Introduction

Speech technology is increasingly getting embedded in everyday life, with its applications spanning from critical domains like medicine, psychiatry, and education, to more commercial settings. This rapid growth has often largely contributed to the successful use of deep learning in modelling large amounts of speech data. However, the performance of speech technology related applications also largely varies depending on the demographics of the population the technology has been trained on and is applied to. That is, inequity in speech technology appears across age, gender, people with vocal disorders or from atypical populations, people with non-native accents.

This interdisciplinary session will bring together researchers working on child speech from speech science and speech technology. In line with the theme of Interspeech 2024, Speech and Beyond, the proposed session will address the limitations and advances of speech technology and speech science, focusing on child speech. Furthermore, the session will aid in the mutual development of speech technology and speech science for child speech, while benefiting and bringing both the communities together.

Topics

A series of oral presentations (or posters) on the following topics of interest (but not limited to):

• Using speech science (knowledge from children’s speech acquisition, production, perception, and generally natural language understanding) to develop and improve speech technology applications.
• Using techniques used for developing speech technology to learn more about child speech production, perception and processing.
• Computational modelling of child speech.
• Speech technology applications for children including (but not limited to), speech recognition, voice-conversion, language identification, segmentation, diarization etc.
• Use and/or modification of data creation techniques, feature extraction schemes, tools and training architectures developed for adult speech for developing child speech applications.
• Speech technology for children from typical and non-typical groups (atypical, non-native speech, slow-learners, etc.)

Website (if available)

https://sites.google.com/view/sciencetech4childspeech-is24/home

Organizers (in alphabetical order)

• Nina R. Benway (University of Maryland A. James Clark School of Engineering, USA)
• Odette Scharenborg (Delft University of Technology, the Netherlands)
• Sneha Das (Technical University of Denmark, Denmark)
• Tanvina Patel (Delft University of Technology, the Netherlands)
• Zhengjun Yue (Delft University of Technology, the Netherlands)

Introduction

Significant strides in performance have been made by large-scale automatic speech recognition (ASR) models, resulting in low error rates for high-resource languages. This success is attributed to their training in a semi/unsupervised fashion on datasets that predominantly feature a concentrated set of languages. With this session, we aim to encourage researchers to propose approaches which can improve the performance of under-utilized languages with large-scale pretrained ASR models. While high-resource language performance can be continuously improved with scaling datasets and models, the same approach cannot be readily applied to under-represented languages, typically characterized by their low-resource nature. Therefore, a greater emphasis on technical advancements is necessary to achieve low error rates for under-represented languages. Building upon the success of previous Interspeech benchmarks like SUPERB and ML-SUPERB, this session extends the exploration to tackle the challenges linked with languages lacking prominence in the training data of pretrained models. Notably, there exists a substantial variance in performance across languages in multilingual benchmarks such as Fleurs. Thus, with this special session, we provide a platform for focused study of analysis and improvements for ASR for languages which are under-utilised in large space speech pretrained models. Additionally, we extend an invitation to researchers to contribute new datasets in under-represented languages, fostering collaboration within the research community and enabling substantial advancements in this domain.

Topics

Different fine-tuning, adaptation approaches on Self-supervised learning (SSL) models for ASR such as –

• Adapter / prompt tuning
• re-scoring approaches
• data augmentation
• zero-shot/few-shot transfer
• test-time adaptation
• pseudo-label training
• ensembling

Analysis and datasets for under-utilised languages with SSL models

• Detailed analysis of performance with pretrained models with under-represented datasets
• Open sourcing new low-resource datasets and studying the performance of large pre-trained models

Website:

https://sites.google.com/view/is24-ssl-ul

Organizers:

• Brian Kingsbury (IBM Research, USA)
• Thomas Hain (The University of Sheffield, UK)
• Hung-yi Lee (National Taiwan University, Taiwan)
• David Harwath (The University of Texas at Austin, USA)
• Peter Bell (University of Edinburgh, UK)
• Jayadev Billa (Information Sciences Institute, University of Southern California, USA)
• Chenai Chair (Mozilla Foundation, South Africa)
• Prasanta Kumar Ghosh (Indian Institute of Science, Bangalore, India)

Introduction

Recent technological advancements, like foundational speech and large language models, are leading to the rapid integration of speech and language technologies into healthcare. This presents many exciting opportunities for personalized care, early disease detection, and improved communication between patients and providers. However, this integration also poses unique challenges regarding data quality, clinical translation, and responsible application of language models.

