Aligning Stuttered-Speech Research with End-User Needs: Scoping Review, Survey, and Guidelines

1. Problem Statement

Despite increased research interest in atypical speech, current speech technology (specifically Automatic Speech Recognition - ASR, and stutter detection) suffers from a critical misalignment with the needs of its primary end-users: People Who Stutter (PWS) and Speech-Language Pathologists (SLPs).

Key issues identified include:

• Lack of Interdisciplinary Dialogue: Research is often driven by technical benchmarks rather than clinical or user-centric use cases.
• Task Ambiguity: Terms like "stuttering" and "disfluency" are used interchangeably, and critical distinctions between tasks (e.g., detection vs. classification, verbatim vs. intended transcription) are often obscured in literature.
• Evaluation Gaps: Standard evaluation metrics (e.g., Word Error Rate) often fail to capture the specific needs of stakeholders (e.g., SLPs need verbatim transcripts for diagnosis; PWS need intended transcripts for communication).
• Limited Collaboration: Fewer than 20% of studies report meaningful stakeholder involvement in problem definition, data collection, or evaluation.

2. Methodology

The authors employed a mixed-methods approach combining a scoping review of the literature with a quantitative and qualitative survey of stakeholders.

A. Scoping Review (Literature Mapping)

• Corpus: 228 papers published between 2010 and October 2025, identified via Semantic Scholar keyword searches.
• Annotation Protocol: Papers were manually annotated across four dimensions:
  1. Research Area: Mapped to a proposed taxonomy (see Contributions).
  2. Language Coverage: Analysis of supported languages.
  3. Stakeholder Involvement: Tracking collaboration with PWS/SLPs.
  4. Open Science: Availability of code, data, and models.
• Taxonomy Development: The authors developed a standardized task taxonomy to disentangle confounded research areas (e.g., separating "Stutter Identification" from "Speech Recognition").

B. Stakeholder Survey

• Participants: 70 respondents total:
  • 40 PWS: Diverse, multilingual, mostly aged 18–44.
  • 30 SLPs: Highly experienced (many with >10 years of practice), working in multilingual contexts.
• Instrument: Two complementary questionnaires (English/Spanish) covering demographics, stuttering profiles, current tool usage, perceived needs, and ethical concerns.
• Analysis: Quantitative data (Likert scales) was analyzed using descriptive statistics; qualitative open-text responses were thematically coded.

3. Key Contributions

A. Proposed Research Taxonomy

The paper introduces a standardized taxonomy to clarify the stuttered-speech research landscape (Table 1 in the paper):

• Speech Recognition: Distinguishes between Intended Speech Recognition (removing disfluencies for communication) and Verbatim Speech Recognition (preserving disfluencies for clinical analysis).
• Stutter Identification: Distinguishes between:
  • Classification: Assigning labels to pre-segmented clips (e.g., "stutter" vs. "no stutter").
  • Detection: Locating the temporal bounds of events (time-stamping).
  • Severity Assessment: Assigning a severity score (regression/classification).
• Other Categories: Translation, Speech Correction, Therapy Tools, Data-Centric approaches, and Research-Focused (ethics, explainability).

B. Alignment Analysis

The study explicitly maps research outputs against stakeholder needs, revealing systematic gaps (Table 2 in the paper).

C. Actionable Guidelines

The paper provides concrete recommendations for future research directions, evaluation practices, and interdisciplinary protocols to achieve human-centered design.

4. Key Results & Findings

A. Literature Landscape Findings

• Dominance of Classification: Research is heavily skewed toward Stutter Identification (170/228 papers), specifically classification on pre-segmented audio. True temporal detection is rare (only ~6 papers perform actual detection despite 72 titles claiming "detection").
• Implicit Objectives: Most ASR research implicitly optimizes for Intended Speech Recognition (cleaning the transcript) without explicitly stating this, despite SLPs requiring Verbatim transcripts for clinical work.
• Language Bias: The field is English-centric (152 papers). Non-English languages (German, Mandarin, etc.) are underrepresented, and cross-lingual transfer is minimal.
• Lack of Openness: Only ~10% of papers release code, models, or data. Stakeholder collaboration is the exception (<20%), mostly limited to data curation rather than problem definition.

B. Stakeholder Insights

• Divergent Transcription Needs:
  • PWS: Prefer Intended Transcripts (removing repetitions/blocks) to facilitate communication and reduce anxiety. They report "Impatient ASR" (systems stopping listening during blocks) as a major barrier.
  • SLPs: Prefer Verbatim Transcripts to accurately assess severity, track patterns, and document therapy.
• Preference for Detection over Classification: Both groups strongly prefer tools that indicate when and where stuttering occurs (temporal detection) rather than just whether it occurs (binary classification). They need pattern tracking across contexts and interlocutors.
• The "Readiness-Willingness" Gap: While 63% of SLPs are open to AI, only 20% feel the field is ready. Adoption is conditional on explainability, privacy, and accuracy. SLPs demand to know why a model flags a segment.
• Data Misalignment: Researchers are increasingly relying on synthetic data to simulate stuttering, whereas PWS expressed a high willingness to donate real speech data. Synthetic data fails to capture the nuance and variability of lived stuttering.

C. Identified Gaps (The "Misalignment")

1. Stutter Identification: Literature offers static labels; stakeholders need temporal, pattern-aware detection.
2. Speech Recognition: Literature is ambiguous about intended vs. verbatim goals; stakeholders have distinct, opposing needs for these two modes.
3. Research-Focused: Explainability, ethics, and deployment guidance are severely under-researched compared to model building.
4. Data-Centric: A reliance on synthetic data despite the availability of willing real-world donors.

5. Significance and Recommendations

This paper serves as a critical corrective to the current trajectory of stuttered-speech research, shifting the focus from purely technical optimization to human-centered utility.

Core Recommendations:

1. Diversify and Standardize: Move beyond classification to detection and severity tracking. Adopt the proposed taxonomy to ensure clear communication.
2. Explicate Objectives: Researchers must explicitly state whether their ASR system targets intended or verbatim transcription.
3. Prioritize Explainability: Develop models that provide segment-level explanations to build trust with clinicians.
4. Multilingual Inclusion: Treat multilingual support as a core requirement, not an afterthought.
5. Open Science: Release data, code, and models to enable benchmarking and adaptation.
6. Deep Collaboration: Involve PWS and SLPs from the problem definition stage through to deployment, not just for data collection.

Conclusion:
The paper argues that without systematically incorporating end-user perspectives, stuttered-speech technology will fail to achieve clinical relevance or ecological validity. By aligning research priorities with the specific, often divergent, needs of PWS and SLPs, the field can move toward creating truly useful, trustworthy, and equitable tools.