Chibuzor Okocha/
Interspeech 2026
Interspeech 2026 Research Benchmark
arXiv:2609.07968

Reasoning Beyond Transcription

Can audio-language models reason over disfluent child speech?

Reasoning Beyond Transcription: Audio Language Models on Child Stuttering Speech

Chibuzor Okocha*Christan GrantZoey Liu

University of Florida

DIAGNOSTIC PIPELINE ARCHITECTURE
Zero-Shot Evaluation
INPUT AUDIO STREAMMixed Acoustic Features
Child Speech Disfluency Adult Speech
↓
EVALUATED SYSTEM
AUDIO LANGUAGE MODEL

Direct cross-modal audio processing without intermediate human-written transcript

↓
DIAGNOSTIC PROBES
“Who is speaking?”
“What did the child mean?”
“Can the model reason?”
EXPLORE DIAGNOSTIC FINDINGS

01 / Evaluation Challenge

Why Disfluent Child Speech Challenges Audio Models

Existing speech-language models are primarily trained on fluent adult data. Evaluating ALMs on child stuttering speech presents three compound acoustic and semantic obstacles.

COMPOUND ACOUSTIC STREAM INPUTVoices of Children Who Stutter Benchmark

1. Child Speech

Distinct acoustic pitch, fundamental frequency (F0), unique vocal tract length, and developmental prosodic variations compared to adult benchmarks.

2. Disfluency

Repetitions, prolongations, silent blocks, and filled pauses interrupt word temporal bounds and break standard language model context tokens.

3. Speaker Interference

Interviewer prompts, overlapping dialogue, background scaffolding, and mixed turn-taking disrupt child-focused focus.

Child+Disfluency+Interviewer
AUDIO LANGUAGE MODEL
Direct Semantic Understanding?

02 / Evaluation Tasks

Two Benchmark Tasks for Semantic Evaluation

Instead of standard word error rate (WER), we formulate two end-to-end tasks to probe whether models truly reason over disfluent spoken audio.

TASK 1Free-Form Generation

Child-Focused Semantic Summarization

Given a raw audio recording of a child talking (often mixed with an interviewer), generate a concise summary strictly representing the child's intentions while filtering out adult prompt leakage.

Input:Mixed-speaker raw audio file (no transcript)
Prompt:“Listen to the audio. Summarize what the child intended to convey, ignoring interviewer questions.”
Output:Child-centered summary string
Evaluation Metrics
FluencyReadability
FaithfulnessSemantic match
PurityNo adult leakage
TASK 2Categorical Reasoning

Child Speech Entailment (NLI)

Given raw child audio (premise) and a textual hypothesis, determine whether the hypothesis is entailed by, neutral to, or contradicted by the child's spoken message.

Input:Raw spoken audio + textual hypothesis
Goal:Classify relationship into 1 of 3 categories
ENTAILMENTTrue premise match
NEUTRALUnsubstantiated
CONTRADICTIONDirectly false
Benchmark Standard

Standardized prompts ensure zero-shot evaluation without task-specific fine-tuning.

Interactive Benchmark Explorer

Can You Reason About This Audio?

Test your intuition against state-of-the-art Audio Language Models. Listen to the spoken prompt (or inspect the schematic context), evaluate the hypothesis, and see where ALMs succeed or fail.

RESEARCH PRIVACY & IRB COMPLIANCE NOTICE:To protect child participant privacy, raw clinical audio recordings are restricted. The interactive widget operates in a schematic demonstration mode using placeholder diagnostic profiles. Replace only with authorized public assets.
Case 01: Contradiction Calibration Failure
Difficulty: Medium
Disfluency Type:Repetition
Speaker Setup:Mixed Speaker (Child + Interviewer)
EVALUATION HYPOTHESIS STATEMENT

“The child enjoys playing sports in large noisy groups.”

YOUR DIAGNOSTIC JUDGMENT:
Select a class above to reveal ground truth and model predictions.