This special session focuses on two key themes in speech-healthcare research:

1. From Collection and Analysis to Clinical Translation

This theme explores how various factors impact the analysis, interpretation, and ultimately, the clinical application of speech data and includes, for example:

• Technical and human challenges in speech data collection
• Understanding how health changes affect speech and language mechanisms.
• Investigation of novel methods for capturing, analyzing, and quantifying speech patterns for health assessment
• Novel approaches to utilizing speech technologies for disease detection, monitoring, and intervention
• Considering practical feasibility to ensure findings reliably translate into clinical practice.

2. Speech and Language Technology for Medical Conversations

This theme includes the growing field of “ambient intelligence,” where technologies like speech recognition and natural language processing (NLP) are used, for example, to:

• Automatically transcribe and analyze doctor-patient conversations.
• Generate accurate medical documentation from these conversations.
• Provide feedback to medical students on their communication skills.
• Offer diagnostic support based on spontaneous speech with physicians.
• Automatically detect health disorders from conversation/spontaneous speech.
• Develop novel applications for speech and language technology in healthcare.

By bringing together these themes, this session fosters discussion on innovative ideas, challenges, and opportunities for utilizing speech technologies to advance healthcare.

Website:

https://sites.google.com/view/splang-health-interspeech2024/

Organizers:

• Nicholas Cummins (nick.cummins@kcl.ac.uk) – King’s College London
• Thomas Schaaf (tschaaf@solventum.com) – 3M | Solventum
• Visar Berisha – Arizona State University
• Sneha Das – Technical University of Denmark
• Judith Dineley – King’s College London
• Matt Gormley – Carnegie Mellon University
• Sandeep Konam – HITLOOP
• Daniel Low – Harvard University
• Bahman Mirheidari – University of Sheffield
• Emily Mower Provost – University of Michigan
• Paula Andrea Pérez-Toro – Friedrich-Alexander-Universität Erlangen-Nuremberg
• Thomas Quatieri – MIT Lincoln Laboratory
• Vikram Ramanarayanan – Modality AI
• Chaitanya Shivade – Amazon.com
• Helmer Strik – Radboud University
• Tanya Talkar – Aural Analytics

Introduction

The main topic of this special session is speech-related biosignals, such as of articulatory or neurological activities during speech production or perception. Since these biosignals reflect human speech processes, they can serve as alternative modalities to the acoustic signal for speech-driven systems. Therefore, biosignal-enabled speech systems have the potential to enable spoken communication when the acoustic speech signal is not available or perceivable. For instance, Silent Speech Interfaces are developed to restore the ability of spoken communication for speech-impaired persons, e.g., after a laryngectomy. The aim is to enable spoken communication by predicting the acoustic speech signal from biosignals such as Electromyography (EMG), Electromagnetic Articulography (EMA), or Ultrasound Tongue Imaging (UTI). Likewise, biosignals related to speech perception, such as Electroencephalography (EEG), are investigated for neuro-steered hearing aids to detect selective auditory attention to single out and enhance the attended speech stream. Progress in the field of speech-related biosignal processing will lead to the design of novel biosignal-enabled speech communication devices and speech rehabilitations for everyday situations.

With the special session “Biosignal-enabled Spoken Communication”, we aim to bring together researchers working on biosignals and speech processing in order to exchange ideas on interdisciplinary topics.

Topics

Topics of interest for this special session include, but are not limited to:

• Processing of biosignals capturing respiratory, laryngeal, or articulatory activity during speech, e.g., Electromagnetic Articulography (EMA), Electromyography (EMG), High Speed Nasopharyngoscopy (HSN), Ultrasound Tongue Imaging (UTI), videos of lip movements, etc.
• Speech-related processing of biosignals reflecting brain activity, such as Electroencephalography (EEG), Electrocorticography (ECoG), or functional Magnetic Resonance Imaging (fMRI).
• Application of biosignals for speech processing tasks, e.g., speech recognition, synthesis, enhancement, voice conversion, or auditory attention detection.
• Utilization of biosignals to increase the explainability or performance of acoustic speech processing methods.
• Development of novel machine learning algorithms, feature representations, model architectures, as well as training and evaluation strategies for improved performance or to address common challenges.
• Application of methods successful in acoustic speech processing to biosignals, e.g., self-supervised learning, knowledge distillation, end-to-end training, etc.
• Health-focused applications of biosignal processing, such as speech restoration, training and therapy, or (mental) health assessments.
• Other applications, such as speech-related brain-computer interfaces (BCIs), speech communication in noisy environments, or acoustic-free speech communication for preserving privacy.

Website:

https://www.uni-bremen.de/en/csl/interspeech-2024-biosignals

Organizers:

• Kevin Scheck (University of Bremen, Germany)
• Peter Wu (UC Berkeley, USA)
• Siqi Cai (National University of Singapore, Singapore)
• Yashish Siriwardena (University of Maryland College Park, USA)
• Tanja Schultz (University of Bremen, Germany)
• Gopala Anumanchipalli (UC Berkeley, USA)
• Carol Espy-Wilson (University of Maryland College Park, USA)
• Satoshi Nakamura (Nara Institute of Science and Technology, Japan)
• Prasanta Kumar Ghosh (Indian Institute of Science (IISc), Bangalore, India)
• Alan W Black (Carnegie Mellon University, USA)
• Haizhou Li (National University of Singapore, Singapore)

Introduction:

Large language models (LLMs) have achieved remarkable success in natural language processing through in-context learning. Recent studies have applied LLMs to other modalities including speech, leading to an emerging research topic for speech processing, i.e., spoken language models (SLMs). SLMs simplify the modeling of speech, making it easier to scale up to more data, languages, and tasks. A single model can often perform multiple speech processing tasks such as speech recognition, speech translation, speech synthesis, and natural dialogue modeling. By using pre-trained LLMs, certain types of SLMs exhibit strong instruction-following capabilities. This presents a promising avenue for developing “universal speech foundation models”, which take natural language instructions as input and proficiently execute diverse downstream tasks. This special session will provide an opportunity for researchers working in these fields to share their knowledge with each other, which can greatly benefit the entire speech community.

Topics

This special session aims to promote and advance the study of SLMs. We anticipate the session format to be panel and poster.

We welcome submissions on various topics related to spoken language models, including but not limited to:

• Data creation
• Speech representation learning (e.g., speech tokenizers)
• Modeling architectures and algorithms
• Training strategies (e.g., supervised fine-tuning, reinforcement learning)
• Efficient adaptation of pre-trained models (e.g., adapters, low-rank adaptation)
• Model compression (e.g., pruning, distillation, quantization)
• Novel applications
• Evaluation benchmarks and analysis methods
• Fairness and bias

Website:

https://www.wavlab.org/activities/2024/interspeech2024-slm/

Organizers:

• Yifan Peng (Carnegie Mellon University)
• Siddhant Arora (Carnegie Mellon University)
• Karen Livescu (TTI-Chicago)
• Shinji Watanabe (Carnegie Mellon University)
• Hung-yi Lee (National Taiwan University)
• Yossi Adi (Hebrew University of Jerusalem)

Introduction

Speech foundation models are emerging as a universal solution to various speech tasks. Indeed, their superior performance has extended beyond ASR. For instance, Whisper has proven to be a noise-robust audio event tagger, showcasing its potential beyond its original training objectives. Despite the advancements, the limitations and risks associated with speech foundation models have not been thoroughly studied. For example, it has been found that wav2vec 2.0 exhibits biases in different paralinguistic features, emotions, and accents, while HuBERT lacks noise robustness in certain downstream tasks. Besides this, foundation models present challenges in terms of ethical concerns, including privacy, sustainability, fairness, and safety. Furthermore, risks and biases of one model may propagate in usage alongside other models, especially in a unified framework, such as Seamless.

