A VES ANIMAL VOCALIZATION ENCODER BASED ON SELF-SUPERVISION Masato Hagiwara Earth Species Project

2025-04-30 0 0 278.63KB 5 页 10玖币

侵权投诉

AVES: ANIMAL VOCALIZATION ENCODER BASED ON SELF-SUPERVISION

Masato Hagiwara

Earth Species Project

ABSTRACT

The lack of annotated training data in bioacoustics hinders

the use of large-scale neural network models trained in a

supervised way. In order to leverage a large amount of unan-

notated audio data, we propose AVES (Animal Vocaliza-

tion Encoder based on Self-Supervision), a self-supervised,

transformer-based audio representation model for encoding

animal vocalizations. We pretrain AVES on a diverse set

of unannotated audio datasets and ﬁne-tune them for down-

stream bioacoustics tasks. Comprehensive experiments with

a suite of classiﬁcation and detection tasks have shown that

AVES outperforms all the strong baselines and even the

supervised “topline” models trained on annotated audio clas-

siﬁcation datasets. The results also suggest that curating a

small training subset related to downstream tasks is an efﬁ-

cient way to train high-quality audio representation models.

We open-source our models1.

Index Terms—bioacoustics, self-supervision

1. INTRODUCTION

The use of machine learning, especially deep neural network

models, has become increasingly popular in bioacoustics in

recent years, driven by the increased amount of bioacoustic

data recorded by affordable recording devices [1].

However, when trained in a supervised way, deep neu-

ral networks typically require a large amount of annotated

data, and collecting high-quality annotation in bioacoustics

requires deep expertise and is often costly in terms of the time

required for manual labeling. This lack of labelled data in

bioacoustics hinders the use of large-scale models and most

recent models still rely on species-speciﬁc customized CNNs

(convolutional neural networks) trained on a small amount of

task-speciﬁc annotated data [2]. Although a few studies [3, 4]

pretrained from large datasets such as AudioSet [5], this re-

quires a large annotated dataset of general sound.

One solution to this issue is self-supervision, a type of

machine learning technique where the training signals come

from the data itself in the form of pseudo labels. In recent

years, large-scale self-supervised models and transfer learn-

ing have been hugely successful in related domains such

as NLP [6, 7], computer vision [8, 9], and human speech

1https://github.com/earthspecies/aves

processing [10, 11]. There has been prior work on train-

ing general-domain audio representation models [12, 13],

although those models are not designed speciﬁcally for bioa-

coustics and contain few datasets of animal sound. These

ﬁndings lead us to expect that similar self-supervised ap-

proaches are also effective for bioacoustic tasks.

In this paper, we propose AVES (Animal Vocalization

Encoder based on Self-supervision2), a self-supervised,

transformer-based audio representation model for encod-

ing animal vocalizations for downstream bioacoustic tasks.

For modeling raw waveforms with self-supervision, we use

HuBERT (Hidden Unit BERT) [14], an audio representation

model originally proposed for human speech. HuBERT ob-

tained one of the top results in the SUPERB benchmark [15].

We pretrain AVES on unannotated audio data that include

not only animal vocalizations, but also other sounds (e.g.,

human speech and environmental sound). We emphasize

that the pretraining is purely self-supervised and does not

rely on any annotation, besides choosing which instances to

include. We evaluate AVES on a suite of classiﬁcation and

detection tasks, two of the most common tasks considered

in the bioacoustic literature [2]. Our experiments show that

AVES outperforms all the baselines (including pretrained

ResNet and VGGish [16] models pretrained on millions of

videos), even the supervised topline models trained on anno-

tated audio classiﬁcation datasets.

Our contributions are as follows: with extensive bench-

marking, we show that self-supervised methods originally

proposed for human speech processing can be successfully

applied to the domain of bioacoustics. This paper is also one

of the few ﬁrst studies [2] showing that, given appropriate

pretraining, transformers work well for data-scarce domains

such as bioacoustics, where CNNs are still by far the most

common approach. We open source our pretrained models.

2. METHOD

AVES heavily relies on HuBERT [14], a self-supervised audio

representation model originally proposed for human speech.

2.1. HuBERT pretraining

Unlike human written language where text can be used as

self-supervision signals, human speech and bioacoustics data

2Pronounced /ei vi:z/. “Aves” is the class name for birds, although our

method supports diverse animal species, not just birds.

arXiv:2210.14493v1 [cs.SD] 26 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：

AVES:ANIMALVOCALIZATIONENCODERBASEDONSELF-SUPERVISIONMasatoHagiwaraEarthSpeciesProjectABSTRACTThelackofannotatedtrainingdatainbioacousticshinderstheuseoflarge-scaleneuralnetworkmodelstrainedinasupervisedway.Inordertoleveragealargeamountofunan-notatedaudiodata,weproposeAVES(AnimalVocaliza-tionEncoder...

展开>> 收起<<

A VES ANIMAL VOCALIZATION ENCODER BASED ON SELF-SUPERVISION Masato Hagiwara Earth Species Project.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！

A VES ANIMAL VOCALIZATION ENCODER BASED ON SELF-SUPERVISION Masato Hagiwara Earth Species Project

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: