Loading...
Publication
Separação de fontes de vocalizações ultrassónicas no estudo comportamental de murganhos
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 11 MB | Adobe PDF |
Advisor(s)
Abstract(s)
A análise de vocalizações permite uma compreensão mais profunda do estado emocional
do locutor, sendo influenciada pelas variações emocionais. Este fenómeno não
ocorre apenas em humanos, o mesmo acontece noutros animais, como é o caso dos
murganhos, Mus musculus, utilizados para estudar padrões psicológicos através das
suas vocalizações. Devido às suas características, as comunicações que estabelecem
são realizadas em frequências superiores a 20 kHz, sendo denominadas de Ultrasonic
Vocalizations (USV).
Este projeto visa ajudar no estudo das USV de murganhos, quando não existem
múltiplos microfones. A leitura com apenas um microfone faz com que todas as
vocalizações estejam misturadas, levando a que todas as análises sejam mais demoradas
e manuais. Desta forma, o projeto tem como objetivo ajudar em situações
onde seja estudado um cenário entre progenitora e cria. Assim foi planeado e criado
de um programa capaz de identificar os dois grupos de murganhos: a projenitora e
as respetivas crias.
O programa criado teve como base a mistura de duas áreas da análise de vocalizações,
a de USV de murganhos e a da diarização de locutor. Para tal, o programa
teve de ser capaz de obter características das vocalizações e separá-las em grupos,
utilizando um método de clustering. Desta forma, utilizou-se Mel Frequency Cepstral
Coefficients (MFCC) para obter as características do sinal, que foram utilizados
em dois métodos de clustering, o Gaussian Mixture Models (GMM) e K-means. No
final, foi projetado um gráfico do espectrograma do áudio, onde se pode verificar os
resultados de ambos os métodos.
Os resultados obtidos foram promissores, mas não definitivos. Os dois métodos
experimentados tiveram resultados semelhantes, analisando corretamente por volta
de 80% das vocalizações das crias, 75% das da progenitora e 60% em situações que
estão os dois a comunicar.
Apesar do sucesso nos resultados, não deixa de ser necessário melhorias em projeções
futuras, onde seja feita uma análise com mais dados, seja realizado um tratamento
mais cuidado dos dados para a obtenção de características ou sejam abordados
outros métodos de clustering que melhor se possam adequar, podendo passar por métodos mais avançados como por Deep Learning.
The analysis of vocalizations allows a deeper understanding of the speaker’s emotional state, as it is influenced by emotional variations. This phenomenon does not occur only in humans; it is also observed in other animals, such as mice, Mus musculus, which are often used to study psychological patterns through their vocalizations. Due to their specific characteristics, their communications occur at frequencies above 20 kHz, known as Ultrasonic Vocalizations (USV). This project aims to assist in the study of mice USV where multiple microphones are not available. Recording with only one microphone results in mixed vocalizations from different individuals, which makes all analyses more time-consuming and manual. Therefore, this project was designed to help with this analysis stage, specifically in scenarios involving a mother and her pups. As such, a program was planned and developed to identify the two groups of mice. The program created is based by combining two areas of vocalization analysis: mouse USV and speaker diarization. For this, the program needed to be able to extract vocalization features and group them using a clustering method. Thus, Mel Frequency Cepstral Coefficients (MFCC) was used to extract signal features, which were then used in two clustering methods: Gaussian Mixture Models (GMM) and K-means. Finally, a spectrogram of the audio was plotted, where the results of both methods can be observed. The results obtained were promising, but not definitive. Both methods produced similar results, correctly analyzing around 80% of the vocalizations made by the pups, also 75% that were made by the mother and only 60% of those that were made by both simultaneously. Despite the success of the results, it’s suggested that improvements should be done in future studies. Such work could include analyzing more data, refining data processing for feature extraction or exploring other clustering methods that may be more suitable, including advanced methods such as Deep Learning.
The analysis of vocalizations allows a deeper understanding of the speaker’s emotional state, as it is influenced by emotional variations. This phenomenon does not occur only in humans; it is also observed in other animals, such as mice, Mus musculus, which are often used to study psychological patterns through their vocalizations. Due to their specific characteristics, their communications occur at frequencies above 20 kHz, known as Ultrasonic Vocalizations (USV). This project aims to assist in the study of mice USV where multiple microphones are not available. Recording with only one microphone results in mixed vocalizations from different individuals, which makes all analyses more time-consuming and manual. Therefore, this project was designed to help with this analysis stage, specifically in scenarios involving a mother and her pups. As such, a program was planned and developed to identify the two groups of mice. The program created is based by combining two areas of vocalization analysis: mouse USV and speaker diarization. For this, the program needed to be able to extract vocalization features and group them using a clustering method. Thus, Mel Frequency Cepstral Coefficients (MFCC) was used to extract signal features, which were then used in two clustering methods: Gaussian Mixture Models (GMM) and K-means. Finally, a spectrogram of the audio was plotted, where the results of both methods can be observed. The results obtained were promising, but not definitive. Both methods produced similar results, correctly analyzing around 80% of the vocalizations made by the pups, also 75% that were made by the mother and only 60% of those that were made by both simultaneously. Despite the success of the results, it’s suggested that improvements should be done in future studies. Such work could include analyzing more data, refining data processing for feature extraction or exploring other clustering methods that may be more suitable, including advanced methods such as Deep Learning.
Description
Keywords
USV Murganhos Clustering GMM K-menas MFCC Sonogramas Mice Sonograms