Thus, it is necessary to investigate speech foundation models for de-biasing (e.g., consistent accuracy for different languages, genders, and ages), enhancing actuality (not making mistakes in critical applications), preventing malicious applications (e.g., using a TTS to attack speaker verification systems, not to use for surveillance), and addressing various other aspects.

In this special session, we specialize in responsible aspects of speech foundation models, which are not adequately covered by regular sessions. We aim to facilitate knowledge sharing from diverse speech areas and pioneer discussions on both tech and non-tech issues. Furthermore, in line with the IS 2024 “Speech and Beyond” theme, we aim to foster connections with other communities such as NLP and ML, which have long been investigating responsible and trustworthy models. Position papers from those communities with theoretical or conceptual justifications for bridging the gap between speech and NLP/ML are also welcome.

Topics

• Limitations of speech foundation models and/or their solutions
  • inability to capture certain information
  • biases and inconsistent performance for different speech types
  • risks that propagate in actual use
• Better utilization of speech foundation models
  • adaptation methods for low-resource/out-of-domain speech for fairness(e.g., parameter-efficient tuning)
  • joint training of multiple tasks for reliable and holistic speech perception
  • integration with language foundation models to address the limitations of speech (e.g., whisper + llama)
• Interpretability, generalizability, and robustness of speech foundation models
  • whether and/or why the models perform well or poorly in certain types of speech, tasks, or scenarios (e.g., noisy, prosodic, pathological, multi-talker, far-field speech)
• Foundation models for understudied speech tasks
  • prosody in context
  • dialog behaviors
  • speech-based healthcare
  • emotion in conversations
  • disfluencies and non-verbal vocalizations including filler, backchannels, laughter
• Potential risks of employing speech foundation models
  • privacy breaches through the representations
  • lack of sustainability due to computing resources
  • lack of fairness concerning gender or other speaker characteristics
• Strategies for integrating tech and non-tech elements to ensure model responsibility
  • theoretical or conceptual justifications are welcome

Website:

https://sites.google.com/view/responsiblespeech

Organizers:

• Yuanchao Li (University of Edinburgh)
• Jennifer Williams (University of Southampton)
• Tiantian Feng (USC)
• Vikramjit Mitra (Apple AI/ML)
• Yuan Gong (MIT)
• Bowen Shi (Meta AI)
• Catherine Lai (University of Edinburgh)
• Peter Bell (University of Edinburgh)

Introduction

The application of large language models has significantly advanced artificial intelligence, including developments in language and speech technologies. The purpose of this special session is to promote exploring how to leverage large language models and contextual features for phonetic analysis and speech science. It will bring together experts from diverse fields to discuss the opportunities and challenges arising from the integration of large language models in phonetic research, and draw more attention to AI for science within the phonetics community.

Topics

1. Utilizing large language models in the development of tools and methods for speech science.
2. Employing contextual representations for phonetic analysis.
3. Discovering phonetic and linguistic properties embedded in pretrained and finetuned large language models.
4. Investigating the interpretability and explainability of large language models from the perspective of human speech processes.
5. Establishing connections between contextual representations and traditional phonetic features.
6. Revisiting classic problems in phonetics and phonology through the re-analysis of speech data using large language models and contextual representations.
7. Addressing the challenge of data sparsity in the application of large language models.
8. Leveraging existing large language models to improve understanding of understudied and endangered languages.

Organizers:

• Mark Liberman (University of Pennsylvania)
• Mark Tiede (Haskins Laboratories and Yale University)
• Jianwu Dang (Chinese Academy of Sciences)
• Tan Lee (Chinese University of Hong Kong)
• Jiahong Yuan (University of Science and Technology of China)

Introduction

This first-ever special session on speech and gender at Interspeech brings together scholars from across all of the speech disciplines. In addition to our technical programme, an interdisciplinary discussion panel can provide practical advice for understanding and talking about gender in speech research and dealing with it in practical applications.

Gender is an extremely important speaker characteristic in speech science and speech technology. It has been shown to affect speech production and interaction in meaningful and complex ways. Also essential is the compilation and use of inclusive and representative speech datasets to train and test speech technologies. Gender therefore relates to many popular research areas of the Interspeech community including dataset development, speech production and perception, algorithmic bias, privacy, and applications to healthcare.

