In collaboration with researchers and scholars from Carnegie Mellon University, New Jersey Institute of Technology, Gallaudet University, Apple, Cornell University, University of Illinois Urbana-Champaign, University of California, Santa Cruz, and University of Maryland, College Park, we led a CHI 2025 workshop on “Speech AI for All: Promoting Accessibility, Fairness, Inclusivity, and Equity” in April 2025.

Prelude

As the first (hopefully, of many) CHI workshop focusing on the intersection of AI and speech diversity, the idea of this work emerged from AImpower.org’s internal discussion post CHI ’24 and formalized during the Human-Centered Machine Learning (HCML) workshop hosted by Apple in August 2024. With Kimi Wenzel, Colin Lea, and Raja Kushalnagar at the HCML workshop, we envisioned the scope and goals of our CHI workshop to encompass the technical, socio-structural, and the experiential dimensions of speech AI technology, especially in relation to its fairness and accessibility issues.

Workshop Goals

We have set out three core goals for the workshop:

1. Summarize the latest research and understanding around AI and speech diversity;
2. Create a roadmap for fair and accessible speech AI;
3. Build a multidisciplinary and intersectional community to carry out the roadmap collaboratively.

Since the beginning, we also committed to the following program goals:

• Center the lived experiences and expectations of those most affected by fairness and accessibility issues in speech AI;
• Highlight scholarship and domain expertise beyond the traditional CHI/technical community;
• Learn from culturally and geographically underrepresented communities;
• Minimize participation barriers.

While these program goals certainly contributed to the core goals of the workshop, our commitment — and the extra efforts involved — to these program goals was rooted in the productivity alone but by in a broader vision for procedural justice and power redistribution in AI development.

Program Overview

Our workshop program was intentionally designed to support both our core and program goals, featuring voices from communities that are significantly impacted by speech AI fairness and accessibility issues, yet rarely included in these conversations.

To achieve this, we made concerted efforts to identify and invite keynote and panel speakers and implemented several strategies to reduce economic, accessibility, and language barriers for participation.

For example, anticipating the high cost of conference attendance and international travel, we we secured early sponsorship from Apple (thank you, Apple!) to offer sizable honoraria that helped offset registration and travel expenses. We also leveraged our broader academic and community networks to successfully include three community panelists from Japan, where the conference was held.

Understanding that Japanese panelists may feel discomfort with English as the dominant language, we conducted pre-workshop communication in both English and Japanese (thanks to LLMs!) and hired a professional Japanese–English interpreter to support the community panel.

To ensure optimal accessibility accommodations, Raja, Christian, and I began working with the CHI Accessibility Chairs in December 2024 to secure sign language interpretation and CART services. Our hybrid experience leads, Christian Vogler and Colin Lea, invested tremendous effort into designing an experience that was accessible for both in-person and remote participants, including audio/video recordings of key sessions.

Time	Format	Activity
9:00 -9:15	In-person/ hybrid	Opening Remarks (by Shaomei Wu) Introduce the workshop, outline its goals, and engage participants in ice-breaker activities.
9:15 -10:30	In-person/ hybrid (recorded)	Expert Micro-Keynote Provocations (moderated by Allison Koenecke) Three 20-minute keynotes from experts beyond the HCI community, followed by a 15-minute joint Q&A session.
10:30 -11:10	–	Break
11:10 -12:40	In-person	Poster Presentations (moderated by Jingjin Li) Participants present accepted papers in a poster format, with 3-min introductory presentation each at the beginning of the session.
12:40 – 14:10	–	Lunch Break
14:10 – 15:40	In-person (recorded)	Community Panel (moderated by Kimi Wenzel) First-person testimony and direct engagement with communities impacted by speech AI.
15:40 – 16:20	–	Break
16:20 – 17:20	In-person	Small Group Themed Discussion (moderated by Norman Su) Participants form small groups around interested topics and engage in in-depth, structured discussions.
17:20 – 17:40	In-person (recorded)	Group Sharing & Wrap-Up (moderated by Norman Su) Different groups share the summary and highlights of their discussions with the rest of the workshop.
17:40 – 17:50	In-person	Closing Remarks (by Raja Kushalnagar)
19:00 – 22:00	In-person	Post-workshop Social in at WATAMI Yokohamaekimae (ミライザカ JR横浜西口駅前店)

Speech AI for All Workshop Program In a Glance

We are proud that these efforts were noticed and contributed to a stimulating, inclusive experience for our participants. As one participant shared in the post-workshop survey:

“One thing I really enjoyed about the workshop overall was all the measures taken for accessibility – I have never been to such an accessible event and having the opportunity to collaborate with communities that I wouldn’t otherwise be able to communicate with was truly eye opening.”

I will also share some highlights from each activity of the workshop below.

Micro-keynotes

Recognizing that challenges in speech AI are not solely technical but deeply socio-cultural, we invited experts from fields such as cultural anthropology, disability studies, public policy, and linguistics. These speakers shared critical accounts of speech AI and grounded the follow-up discussion in holistic, socio-technical approaches. Our speakers include:

• Karen Nakamura: In their keynote speech titled “Open the Pod Bay Doors, Hal: I’m sorry. I don’t understand you.” Karen highlighted the regression in modern AI speech models to recognize non-standard speech and the adverse impact on the lives — and in some cases, denying the personhood — of people with speech diversity. As we exploring solutions for these problems, Karen cautioned the tech community for building “disability dongles” or solutions as seperate/beta projects, that result in further marginalization and failed promises.

• Ly Xīnzhèn Zhǎngsūn Brown: Starting their talk by prompting the audience to use the room as they need or prefer, Ly Xīnzhèn gave an overview of ableism and ableist harms within/by AI, and how current laws/policies act (and could act) as a tool of ableist harm. Ly Xīnzhèn ended their talk with a call for disability justice and collective liberation.

• Nicole Holliday: In her keynote titled “I Don’t Like My Conversations Being Judged By an AI”: Bias and Quality in Socially Prescriptive Speech Technologies”, Nicole discussed the use case of speech AI in Socially Prescriptive Speech Technologies (SPST), especially, the integration of SPST in videoconferencing softwares that both reflects and reinforce existing social biases of speakers from different race, gender, and L1 groups. To resist SPST harms, Nicole called for more public education on how human speech works, deeper and broader ethics consideration of SPST in linguistics and industry, increased consumer awareness (and resistance) of such technologies, as well as policy and legal reform to regulate the public deployment of SPST – especially in the employment context.

All of the keynotes were very well received by the participants, with an average rating of 4.4/5 in our post-workshop survey. One participant noted:

“I really enjoyed the diversity of the keynotes coming from people of different schools of thought – I think that diversity in perspectives was really nice to set the stage for the remainder of the workshop.”

Watch the full keynote session below.

Poster Presentation

We have received 22 submissions and accepted 9 for poster presentation. The accepted papers spanned a wide range of disciplines and community concerns:

1. How People Living With Amyotrophic Lateral Sclerosis Use Personalized Automatic Speech Recognition Technology to Support Communication, presented by Richard Cave (University College London)

2. Listening to Bias: Tracing Lineage, Impacts, and Paths Forward, presented by Johann Diedrick (New York University / Social Science Research Council, Just Tech Fellow)

3. Bridging the Speech AI Accessibility Gap for Deaf and Hard of Hearing People, presented by Abraham Glasser (Gallaudet University)

4. Towards Temporally Explainable Dysarthric Speech Clarity Assessment, presented by Chitralekha Gupta (National University of Singapore)

5. Inclusivity of AI Speech in Healthcare: A Decade Look Back, presented by Retno Larasati (The Open University, United Kingdom)

6. Automatic Speech Recognition Model Adaptation for Individuals with Autism Spectrum Disorder, presented by Bowon Lee (Inha University)

7. Addressing Non-Pathological Speech Disfluency in AI, presented Ralph Rose, Ayaka Sugawara, Aina Tanaka (Waseda University)

8. On the Lack of Queer Voices in Diverse Speech Datasets, presented by Brooklyn Sheppard (University of Calgary)

9. Enhancing Mental Health through DSR: Psychological Behavioral Evidence and Future Application Prospects, presented by Su-Jing Wang (CAS Key Laboratory of Behavioral Science, Institute of Psychology)

The poster session was highly interactive, allowing participants to network broadly while engaging in deep 1:1 conversations. Participants rated it an average of 4.8/5 for value. As one participant noted:

I very much enjoyed the poster presentations and themed discussions as they allowed a lot of interaction, discussion and getting to know the other workshop participants.

Community Panel

The community panel is one of my favorite parts of the day. Designed to share personal testimony and contest of the inclusive and accessibility issues of speech AI, we curated an all-star panel with local and international speakers from communities directly impacted by those issues. From left to rights, our panelists are:

• Ken-ichiro YABU (he/his): project researcher with RCAST (Research Center for Advanced Science and Technology) and IOG (Institute of Gerontology) in the University of Tokyo. He has muscle weakness in his whole body and uses an electric wheelchair. He has moderate speech articulation disorder.
• Yukinori Tamaki (he/his): social worker and the representative director of Hyogo Prefecture Consultation Support Network, General Incorporated Association. He has cerebral palsy and appeared on NHK Educational TV’s Baribara for 16 years.
• Yang Yang (she/her): Certificated Speech Therapist based in Tokyo, Japan. She was born and raised in Shanghai, China mainland. She has a developmental stutter and is one of the founding members of “Slow Order Cafe”, a project that runs a series of pop-up events staffed entirely by young people who stutter.
• Robin Netzorg (she/her): is a 6th year PhD student at the University of California at Berkeley. Her research focuses on interpretable models of speaker identity that quantifies expert perception of voice quality, with an aim to apply these methods to gender-affirming voice training and speech therapy as assistive tools.
• Abraham Glasser (he/his) is a born-Deaf, native American Sign Language (ASL) signer who occasionally uses his voice that has a Deaf accent. He is an Assistant Professor in the Accessible Human Centered Computing program at Gallaudet University. His research investigates technologies, AI, and accessibility for people who are Deaf or Hard of Hearing, including speech, sign language, and immersive technologies.

Moderated by Kimi Wenzel and translated by Robert John (our Japanese – English interpreter), the panelists shared their experiences and perspectives on the following topics:

• Lived experiences and scenarios when speech AI creates new challenges for them;
• Current work-arounds people deploy to address these challenges;
• The kinds of changes people want to see in speech AI products and developments.
• Positive experience with speech AI (for almost all panelists, the answer is “none” 😔).

Panelists expressed frustration with current systems, especially those deployed without consent in public or commercial services. Many reported avoiding speech AI altogether, preferring alternative modes of interactions.

Watch the recorded panel session below to learn more about the panel’s perspectives.

Themed Discussion

In our final activity, participants broke into small groups under three key themes:

• Inclusion and Representation: How can we centralize involvement of those who are most impacted be Speech AI? How can we meaningfully involve and include folks?
• A Roadmap for Fair Speech AI: What are guiding principles to building long-term solutions regarding Speech AI? For example, what design principles or metrics and measurements can we develop?
• Tensions in the “All” of Speech AI for All: What are the tensions (e.g., debates about synthesized speech, the role of transcripts for stuttered speech) within the “all” – communities impacted with Speech AI? How do we navigate that plurality? How can we productively and positively move forward?

Watch the session summary in the recorded video below.

Looking Ahead

While we enjoyed the intellectual and communal synergy at the workshop, we also realized that a one-day, 8-hour workshop is not sufficient to fully form a community and build a comprehensive roadmap together. We need more time and more space for unstructured conversations (small talks!), thoughtful debates that allow us to engage with the topic and with each other in a deeper, more personal way. This sentiment is shared by many of our participants, as the top feedback we received in the post-workshop survey, is to give people more time: more time to ask questions to the speakers, more time to engage in the themed discussions, more time to engage with the posters, more time to freely network, more time to enjoy lunch/coffee breaks.

Inspired by the great energy and momentum of the first Speech AI for All workshop, we will plan the future events with more unstructured time and deeper focus on specific problem areas to allow for deeper conversations and concrete discoveries.

While this year’s workshop highlights current issues and challenges in speech AI fairness and accessibility, it is not a rejection to speech AI technology but the call for inclusive development and broad assessment before deployment. By working together today, we can build speech AI systems that empowers, rather than alienate, the people they are meant to serve.