서브메뉴
검색
Learning Audio Foundation Models for Reasoning
Learning Audio Foundation Models for Reasoning
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103610
- ISBN
- 9798290940250
- DDC
- 534
- 저자명
- Deshmukh, Soham.
- 서명/저자
- Learning Audio Foundation Models for Reasoning
- 발행사항
- [Sl] : Carnegie Mellon University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 248 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-02, Section: B.
- 주기사항
- Advisor: Raj, Bhiksha.
- 학위논문주기
- Thesis (Ph.D.)--Carnegie Mellon University, 2025.
- 초록/해제
- 요약Sound understanding, a field of machine learning, has applications across healthcare, noise monitoring, bioacoustics, home security, and music. Traditionally, task-specific models were developed for each application, requiring data collection, annotation, model training, evaluation, and deployment. This iterative process, which involves designing architectures, datasets, and objectives, is repeated for each new domain, task, or application.To streamline this process, recent research has focused on pretraining large-scale datasets to learn general-purpose audio representations or embeddings. These pretrained models can then be adapted to new tasks by training linear or shallow classifiers on top of the learned embeddings. Pretrained models with fine-tuning consistently outperform task-specific models in downstream tasks, making them a cornerstone of modern sound understanding.Given the success of pretraining and scaling (both in compute and data), this thesis investigates two central questions: First, what are effective learning or pretraining methods to enable the scaling of audio models? Second, once pretrained, how can the new abilities be analyzed, evaluated, and subsequently improved? These questions are interdependent, with advancements in one area feeding into and enhancing the other. The resulting models, termed Audio Foundation Models (AFMs), are pretrained on large-scale data and develop novel capabilities for understanding and reasoning about sound. This thesis aims to devise such models capable of understanding and reasoning about audio.In the first part of this thesis, we introduce scalable learning methods to develop AFMs. Specifically, we explore audio-language pretraining to connect auditory perception with linguistic understanding. We propose three learning methods and subsequently models. First, we propose CLAP (Contrastive Audio-Language Pretraining), which learns multimodal audio-text representations through contrastive learning. CLAP excels in retrieval and indexing tasks and supports zero-shot classification, where test-time classes are user-defined. Second, we introduce audio-conditioned text generation and subsequently Pengi, an audio-language model that frames audio tasks as text-generation tasks. By combining an audio encoder with a language model, Pengi enables open-ended tasks such as audio captioning and question-answering while maintaining strong performance on close-ended tasks. Finally, we introduce Mellow, a small audio-language model designed to address Pengi's limitations in detailed reasoning and question-answering, thereby improving audio-grounded reasoning capabilities.These scalable learning methods enable training on millions of audio-text pairs, resulting in Audio Foundation Models with emergent capabilities. These models often demonstrate abilities they were not explicitly trained for, making them valuable for addressing new domains and tasks. However, discovering and benchmarking these capabilities is critical to ensuring their safe and effective use in real-world applications. In the second part of this thesis, we focus on the reasoning and analytical abilities of Audio Foundation Models. First, we demonstrate that AFMs can serve as reference-free audio quality estimators using antonym prompting strategies. Second, we define logical reasoning for AFMs and introduce an audio entailment task to measure deductive reasoning, identifying key limitations of existing models. To enhance reasoning performance, we propose the "caption-before-reason" method, which can be applied directly at test time. Finally, we explore comparative reasoning across multiple audios, introducing two new datasets with detailed explanations and proposing the ADIFF model, which improves reasoning over state-of-the-art AFMs.While pretraining on large datasets is effective, it is resource-intensive and does not fully address long-tail and out-of-distribution scenarios. To address these challenges, in the third and final part of this thesis, we propose methods to adapt AFM to new domains with minimal resources. These methods are categorized based on the required data and computational budget. First, we introduce test-time adaptation, which adapts models to new data at test time using unlabeled audio. Second, we show that few-shot learning in Pengi can improve performance with limited labeled examples. Finally, we propose a method that leverages CLAP's multimodal space to enable missing-modality training for text generation tasks.In summary, this thesis presents scalable learning methods, evaluation benchmarks, and key components for developing Audio Foundation Models capable of reasoning. These models have diverse applications, including media retrieval, creative audio generation, healthcare assistance, and predictive maintenance in industrial settings. Looking ahead, we believe this work is crucial in the inevitable rise of audio agents, enabling embodied AI systems to understand and reason about their environments using rich acoustic information.
- 일반주제명
- Acoustics
- 일반주제명
- Computer engineering
- 일반주제명
- Electrical engineering
- 키워드
- Audio processing
- 기타저자
- Carnegie Mellon University Electrical and Computer Engineering
- 기본자료저록
- Dissertations Abstracts International. 87-02B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357864
■00520260202103610
■006m o d
■007cr#unu||||||||
■020 ▼a9798290940250
■035 ▼a(MiAaPQ)AAI32043167
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a534
■1001 ▼aDeshmukh, Soham.