Topics

This special session highlights the importance of gender as a speaker characteristic and provides a unique opportunity to address theoretical and practical questions relating to sex and gender in speech technology and speech science. Submissions can be full research papers or position papers.

Topics include:

• Empirical investigations of algorithmic bias
• Developing and using inclusive datasets
• Gender, sexuality and other types of diversity in language and speech production and processing
• Ensuring privacy and diversity for high-stakes speech applications
• Applications for gender representation in media
• Applications related to inclusion and healthcare (e.g., voice and speech therapy) related to gender and sexuality

Organizers:

• Nina Markl (University of Essex, United Kingdom)
• Cliodhna Hughes (University of Sheffield United Kingdom)
• Odette Scharenborg (Delft University of Technology The Netherlands)

Deadline: 2 March 2024

Submissions: https://interspeech2024.org/paper-submission/

Introduction

The main goal of the special session is to bring together researchers who use computational
modeling as a means to study human language acquisition, perception, or production.
Computational modeling is a research method that complements empirical and theoretical
research on the aforementioned topics by creating algorithmic implementations of theories
on human language processing, thereby enabling testing of the feasibility and scalability of
these theories in practice. Models can also act as learnability proofs by demonstrating how
language behaviors are achievable with a specific set of cognitive processes (“innate
mechanisms and biases”), input/output data (“the environment model”), and potential
physiological constraints.
Due to the interdisciplinary nature of spoken language as a phenomenon, computational
modeling research is currently dispersed across different conferences and workshops in areas
such as speech technology, natural language processing, phonetics, psycholinguistics,
cognitive science, and neuroscience. Our primary aim is to bring together people working in
these different areas but who use computational modeling as a part of their work.
Recent advances in machine learning and statistical modeling have resulted in increasingly
powerful computational models capable of explaining several aspects of human language
behaviors. This means that the methodologies used in computational modeling and
contemporary speech technologies are also becoming increasingly similar, such as using selfsupervised
learning to learn useful speech representations. At the same time, computational
models of human speech processing benefit from technical expertise in speech processing
and theoretical understanding of speech as a phenomenon—traits that are characteristic of
the Interspeech community. Therefore, another goal of this special session is also to
encourage and facilitate cross-disciplinary collaboration between technology and science –
oriented researchers in the study of human language.
Topics
We will accept submissions from any of the three areas and papers bridging them: language
acquisition, speech perception, and speech production. The key requirement is that the
submitted papers should use computational models to study human language processing,
provide novel theoretical or conceptual insights to computational research, or describe novel
technical tools, datasets, or other resources that support computational modeling research.
The emphasis will be on speech behaviors and studies involving speech audio, articulatory
data, and other biosignals.
Special session at Interspeech 2024, 1–5 September, 2024, Kos Island, Greece:
Potential topics include both L1 and L2 processing, such as (but not limited to):
● Models of sub-word, lexical, and/or syntactic perceptual learning
● Models of articulatory learning
● Models of spoken word recognition or other aspects of speech perception
● Models of speech planning and motor control
● Multimodal models (e.g., visually-grounded speech learning, audiovisual perception)
● Articulatory models
● Models bridging language acquisition and adult speech behaviors
● Models of bilingual acquisition, comprehension and production
● Models of dyadic or multi-agent interaction
● Embodied models
● Self-supervised learning of speech representations
● Bayesian models of language acquisition and speech communication
● Data, evaluation practices, or other tools and resources for computational modeling
of human speech behaviors.
We also encourage machine learning and speech technology researchers to submit their
technical work on unsupervised or self-supervised acquisition of speech units (e.g., phones,
syllables, words, syntax) and analysis of language representations in deep learning models, if
these models can explain acquisition of language structures without supervised learning.

Organizers:

• Okko Räsänen (Tampere University, Tampere, Finland)
• Thomas Hueber (GIPSA-lab, Grenoble, France)
• Marvin Lavechin (GIPSA-lab, Grenoble, France)