Wild Dolphin Venture 11-751 Discourse Last Venture

Presentation Transcript

Wild Dolphin Project 11-751 Speech Final Project by Jiazhi Ou jzou@cs.cmu.edu Tal Blum blum@cs.cmu.edu

Outline Wild Dolphin Project, Dolphin Speech Data, Labeling, Labeling issues Previous work Models preparing Experiments & Results Conclusions

The Wild Dolphin Project (WDP) The Wild Dolphin Project (WDP), established by Dr. Denise Herzing in 1985, is occupied with an aspiring, long haul logical investigation of a particular case of Atlantic spotted dolphins that live 40 miles off the shore of the Bahamas, in the Atlantic Ocean. For around 100 days every year, Phase I examine has included the capturing, recording, and sound taping of a gathering of occupant dolphins, expecting to find out about their lives. http://www.wilddolphinproject.org/index.cfm

Dolphin's Speech Dolphin's Speech is altogether different than man's discourse Range of frequencies is more extensive Two components for delivering sound at the same time Directionality of a portion of the frequencies Carried in water Can travel substantial separations

Dolphin's Speech(2) Is utilized for: Identification Communicating Fighting Defending Courting Warning Calling Hunting

Dolphin's Speech(3) 3 primary sorts Whistles Signature Non-signature Clicks Spike trains

What do we know Not much We realize that every dolphin has a one of a kind shriek called signature shriek. The mark shriek is like those that are in close contact with the infant dolphin

Data 164 records containing hints of one dolphin whose name is known. Normal record length is 7 sec Total information length under 20 minutes out of which about half is quiet The information does not contain the majority of the pertinent frequencies

Labeling Dolphin Names Dolphin ID extend Pause, Noise, Dolphin Signature Whistles, Dolphin Non-Signature shrieks.

Labeling Problems How would we recognize those 2 shrieks? How to recognize shrieks and non-shrieks? They co-happen How to decide the span of the name? Ought to close marks be named as one name? This affects the model Some signs are frail, most likely because of an adjustment in the dolphins bearing

Mapping from Labels to Models

Label Statistics

Previous Work Dolphin-ID Project by Tanja, Alan and Yue Task: To recognize dolphin ID utilizing their mark shrieks 51 named documents by Alan 13 HMMs: 10 for every dolphin + DOLPHIN, PAUSE, and GARBAGE Use Janus to do preparing and testing Try various types of elements

Our Work Model Generalized Signature Whistles Label More Files Create HMMs for mark shrieks, non-signature shrieks, waste, and respite Train and test the HMMs utilizing Janus Evaluate the test comes about with our own particular technique Compare diverse model choices

Signal Processing Tanja scripts Down inspecting High Pass Filter FFT LDA

b m e m e HMM Topologies Signature Whistles Non-Signature Whistles Garbage Pause (Water)

Model Selection Scheme 1 Signature Whistles, Non-Signature Whistles, GARBAGE, PAUSE Scheme 2 Signature Whistles, GARBAGE, PAUSE Scheme 3 10 HMMs (one for every dolphin), GARBAGE, PAUSE

Evaluation We can not utilize WER here since there are no words, just sections. The strategy we utilized was to register a perplexity network over shrouded states. Janus treat hush contrastingly and doesn't indicate quiet grouping which convolutes the assessment.

Experiments Data 162 marked records were utilized Half of the information for preparing, half to test Swap the preparation set and test set 162 test outcomes every together Feature The same as those in dolphin-ID extend Model Selection 3 unique plans

Results – Scheme 1

Results – Scheme 2

Results – Scheme 3

Analysis of Results You can just get in the same class as your names Scheme 3 is the best to adjust signature shrieks - speaker subordinate Scheme 1 is the most exceedingly terrible – insufficient information to display non-signature shrieks and waste Scheme 2 is in the center – speaker autonomous Pause is the most hard to show – It contains every single distinctive thing. We displayed it with just 1 state

Conclusion Analyzing dolphin sounds is very unique in relation to breaking down human discourse. The techniques utilized must be conformed to the attributes of the dolphin sounds. There is a considerable measure of work to be done in the flag preparing stage Partly managed preparing It may be better just to develop a model for the names we are certain and let the model realize what are mark shrieks or units that separate between various names.

We additionally attempted … One-state show for non-signature shrieks, waste, and respite - Segmentation blame in preparing "Circle back" model for mark shrieks - The circle back move has no effect

Acknowledgment Tanja Schultz Yue Pan Alan W Black Szu-Chen Stan Jou Hua Yu

Thank You! Jiazhi Ou Tal Blue {jzou, tblum}@cs.cmu.edu