FluencyBank: Annotated Stuttering Speech Corpus

• FluencyBank is a large, annotated speech corpus featuring 25 hours of stuttered and non-stuttered speech collected from a diverse group of 48 English speakers.
• It employs multi-tiered annotation protocols that capture primary dysfluencies, secondary behaviors, and tension levels, ensuring detailed temporal and qualitative insights.
• The corpus underpins both rule-based and advanced neural models for stuttering detection, segmentation, and severity assessment, enhancing clinical and technological applications.

FluencyBank is a large, systematically annotated corpus of stuttered and non-stuttered speech that serves as a critical resource for the computational modeling, clinical assessment, and automatic detection of stuttering and related dysfluencies. It is utilized widely in both clinical linguistics and speech technology research for benchmarking, development, and evaluation of stuttering detection, segmentation, transcription, and intervention systems.

1. Corpus Scope, Structure, and Annotation

FluencyBank comprises approximately 25 hours of spontaneous and elicited stuttered speech from 48 English speakers aged 10–70, with male and female speakers nearly balanced. The data includes monologues, interviews, and conversational excerpts, covering a range of linguistic contexts and severities (Xu et al., 15 Jan 2026). Sessions are transcribed in the CHAT format, which encodes event-level time-stamped annotations for disfluency types including sound repetitions (SR), word repetitions (WR), prolongations (Prol), and blocks (silent/audible) (Zhang, 21 Aug 2025). Speaker metadata (e.g., gender, ID) is included; age and demographic variables are available but not always consistently reported in downstream research (Xu et al., 15 Jan 2026).

Over multiple curation and annotation cycles, FluencyBank has evolved from single-label primary dysfluency assignment towards rich, multi-dimensional, clinically grounded annotation. The most recent standard, implemented via ELAN tiered annotation, differentiates:

• Primary dysfluencies: Syllable Repetition (SR), Incomplete Syllable Repetition (ISR), Multi-syllable Unit Repetition (MUR), Prolongation (P), Block (B) (LBDL taxonomy).
• Secondary behaviors: Verbal escape (V), Facial grimace (FG), Head movement (HM), Extremity movement (ME) (SSI-4 taxonomy).
• Tension levels: Ordinal 0–3 score reflecting Boey's scale for visible tension (Valente et al., 31 May 2025).

A consensus test set annotates 732 disfluency moments, cross-tabulated by primary type and secondary behavior. Table 1 summarizes the current annotation schema:

Dimension	Values
Primary Disfluency	SR, ISR, MUR, P, B
Secondary Behavior	V, FG, HM, ME
Tension Level	0, 1, 2, 3

[(Valente et al., 31 May 2025), Table 1]

2. Annotation Protocols and Data Quality

Expert annotation is central to FluencyBank's reliability and granularity. Three clinicians, each with significant experience in speech-language pathology (SLP), independently label recordings using ELAN’s multi-tier audio-video interface, with careful calibration on pilot data to standardize onset/offset boundary rules and annotation conventions (Valente et al., 31 May 2025). The protocol includes:

• Span marking using waveform display, with specific onset/offset rules for each event type.
• Assignment of LBDL, SSI-4, and Boey tension labels per span.
• Consensus adjudication for high-disagreement cases, yielding a gold-standard test set through joint audiovisual review.

Inter-annotator agreement (IAA) for time spans reaches Krippendorff’s $\alpha = 0.68$, while secondary behaviors are consistently marked (e.g., facial grimace, $\text{FG}$: $n=563$ in 732 segments). Tension annotation shows lower agreement ($\alpha=0.18$), reflecting the inherent subjectivity of this ordinal rating (Valente et al., 31 May 2025).

Table 3 illustrates gold-standard consensus event counts:

Type	Total	V	FG	HM	ME	None
SR	190	23	114	38	1	53
ISR	143	29	106	25	0	23
MUR	94	10	64	16	0	21
P	93	2	62	34	0	18
B	265	83	202	75	1	21

[(Valente et al., 31 May 2025), Table 3]

3. Applications: Detection, Segmentation, and Assessment

FluencyBank enables development and benchmarking of both conventional rule-based and state-of-the-art neural models for dysfluency detection, segmentation, and severity assessment.

3.1. Rule-Based and Interpretable Models

Rule-based frameworks leverage corpus-level statistics and segment-level annotations to construct transparent dysfluency detectors. These systems implement speaker-rate adaptive thresholds for event duration, spectral correlation, rhythm, and phonemic alignment, achieving robust results. For example, state-of-the-art rule-based detection on FluencyBank attains F1 = 0.83 (overall), with prolongation detection reaching 97–99% accuracy (Zhang, 21 Aug 2025). Hierarchical logic resolves overlaps: blocks take precedence over repetitions and prolongations, and minimum temporal separation is enforced.

Clinicians rate the interpretability and tunability of these systems highly (trust: 4.2/5 for rule-based, $\kappa=0.74$ machine-SLP agreement) and routinely adjust event thresholds to individualize assessment (Zhang, 21 Aug 2025).

3.2. Neural and Hybrid Models

FluencyBank has been used to validate neural architectures including wav2vec 2.0 fine-tuning (Bayerl et al., 2022), multi-agent text repair pipelines (Xu et al., 15 Jan 2026), and end-to-end waveform-to-waveform translation (StutterZero, StutterFormer) (Xu, 21 Oct 2025). For dysfluency detection, embeddings derived from fine-tuned wav2vec 2.0 (MTL) yield up to 27% F1 gains over the vanilla baseline, particularly for interjections, prolongations, and sound repetitions. Table: [(Bayerl et al., 2022), Table 1].

System	Blocks	Interj.	Prol.	Snd Rep.	Wd Rep.
W2V2-BASE	0.30	0.70	0.51	0.50	0.39
W2V2-STL	0.31	0.83	0.52	0.40	0.40
W2V2-MTL	0.33	0.84	0.60	0.60	0.43

[(Bayerl et al., 2022), Table 1]

Speech segmentation frameworks such as StutterCut extend FluencyBank’s capabilities through the FluencyBank++ extension, which adds frame-level boundary annotation for four dysfluency types across 3,017 five-second clips. StutterCut frames segmentation as a graph partitioning task, using whisper-derived embeddings, uncertainty-gated classifier outputs, and Normalized Cut optimization. It achieves time-F1 = 69.3% (±4.9) and onset error ≈ 0.3 s, surpassing prior baselines (Ghosh et al., 4 Aug 2025).

3.3. FluencyBank in Severity Assessment and Prediction

Severity scoring leverages FluencyBank’s three-dimensional annotation, combining stuttered syllable frequency, event duration, and tension level with clinician-defined weights (extensions of the SSI-4 framework) (Valente et al., 31 May 2025). For event prediction (rather than detection), models trained on external corpora (e.g., SEP-28k) but evaluated on the pediatric “Teaching” subset of FluencyBank reach AUC_event = 0.674 and AUC_preblock = 0.655, confirming generalization and the presence of detectable prosodic precursors to severe stuttering events (Kozak, 30 Apr 2026).

4. FluencyBank in Speech Correction and Accessibility

FluencyBank is repeatedly positioned as the gold-standard real-world testbed for evaluating stutter-corrective speech recognition and synthesis pipelines. In multi-stage systems such as STEAMROLLER, FluencyBank recordings are input for evaluating ASR followed by collaborative text repair (e.g., GPT-4o), text-to-speech (TTS) synthesis (e.g., StyleTTS2), and downstream ASR fine-tuning on repaired speech. Baseline ASR models (data2vec, wav2vec2, Whisper) show Word Error Rates (WER) up to 31.6%, which drop to 19.7% after full pipeline repair. More severe stuttering segments see WER reductions from 33.5% to 22.6% (Xu et al., 15 Jan 2026).

End-to-end models (StutterZero, StutterFormer) evaluated on 800 FluencyBank segments (Voices-AWS subset) achieve 0.161 and 0.120 mean WER respectively, substantially outperforming Whisper-Medium's 0.361. BERTScore improvements exceed 30 percentage points, suggesting strong preservation of semantic content and naturalness (Xu, 21 Oct 2025).

These results underscore FluencyBank’s role in benchmarking the robustness of speech technology for people who stutter, especially under varied real-world conditions (microphone, room, background noise, prosody).

5. FluencyBank Extensions and Emerging Benchmarks

Multiple research groups have extended FluencyBank to provide more challenging and diagnostically relevant benchmarks:

• FluencyBank++ (frame-level annotation for segmentation) enables direct evaluation of onset detection and precise segment labeling, filling the gap between clip-level and real-time operation for therapy support (Ghosh et al., 4 Aug 2025).
• Consensus-labeled test sets permit rigorous, clinically relevant comparison of detection and segmentation models, enabling analysis of both automated and clinician performance (Valente et al., 31 May 2025).

Critical design features—annotation guidelines, cross-annotator calibration, and multi-modal (audio-video) input—have increased the corpus’s granularity and clinical utility.

6. Clinical Integration, Transparency, and Limitations

FluencyBank’s real-world segmentation, secondary behavior annotation, and open protocol for patient-specific tuning facilitate its direct integration into clinical workflows. Interpretability is explicitly supported through segment-level rule triggers and speaker-adaptive thresholds (Zhang, 21 Aug 2025). Inter-rater agreement and trust metrics show that SLPs endorse rule-based systems for their transparency, while neural models provide scalability and cross-population generalization (Zhang, 21 Aug 2025, Bayerl et al., 2022).

Notable limitations include subjectivity in tension-level labeling (α = 0.18), partial coverage of certain demographics (e.g., underpowered adult cluttering subset), and variable completeness of metadata (e.g., age, ethnicity) (Valente et al., 31 May 2025, Kozak, 30 Apr 2026). Nevertheless, the corpus continues to evolve, incorporating new annotation standards and evaluation practices.

7. Summary and Prospective Impact

FluencyBank is foundational for advancing automated stuttering assessment, intervention, and accessibility. Its evolution from event-labeled speech to three-dimensional, multi-modal clinical annotation directly reflects the increasing sophistication of both clinical and machine learning approaches. The corpus supports a wide spectrum of research—from interpretable rule-based frameworks to end-to-end neural conversion—while maintaining relevance to real-world therapeutic contexts and the lived experiences of people who stutter. Ongoing development of more nuanced subsets and richer annotation will likely increase its impact on inclusive speech technologies and clinical best practices (Valente et al., 31 May 2025, Zhang, 21 Aug 2025, Xu, 21 Oct 2025, Bayerl et al., 2022, Xu et al., 15 Jan 2026).