distributed cluster, connected wirelessly.
Some points to keep in mind:
1. Android and iOS have different hardware. So far what I concluded is that, if we go into features like Energy, power, RMS, peaks, dB etc. They don't give that steady result since hardware is changed and each phone computes energy of same voice (for testing same recorded voice) at different levels.
2. Used normalization techniques ,but it didn't help in that way.
3. Read somewhere that when you dont know the min max value of certain dataset Z-score is best option, also tried to implement by using circular buffer to store pevious one second of data. It improved selection but not upto expectation. Still gives varying results.
Any help will be appreciated.
Regards,
Khubaib Ahmad
visit the following link below:
https://tel.archives-ouvertes.fr/tel-01883828/document
http://sensorlab.cs.dartmouth.edu/pubs/miluzzo-phd-dissertation.pdf
- Badges
- Science method