Source-linked AI summary
Towards participatory speech dataset curation: A queer case study and conceptual framework
Brooklyn Sheppard, Anaelia Ovalle, Adina Williams, Levent Sagun
TL;DR
Speech dataset creation must account for queer speakers’ concerns about AI, identity disclosure, and data use. Drawing on prior participatory efforts, the paper proposes a framework for community-led curation. It organizes engagement around community, project formulation, participation, and personal autonomy, with these dimensions informing one another.
Problem
Speech collection practices and limited queer representation leave challenges for engaging queer speakers, whose identity labels and voice data can raise sensitivity and privacy concerns.
Source: §1 · p. 1, §2.1 · p. 1, §2.2 · p. 2Method
The paper reviews participatory AI and speech-data efforts through a queer case study and develops a co-design and knowledge-sharing framework for speech-data curation.
Source: §3 · pp. 2–3, §4 · p. 3Results
The proposed framework links community definition, project formulation, modes of participation, and personal autonomy in an ongoing two-way process.
Source: §4 · p. 3, §4.5 · p. 4Takeaways & Limitations
The framework offers AI researchers and queer community members guidance for participatory speech dataset creation, while inviting refinement across use cases and speech communities.
Source: §5 · p. 4The framework remains subject to iterative refinement for varied use cases and speech communities.
Source: §5 · p. 4
Abstract
from arXiv · showhide
In this paper, we motivate the need for a participatory speech dataset creation framework through a case study of the LGBTQIA+, or queer, community - a community with documented concerns about AI and reported harms, including attempts to develop 'gaydar' technologies that purportedly identify individuals as queer. We review common speech data collection practices, why these methods may be unsuitable for engaging with queer speakers, and discuss previous efforts in participatory AI with queer community engagement, as well as participatory endeavours specific to speech data collection for other marginalized communities. From this review, we develop a conceptual framework for participatory speech data curation by, for, and with marginalized communities drawing on insights from co-design and knowledge sharing. We propose a framework comprising overlapping and two-way processes of defining a community, project formulation, modes of participation, and personal autonomy.
1. Introduction
Speech data collection must balance quality and diversity with the labor and sensitivity involved in recording and annotating speakers. The paper argues for participation from dataset creation onward and proposes a framework informed by queer speakers’ experiences.
1. Introduction
- Collecting speech that is both high-quality and diverse is difficult, requiring substantial community labor for annotation and care with identity labels that may disclose queer membership.Gender annotations can identify people who do not identify within the gender binary as members of the queer community.Source: §1 · p. 1
- Crowdsourcing efficiently creates speech datasets, but temporary engagement without meaningful long-term collaboration raises concerns about “participation washing.”Participatory AI instead brings affected communities and other stakeholders together to design AI collaboratively.Source: §1 · p. 1
- Most datasets with speaker-gender labels omit speakers outside the gender binary; those that include them contain only a small number. MAGES focuses on gender-diverse representation and specific use cases.Source: §1 · p. 1
- Participatory speech dataset curation should involve affected communities from dataset creation onward, and the paper proposes a framework for curation by, for, and with marginalized communities.The framework is motivated through the queer community as a case study.Source: §1 · p. 1
2. Challenges in speech data diversity: The case of queer speakers
Queer speakers are an important case for speech-data inclusion because identity labels, changing voices, limited representation, and documented AI harms create distinctive challenges. The paper reviews these concerns and the limits of existing evidence about speech technologies’ performance for gender-diverse speakers.
2.1. The queer community & their voice(s)
- Queer speakers’ voices may change over time or context, while gender identity annotations can disclose marginalized status in a way that makes such data sensitive.The paper defines queer as people who self-identify as having non-normative sexual identity, gender identity, and/or sexual characteristics; voice changes may relate to voice coaching, hormone therapy, or communicating within versus outside the community.Source: §2.1 · pp. 1–2
- Speech-recognition and categorization models show demographic biases; ASR performs significantly worse on female voices when female voices are underrepresented in training, while evidence beyond binary gender remains preliminary.In a three-category study, no tested model favored the “other” category; language-specific samples ranged from 8 to 86 speakers, and the authors called for more data.Source: §2.1 · p. 2
2.2. The queer community and AI
- Queer communities have faced reported AI harms, including purported “gaydar” technologies, targeting of trans healthcare facilities, and disproportionate toxicity flagging of queer dialects.Interviews with trans and non-binary voice-assistant users also report privacy concerns, discomfort with visibility to developers, and distrust of AI practices.Source: §2.2 · p. 2
- Speech datasets have often used perceived gender and binary labels, risking misgendering and erasure; longstanding distrust and limited queer representation make this a challenging case for participatory collection.Self-reported gender identity is becoming more common in speech datasets.Source: §2.2 · p. 2
3. Previous efforts in participatory AI
Prior participatory work offers examples of community consultation and co-design, but leaves questions about community influence over speech-data projects. The paper draws on these efforts and reciprocal field-linguistics practices to motivate shared project design and knowledge exchange.