■24510▼aLearning Audio Foundation Models for Reasoning
■260 ▼a[Sl]▼bCarnegie Mellon University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a248 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-02, Section: B.
■500 ▼aAdvisor: Raj, Bhiksha.
■5021 ▼aThesis (Ph.D.)--Carnegie Mellon University, 2025.
■520 ▼aSound understanding, a field of machine learning, has applications across healthcare, noise monitoring, bioacoustics, home security, and music. Traditionally, task-specific models were developed for each application, requiring data collection, annotation, model training, evaluation, and deployment. This iterative process, which involves designing architectures, datasets, and objectives, is repeated for each new domain, task, or application.To streamline this process, recent research has focused on pretraining large-scale datasets to learn general-purpose audio representations or embeddings. These pretrained models can then be adapted to new tasks by training linear or shallow classifiers on top of the learned embeddings. Pretrained models with fine-tuning consistently outperform task-specific models in downstream tasks, making them a cornerstone of modern sound understanding.Given the success of pretraining and scaling (both in compute and data), this thesis investigates two central questions: First, what are effective learning or pretraining methods to enable the scaling of audio models? Second, once pretrained, how can the new abilities be analyzed, evaluated, and subsequently improved? These questions are interdependent, with advancements in one area feeding into and enhancing the other. The resulting models, termed Audio Foundation Models (AFMs), are pretrained on large-scale data and develop novel capabilities for understanding and reasoning about sound. This thesis aims to devise such models capable of understanding and reasoning about audio.In the first part of this thesis, we introduce scalable learning methods to develop AFMs. Specifically, we explore audio-language pretraining to connect auditory perception with linguistic understanding. We propose three learning methods and subsequently models. First, we propose CLAP (Contrastive Audio-Language Pretraining), which learns multimodal audio-text representations through contrastive learning. CLAP excels in retrieval and indexing tasks and supports zero-shot classification, where test-time classes are user-defined. Second, we introduce audio-conditioned text generation and subsequently Pengi, an audio-language model that frames audio tasks as text-generation tasks. By combining an audio encoder with a language model, Pengi enables open-ended tasks such as audio captioning and question-answering while maintaining strong performance on close-ended tasks. Finally, we introduce Mellow, a small audio-language model designed to address Pengi's limitations in detailed reasoning and question-answering, thereby improving audio-grounded reasoning capabilities.These scalable learning methods enable training on millions of audio-text pairs, resulting in Audio Foundation Models with emergent capabilities. These models often demonstrate abilities they were not explicitly trained for, making them valuable for addressing new domains and tasks. However, discovering and benchmarking these capabilities is critical to ensuring their safe and effective use in real-world applications. In the second part of this thesis, we focus on the reasoning and analytical abilities of Audio Foundation Models. First, we demonstrate that AFMs can serve as reference-free audio quality estimators using antonym prompting strategies. Second, we define logical reasoning for AFMs and introduce an audio entailment task to measure deductive reasoning, identifying key limitations of existing models. To enhance reasoning performance, we propose the "caption-before-reason" method, which can be applied directly at test time. Finally, we explore comparative reasoning across multiple audios, introducing two new datasets with detailed explanations and proposing the ADIFF model, which improves reasoning over state-of-the-art AFMs.While pretraining on large datasets is effective, it is resource-intensive and does not fully address long-tail and out-of-distribution scenarios. To address these challenges, in the third and final part of this thesis, we propose methods to adapt AFM to new domains with minimal resources. These methods are categorized based on the required data and computational budget. First, we introduce test-time adaptation, which adapts models to new data at test time using unlabeled audio. Second, we show that few-shot learning in Pengi can improve performance with limited labeled examples. Finally, we propose a method that leverages CLAP's multimodal space to enable missing-modality training for text generation tasks.In summary, this thesis presents scalable learning methods, evaluation benchmarks, and key components for developing Audio Foundation Models capable of reasoning. These models have diverse applications, including media retrieval, creative audio generation, healthcare assistance, and predictive maintenance in industrial settings. Looking ahead, we believe this work is crucial in the inevitable rise of audio agents, enabling embodied AI systems to understand and reason about their environments using rich acoustic information.
■590 ▼aSchool code: 0041.
■650 4▼aAcoustics
■650 4▼aComputer engineering
■650 4▼aElectrical engineering
■653 ▼aAudio foundation models
■653 ▼aAudio processing
■653 ▼aAudio-language models
■653 ▼aAudio-text learning
■653 ▼aMachine listening
■653 ▼aMultimodal learning
■690 ▼a0800
■690 ▼a0986
■690 ▼a0544
■690 ▼a0464
■71020▼aCarnegie Mellon University▼bElectrical and Computer Engineering.
■7730 ▼tDissertations Abstracts International▼g87-02B.
■790 ▼a0041
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357864▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


