micai13 turdus migratorius (2)

15
Emotion based features of bird singing for Turdus migratorius identification Toaki Villarreal, Caleb Rascón and Ivan Meza Universidad Nacional Autónoma de México http://golem.iimas.unam.mx/ GrupoGolem

Upload: grupo-golem-dcc-iimas-unam

Post on 15-Jun-2015

46 views

Category:

Technology


1 download

DESCRIPTION

Talk given in MICAI 2013: Emotion based features of bird singing for Turdus migratorius identification

TRANSCRIPT

Page 1: Micai13   turdus migratorius (2)

Emotion based features of bird singing for Turdus migratorius identification

Toaki Villarreal, Caleb Rascón and Ivan MezaUniversidad Nacional Autónoma de México

http://golem.iimas.unam.mx/

GrupoGolem

Page 2: Micai13   turdus migratorius (2)

Motivation

● Ecosystems change constantly and with this its inhabitants

● In the case of birds, to understand such changes specialists monitor populations

● Bioacoustics monitors by means of the singing of birds

Page 3: Micai13   turdus migratorius (2)

Challenges

● It is time consuming

● It requires a specialist

● Recording área limited

Page 4: Micai13   turdus migratorius (2)

On the other hand

Automatic emotion recognition in speech has the goal:

With a recording of speech recognise the emotional state of the speaker

Page 5: Micai13   turdus migratorius (2)

Our goal

With a recording of bird singing recognise the bird species

*We are interested in the bird species, not the emotional state of the bird

Page 6: Micai13   turdus migratorius (2)

Proposal

Develop a monitoring system for birds based on what we know for emotion recognition

At this point:

● We focus on Turdus migratorius● It has to be live and had a good performance● Module of Acoustic Identification of Birds

Page 7: Micai13   turdus migratorius (2)

Bird songs

Page 8: Micai13   turdus migratorius (2)

Emotion based: procedure

● Identify a candidate composed by multiple frames (usually turns in a conversation)

● Extract a representation per frame

● Use additive functions to represent the segment as a vector

● Use a classification technique

Page 9: Micai13   turdus migratorius (2)

For the birds

● Energy based sound activity, syllabifier

● Standard MFCC (13 valores), 1st derivation

● Mean, Std. var, 1st, 2nd and 3rd , min, max, skewness, kurtosis

● Support Vector Machine

Page 10: Micai13   turdus migratorius (2)

The system

Page 11: Micai13   turdus migratorius (2)

MIAA

Page 12: Micai13   turdus migratorius (2)

Evaluation performance

Segment based evaluation, recordings with only Turdus migratorius, GS labellings available

Precision 87.49%

Recall 75.15%

F1-score 78.30%

Precision 73.94%

Recall 64.64%

F1-score 67.34%

Syllabifier Classifier

Page 13: Micai13   turdus migratorius (2)

What about other species?

Segment based evaluation, no GSSix species: Turdus rufopalliatus, Myadestes occidentalis, Thryomanes bewickii, Cardinalis cardinalis and Toxostoma curvirostre

Precision 83.23%

Recall 83.23%

F1-score 83.23%

Page 14: Micai13   turdus migratorius (2)

Experiments

Which features are more helpful?

Page 15: Micai13   turdus migratorius (2)

Conclusions

● We were able to identify the Turdus migratorius

● Our approach is inspired by emotion based

identification systems

● We showed that MFCC are good enough

● The module is functional and works live on

the MIAA module

● … lots of future work