3. Previous efforts in participatory AI
- PARQAIR-MH uses repeated anonymous surveys to build consensus on queer community preferences for mental-health AI, though community experts’ influence on later model design or data collection is limited.The process produces a toolkit for AI researchers and health policy institutions.Source: §3 · p. 2
- Speech communities both provide recordings and annotations and bear voice-identification risks, motivating consent, privacy protections, and transparency about represented communities.Voice data may inadvertently reveal membership in a marginalized community and increase the likelihood of harm.Source: §3 · p. 2
- StammerTalk engaged the stuttering community in a co-designed dataset that community members advocated releasing publicly, and a post-collection survey reported feelings of empowerment among speakers and researchers.The core research team formulated the project design independently of speakers, leaving a distinction between data collectors and contributors.Source: §3 · pp. 2–3
- Participatory field linguistics shifts from one-way extraction toward reciprocal knowledge sharing, involving language communities in project goals and outcome evaluation.These practices inform the paper’s proposal for participatory speech dataset creation by, for, and with queer communities.Source: §3 · p. 3
4. Conceptual framework for community-led speech data curation
The framework places community needs and lived experience at the center of speech-data curation, linking project formulation, participation, and personal autonomy. Its phases are interdependent and may require ongoing reconsideration as project choices and community needs shape one another.
4. Conceptual framework for community-led speech data curation
- The framework favors co-design where feasible, situating project goals within the community and connecting project scope to participation choices and individual control over contributions.It contrasts one-way researcher-to-output collection with nested, bidirectional collaboration and co-design.Source: §4 · p. 3, Figure 1 · p. 3
4.1. Community
- Community definition asks whom a project will serve and who may be left out, including people at intersecting identities, language groups, or dialects.The paper notes that a project centered on Queer in AI may skew toward Global North and predominantly English-speaking participants, and recommends making absent perspectives explicit.Source: §4.1 · p. 3
4.2. Project formulation
- Project formulation considers speech genres and their uses and risks, including the possibility that intimate conversational recordings may deter participation or require removal.The framework also asks whether data will be public, how it may be used, and how it will be stored and maintained, including whether contributors can revoke data and who retains access.Source: §4.2 · pp. 3–4
4.3. Modes of participation
- Modes of participation include contributing recordings, annotating speech, organizing, and navigating collaboration, communication preferences, and group power dynamics.Contributions may be acknowledged through payment or authorship [33], depending on collaborators’ preferences.Source: §4.3 · p. 4
4.4. Personal autonomy
- Personal autonomy concerns each contributor’s empowering mode of engagement, while some project decisions may require community consensus and others allow individual choice.Examples of individual choices include participation mode, acknowledgment, and the freedom to revoke data; choices can also affect the community downstream.Source: §4.4 · p. 4
4.5. A two-way process
- The phases are not unidirectional: project choices can prompt reconsideration of which communities are involved, while risk can shape sharing and security decisions.The framework calls for continuously considering community, project formulation, participation modes, and personal autonomy throughout dataset creation.Source: §4.5 · p. 4
5. Conclusion and future directions
The paper offers researchers and queer community members collaborative, two-way knowledge-sharing guidelines for participatory speech dataset creation. It presents queer speakers as a case study while proposing broader applicability across communities and use cases.
5. Conclusion and future directions
- The framework provides AI researchers and queer community members guidelines for participatory speech dataset creation through collaborative, two-way knowledge sharing.It is presented as a case study of the queer community, with potential application beyond it.Source: §5 · p. 4
- The framework could be iteratively refined across speech communities and use cases, including through reflection on its use and community feedback.Source: §5 · p. 4
- The authors hope the framework will encourage collaborative consideration of more queer perspectives in speech technology research and development.Source: §5 · p. 4