03 / Experimental Methodology

Three Diagnostic Reasoning Paths

By comparing end-to-end ALMs against cascaded pipelines and human transcript oracle baselines, our benchmark pinpoints whether failures stem from acoustic feature extraction or semantic reasoning.

PATH A — AUDIO NATIVE (END-TO-END ALM)
Primary Target
Raw Disfluent AudioContinuous Waveform
Audio Language ModelJoint Encoder & LLM
Semantic PredictionSummary / NLI Class
PATH B — ASR + LLM CASCADE
Cascaded Baseline
Raw AudioChild Speech
ASR ModelSpeech-to-Text
Text LLMReasoning Engine
PredictionOutput Label
PATH C — TRANSCRIPT ORACLE
Upper Bound Ceiling
Human TranscriptClean Ground Truth Text
Text LLMPure Language Reasoning
Oracle PredictionReasoning Capacity

Diagnostic Error Localization Strategy

By comparing Path A (Direct ALM) against Path B (ASR Cascade) and Path C (Oracle Transcript), we decouple whether reasoning failures are driven by Acoustic Processing Disruption or Semantic Reasoning Bottlenecks.

04 / Corpus & Benchmark

Benchmark at a Glance

Constructed from the clinical Voices of Children Who Stutter corpus (FluencyBank), systematically structured across acoustic density levels and reasoning difficulties.

22Child Participants
44Total Audio Sessions
22/22
Single-Spk / Mixed Interview
5-10 minsPer Session Duration
INTERACTIVE REASONING DIFFICULTY CONTROLE

Entailment Reasoning Complexity Granularity

Tier Level: Easy

Evidence can be identified directly from a single distinct utterance.

Avg Model Accuracy:44.9%
Avg Macro F1:0.343
CONCEPTUAL SPATIAL-TEMPORAL REASONING SCOPE
Direct Single Utterance Match“I like blue bikes.” → Direct match with minimal temporal span.
Scope: Single Local Temporal Window

05 / Evaluated Systems

Model Landscape & Taxonomy

A comprehensive evaluation of state-of-the-art Audio Language Models alongside established ASR + Text LLM cascades.

Zero-shot evaluation with fixed task prompts

Direct End-to-End Audio Language Models (ALMs)

Process continuous acoustic features directly without text transcription

Audio Flamingo 3🤗
NVIDIA ResearchMultimodal Audio-LM
Audio Flamingo 2🤗
NVIDIA ResearchMultimodal Audio-LM
Kimi-Audio🤗
Moonshot AINative Audio LM
Qwen2.5-Omni🤗
Alibaba QwenOmni Speech-Text LM
Qwen2-Audio🤗
Alibaba QwenSpeech LM
SALMONN🤗
Tsinghua / ByteDanceSpeech-Audio LM
GAMA🤗
UMD / Speech LabGeneral Audio LM
MIT CSAILListen-Think-Understand

Cascaded Pipeline Components (ASR + Text LLM)

Separate acoustic transcription from downstream textual reasoning

ASR FRONTENDS
Whisper Large-v3OpenAI
Granite SpeechIBM
TEXT REASONING BACKENDS
Qwen 2.5Alibaba
Llama 3.2Meta
MistralMistral AI

06 / Headline Benchmark Findings

Direct Audio Models vs. Cascaded Pipelines

Evaluating whether end-to-end continuous audio processing matches traditional ASR transcription combined with high-capacity text LLMs.

STRONGEST END-TO-END ALM

Qwen2.5-Omni

Overall Accuracy0.681
Macro F1 Score0.683
+5.8% Accuracy Gap
Cascade Advantage
STRONGEST ASR + LLM CASCADE

Whisper + Qwen 2.5

Overall Accuracy0.739
Macro F1 Score0.737

Primary Scientific Message

Providing a strong text LLM with a transcript recovers substantial semantic reasoning performance (+5.8% accuracy), indicating that modern ALM failures stem primarily from disfluent acoustic/speaker processing rather than semantic reasoning capabilities.

Critical Calibration Failure Mode

High Accuracy Can Hide a Failure Mode

Standard aggregate accuracy metrics mask severe class-wise calibration imbalances. Several state-of-the-art ALMs collapse onto entailment predictions, achieving baseline accuracy while failing entirely on contradiction detection.

Model SystemOverall AccMacro F1Entailment AccContradiction Acc
Qwen2.5-Omni68.1%0.68381.2%55.0%
Kimi-Audio64.7%0.58275.4%81.1%
Audio Flamingo 338.6%0.28097.1%11.9%
SALMONN33.4%0.18099.0%0.0%
LTU33.0%0.17098.5%0.0%
Zero-Contradiction Collapse (SALMONN & LTU = 0.000):Models like SALMONN and LTU exhibit 98-99% entailment accuracy but achieve exactly 0.0% contradiction accuracy, demonstrating that high aggregate accuracy can stem from extreme prediction bias rather than true semantic comprehension.
“Aggregate accuracy alone can obscure semantic calibration failures.”

07 / Impact of Reasoning Complexity

Reasoning Difficulty Breakdown

Evaluating performance across Easy, Medium, and Hard inference demands.

EASYTier Benchmark

One Utterance Direct Retrieval

Evidence can be identified directly from a single distinct utterance.

Avg Model Accuracy:44.9%
Avg Macro F1:0.343
MEDIUMTier Benchmark

Local Context Integration

Requires paraphrasing and light inference across adjacent utterances.

Avg Model Accuracy:43.9%
Avg Macro F1:0.357
HARDTier Benchmark

Long-Range Integration & Speaker Filtering

Requires integrating multi-utterance narrative while filtering adult interviewer speech.

Avg Model Accuracy:41.7%
Avg Macro F1:0.327
Nuanced Scientific Finding: Modest Monotonic Trend vs Persistent Class Bias

Performance changes modestly across difficulty tiers (Easy: 44.9% → Hard: 41.7%). The primary bottleneck is not a dramatic collapse on longer context, but rather that systemic class calibration biases (overpredicting entailment) persist uniformly across all difficulty tiers.

08 / Acoustic Stress Testing

Disfluency Stress Test & Density Analysis

Measuring model reasoning stability across disfluency types and acoustic disfluency density bins.

ACOUSTIC DISFLUENCY PHENOMENA IN BENCHMARK
FP:Filled Pause“um... uh...”
REP:Repetition“I-I-I want”
PRO:Prolongation“s...ssss-sun”
BLK:Silent Block“[pause] ...ball”
MIX:Mixed DisfluencyCompound patterns

Performance Across Disfluency Density Bins

Ratio of disfluent events to total word count

Low (~2.3%)2.3%
54.2%

Minimal disfluencies (isolated pauses)

Moderate (~4.7%)4.7%
56.7%

Frequent repetitions / single prolongations

High (~8.6%)8.6%
66.2%

Dense repetitions and part-word blocks

Severe (~17.0%)17.0%
52.9%

Severe stuttering with multi-factor disfluencies

Non-Monotonic Empirical Trend:Performance does not follow a simple linear decreasing trajectory across disfluency density. Rather, performance remains variable across low-to-moderate disfluency densities, with severe degradation occurring primarily at the highest density bin (17.0% disfluency density).
Qualitative Diagnostic Gallery

Diagnostic Failure Explorer

Filter case studies by error taxonomy to analyze acoustic processing breakdowns, speaker leakage, and contradiction calibration failures.

Filter:
Contradiction BiasDifficulty: Medium

Case 01: Contradiction Calibration Failure

“The child enjoys playing sports in large noisy groups.”

Ground Truth: CONTRADICTIONInspect Case →
Speaker LeakageDifficulty: Hard

Case 02: Adult Interviewer Interference Leakage

“The child travelled to Florida last summer.”

Ground Truth: NEUTRALInspect Case →
High DisfluencyDifficulty: Easy

Case 03: Severe Disfluency Breakdown

“The child's favorite toy is a dinosaur.”

Ground Truth: ENTAILMENTInspect Case →
Semantic FailureDifficulty: Hard

Case 04: Long-Range Narrative Inference

“The child owns a pet dog named Max.”

Ground Truth: ENTAILMENTInspect Case →

09 / Summarization Benchmark

Child-Focused Summarization Metrics

Evaluating free-form generation quality across Fluency, Faithfulness, and Purity (resisting adult interviewer prompt leakage).

Sample 01: School Activity Narrative (Placeholder Schematic)Child (Age 8) + Interviewer
HUMAN REFERENCE / ORACLE SUMMARY

“The child explains that they enjoy science experiments at school, specifically growing plants in small cups, and want to show their project to their grandmother.”

EVALUATED MODEL SUMMARIES & METRICS:
Audio Flamingo 3Top Performer

“The child likes science class experiments involving plant growth and wants to share the result with their grandmother.”

Fluency:High
Faithfulness:High
Purity:High (No Leakage)
Kimi-AudioTop Performer

“The child enjoys science at school with plant cups and hopes to show their family.”

Fluency:High
Faithfulness:High
Purity:High (No Leakage)
Qwen2.5-Omni

“The interviewer asks about science class. The child mentions planting seeds in cups and talking to their grandmother.”

Fluency:High
Faithfulness:Moderate
Purity:Moderate
Strongest Reported Summarization Systems:Audio Flamingo 3 and Kimi-Audio demonstrated the highest overall summarization quality, maintaining high semantic faithfulness while successfully isolating the child's intentions from adult interviewer scaffolding.
Research Roadmap

Next Steps & Future Research Directions

Addressing current ALM diagnostic vulnerabilities requires joint acoustic modeling, calibration refinement, and speaker-targeted processing.

Acoustic ModelingPILLAR 01

Disfluency-Aware Audio Pre-training

Incorporating clinical disfluent child speech acoustics (repetitions, prolongations, blocks) directly into self-supervised audio encoder pre-training loss functions rather than treating them as out-of-distribution noise.

Open Benchmark Frontier
Calibration & DPOPILLAR 02

Calibrated Entailment & Contrastive Alignment

Mitigating severe entailment overprediction bias through Direct Preference Optimization (DPO) and class-balanced contrastive alignment to ensure robust contradiction classification.

Open Benchmark Frontier
Speaker DiarizationPILLAR 03

Target-Speaker Acoustic Extraction

Integrating frontend acoustic speaker separation and diarization directly into the audio encoder representations to prevent adult interviewer prompt leakage in mixed-speaker clinical interviews.

Open Benchmark Frontier
Context MemoryPILLAR 04

Long-Span Temporal Block Retention

Developing continuous temporal memory representations capable of maintaining narrative state across extended silent stuttering blocks and multi-second disfluent pauses.

Open Benchmark Frontier

10 / Summary & Conclusions

Key Research Takeaways

Three core conclusions from our diagnostic evaluation on child stuttering speech.

TAKEAWAY 01

Useful High-Level Extraction Capabilities

Modern Audio Language Models can extract useful high-level semantic meaning directly from disfluent child speech without requiring clean intermediate text transcripts.

Interspeech Benchmark
TAKEAWAY 02

Acoustic & Speaker Interference Bottlenecks

Mixed adult-interviewer dialogue and disfluent acoustic inputs (repetitions, blocks) remain primary sources of failure, where cascades with strong ASR still hold an advantage.

Interspeech Benchmark
TAKEAWAY 03

Class-Wise Calibration Bias Hidden by Accuracy

Aggregate accuracy obscures severe calibration collapse. Models frequently overpredict entailment (achieving up to 99% entailment accuracy while scoring 0% on contradictions).

Interspeech Benchmark
CENTRAL FINDING
“Reasoning beyond transcription requires more than hearing the words.”

Interspeech 2026 • Audio Language Models on Child Stuttering Speech