LEARNABLE ACOUSTIC FRONTENDS IN BIRD ACTIVITY DETECTION Mark Anderson Naomi Harte SIGMEDIA Lab

2025-05-02 0 0 1.25MB 5 页 10玖币

侵权投诉

LEARNABLE ACOUSTIC FRONTENDS IN BIRD ACTIVITY DETECTION

Mark Anderson, Naomi Harte

SIGMEDIA Lab

School of Engineering

Trinity College Dublin, Ireland

{andersm3, nharte}@tcd.ie

reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or

reuse of any copyrighted component of this work in other works.

ABSTRACT

Autonomous recording units and passive acoustic monitor-

ing present minimally intrusive methods of collecting bioa-

coustics data. Combining this data with species agnostic bird

activity detection systems enables the monitoring of activity

levels of bird populations. Unfortunately, variability in am-

bient noise levels and subject distance contribute to difﬁcul-

ties in accurately detecting bird activity in recordings. The

choice of acoustic frontend directly affects the impact these

issues have on system performance. In this paper, we bench-

mark traditional ﬁxed-parameter acoustic frontends against

the new generation of learnable frontends on a wide-ranging

bird audio detection task using data from the DCASE2018

BAD Challenge. We observe that Per-Channel Energy Nor-

malization is the best overall performer, achieving an accu-

racy of 89.9%, and that in general learnable frontends sig-

niﬁcantly outperform traditional methods. We also identify

challenges in learning ﬁlterbanks for bird audio.

Index Terms—Bird Activity Detection, Bioacoustics,

Learnable Frontend, PCEN, LEAF, STRF, TD Filterbanks

1. INTRODUCTION

Monitoring of bird populations is of increasing importance as

biodiversity continues to decline and the impacts of climate

change become more evident [1]. Many population studies

are utilising passive acoustic monitoring and autonomous

recording units to record bioacoustic audio [2, 3], which

present non-intrusive methods for carrying out research on

bird populations [4]. The results of these studies can serve as

indicators of the effects of climate change [5].

Audio is highly suitable for monitoring bird populations

as many species are identiﬁable by their vocalisations and vi-

sual analysis can be difﬁcult. In the age of Deep Learning,

spectrogram representations of the audio are the most com-

mon input features to a system [6]. This form of representa-

tion is easily interpretable by humans and also allows Con-

volutional Neural Networks to extract spectro-temporal pat-

terns.

The primary parameters governing the spectrogram rep-

resentation of audio are the window size (determining the

time/frequency trade-off), frequency axis scaling and ampli-

tude compression. There is evidence that increased time res-

olution favours bird audio [7] but there is little consensus on

the best representation of the frequency axis, whether it is

linear, logarithmic or mel scaling [6]. There is further de-

bate on whether mel scaling is suitable for bioacoustics, as

it is derived from human auditory perception [8]. Bird au-

dio is typically in the range of 800-8000Hz, and mel scaling

focuses much of the energy in the lower bands. Log compres-

sion is loudness dependant and can increase the appearance

of noise in the spectrogram. The robustness of the chosen

representation to environmental noise and loudness variation

is paramount. In spite of this, standard log-mel spectrograms

used in human speech applications are ubiquitous in the anal-

ysis of bird vocalisations [6].

New trainable frontends have been developed for a range

of audio applications which can learn the ﬁlterbanks [9, 10],

perform loudness normalisation and noise reduction [11, 10],

or exploit patterns in temporal and frequency modulation

[12]. Per-Channel Energy Normalisation has been employed

recently in bioacoustics [13, 14] and the remaining frontends

have been tested on various bird audio tasks, including bird

activity detection. However, no dedicated comparative as-

sessment of the suitability of these methods in bird activity

detection has been done. As a vital ﬁrst step to any subse-

quent population analysis, this is crucial for judging their

suitability for bird bioacoustics work.

This paper compares traditional and learnable frontends

using the same datasets and model architecture. We evalu-

ate performance on a species-agnostic bird activity detection

task. The datasets contain a large variety of bird species and

call types, as well as many noise sources (e.g. wind, trafﬁc,

human activity and speech). We provide the most comprehen-

sive, up-to-date investigation on the suitability of learnable

frontends in bird audio detection to date. Section 2 details the

frontends. In Section 3 we outline our methodology for eval-

uation. In Section 4, we present and discuss the experimental

results and Section 5 concludes the paper.

arXiv:2210.00889v1 [eess.AS] 3 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

LEARNABLEACOUSTICFRONTENDSINBIRDACTIVITYDETECTIONMarkAnderson,NaomiHarteSIGMEDIALabSchoolofEngineeringTrinityCollegeDublin,Irelandfandersm3,nharteg@tcd.ie©2022IEEE.Personaluseofthismaterialispermitted.PermissionfromIEEEmustbeobtainedforallotheruses,inanycurrentorfuturemedia,includingreprinting/repub...

展开>> 收起<<

LEARNABLE ACOUSTIC FRONTENDS IN BIRD ACTIVITY DETECTION Mark Anderson Naomi Harte SIGMEDIA Lab.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

LEARNABLE ACOUSTIC FRONTENDS IN BIRD ACTIVITY DETECTION Mark Anderson Naomi Harte SIGMEDIA Lab

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: