PaperlayerView exact revision on arXiv

Source-linked AI summary

Selective Listening: Mechanism-Guided Control of Audio Influence in Large Audio-Language Models

Yulin Sun, Kele Xu, Yong Dou

arXiv:2610.11196v1cs.SDcs.CLcs.LG

TL;DR

Task-irrelevant audio can change text-reasoning decisions, and aggregate Accuracy may conceal those paired shifts. The paper localizes actionable late audio pathways and introduces ICAP-Gate to condition their influence on task demand. Across the evaluated LALMs and conditions, it reports lower paired-drift point estimates than ungated inference while matching performance on the evaluated ASR task.

  • Problem

    Task-irrelevant audio can alter text-reasoning decisions, and aggregate Accuracy can hide paired drift when audio-induced repairs and damages cancel.

    Source: §1 · p. 1

  • Method

    The paper uses paired evaluation and targeted interventions to identify actionable late audio pathways, then applies ICAP-Gate to modulate each model’s pathway according to task demand.

    Source: §1 · pp. 1–2, §3.2 · p. 4

  • Results

    Across four LALMs, ICAP-Gate has lower Influence Rate and Answer Flip point estimates than ungated inference in all 16 full-split evaluations and matches ungated performance on the evaluated ASR task.

    Source: §4.2 · p. 6, §6 · p. 9

  • Takeaways & Limitations

    The findings support selective modality influence control as a design principle for robust multimodal reasoning.

    Source: §6 · p. 9

    The reasoning setup assumes the gold answer is determined by text input, with environmental audio or natural speech providing no answer-bearing information.

    Source: §3.1 · p. 3

Abstract

from arXiv · show

Large audio-language models (LALMs) exploit multimodal evidence, yet task-irrelevant audio can alter text-reasoning decisions when listening is unnecessary. Aggregate Accuracy can hide this paired drift because audio-induced repairs and damages may cancel. Paired drift analysis and targeted interventions identify architecture-specific, intervention-sensitive late audio pathways as actionable control points. We introduce ICAP-Gate, which applies mechanism-guided, task-conditioned control to each model's pathway. Across four LALMs, two reasoning benchmarks, and environmental-sound and natural-speech interference, ICAP-Gate has lower point estimates for Influence Rate and Answer Flip than ungated inference in all 16 full-split model--condition evaluations. Fixed suppression degrades automatic speech recognition (ASR) across all four models, whereas ICAP-Gate matches ungated ASR performance by preserving the pathway for explicit audio-demand instructions. ICAP-Gate has lower paired-drift point estimates than mitigation prompting in all four evaluated settings and provides competitive stabilization relative to eight-sample Self-Consistency while using one generation per query; in controlled ARC measurements, Self-Consistency incurs $7.0$--$9.2\times$ ungated latency. These results establish selective modality influence control as a design principle for robust multimodal reasoning.

1 INTRODUCTION

Task-irrelevant audio can change text-reasoning decisions, while aggregate Accuracy may hide these paired shifts. The paper traces the shifts to actionable late audio pathways and introduces task-conditioned ICAP-Gate to stabilize reasoning while preserving evaluated ASR performance.

1 INTRODUCTION

  • Paired decision drift reveals changes hidden by aggregate Accuracy when audio-induced repairs and damages cancel.The measure compares whether the same text question reaches a different decision with task-irrelevant audio.Source: §1 · p. 1
  • Because audio must remain available for explicit audio-demand instructions and audio-relevant tasks, the paper asks where irrelevant audio becomes actionable and whether its influence can be selectively controlled.The motivating problem is that environmental sounds or speech may provide no answer evidence yet alter a text-reasoning decision.Source: §1 · p. 1
  • Targeted interventions identify intervention-sensitive late audio pathways as routes through which irrelevant audio affects downstream decisions.Qwen2.5-Omni provides the highest-resolution causal localization, with actionable late-path control points recurring in Phi-4-MM and Voxtral-Mini-3B.Source: §1 · pp. 1–2
  • Across four LALMs, ICAP-Gate has lower Influence Rate and Answer Flip point estimates than ungated inference in all 16 full-split evaluations while matching ungated WER on the evaluated ASR task.The evaluations span two reasoning benchmarks and environmental and natural-speech interference.Source: §1 · p. 2

2 RELATED WORK

Prior work covers expanding audio-language capabilities, disruption from irrelevant modalities, and methods for locating and routing internal information. This paper connects intervention-based pathway localization with task-conditioned control of irrelevant audio.

2.1 LARGE AUDIO-LANGUAGE MODELS AND MULTIMODAL REASONING

  • Audio-language systems have expanded from linking pretrained language models to audio and speech representations to broader instruction tuning, long-context reasoning, and end-to-end speech interaction.Recent systems expose distinct interfaces, including Qwen2.5-Omni, Phi-4-Multimodal, and Voxtral.Source: §2.1 · p. 2

2.2 ROBUSTNESS TO IRRELEVANT OR DISTRACTING MODALITIES

  • Prior work finds that task-irrelevant audio can perturb text reasoning, while this paper localizes and controls the consequential pathway according to task demand.Li et al. (2026a) examined silence, synthetic noise, environmental sounds, prompting, and Self-Consistency as inference-time mitigations.Source: §2.2 · p. 2, Figure 2 · p. 3

2.3 MECHANISTIC INTERVENTION AND CONDITIONAL MODALITY CONTROL

  • Mechanistic interventions make behaviorally consequential internal computation actionable, and conditional-computation methods regulate information flow through gates, conditioning, attention, and routing.The paper connects these approaches by modulating each LALM’s architecture-specific pathway through which irrelevant audio affects reasoning.Source: §2.3 · p. 3

3 METHODOLOGY

The methodology pairs text-reasoning decisions across clean, irrelevant-audio, and controlled conditions, then uses paired drift and targeted interventions to locate actionable late audio pathways. ICAP-Gate modulates each model’s pathway according to explicit task demand.

3.1 PROBLEM FORMULATION AND REASONING DRIFT

  • Text-reasoning examples are evaluated with clean, irrelevant-audio, and pathway-controlled inference, with audio providing no answer-bearing information.The aligned conditions use the clean prediction as the paired reference for audio-induced changes.Source: §3.1 · p. 3
  • Paired drift measures correctness-state and answer changes that aggregate accuracy can hide; net transition measures track recovery of clean-correct cases and restoration of clean decisions.The measures distinguish individual changes even when opposing transitions cancel in aggregate accuracy.Source: §3.1 · p. 4

3.2 LATE AUDIO PATHWAYS PROVIDE ACTIONABLE CONTROL POINTS

  • Targeted late-path interventions reverse audio-induced answer changes in Qwen2.5-Omni, and analogous actionable pathways recur in Phi-4-MM and Voxtral-Mini-3B.Qwen2.5-Omni clean-correct-to-wrong and clean-stable cases have similar activation RMS but sharply different clean gold-answer margins.Source: §3.2 · p. 4, Figure 3 · p. 5
  • The localized pathway provides an internal control point for audio influence rather than requiring removal of the audio input.The pathway is also important for explicit audio-demand instructions and the evaluated ASR task.Source: §3.2 · p. 4, §3.3 · p. 5

3.3 ICAP-GATE CONDITIONS PATHWAY INFLUENCE BY TASK DEMAND

  • ICAP-Gate modulates each architecture’s selected audio pathway according to explicit instruction demand, preserving it for audio-understanding tasks and attenuating it otherwise.Routing uses the instruction’s explicit audio-demand cue; online inference uses one routing decision and one temporary pathway modulation, with one generation per query.Source: §3.3 · p. 5

4 EXPERIMENTS

The experiments test whether task-conditioned pathway control reduces audio-induced reasoning drift without compromising audio-dependent recognition. Across four LALMs, ICAP-Gate lowers paired-drift point estimates relative to ungated inference, while comparisons with prompting, Self-Consistency, fixed suppression, and speech audits examine its scope and trade-offs.

4.1 EXPERIMENTAL SETUP

  • The evaluation pairs four LALMs with ARC-Challenge and MMLU under FSD environmental-sound and SIB natural-speech interference, using aligned clean, ungated-audio, and gated predictions.The ASR preservation evaluation uses ungated, fixed-attenuation, and ICAP-Gate inference.Source: §4.1 · p. 5

4.2 ICAP-GATE LOWERS PAIRED-DRIFT POINT ESTIMATES ACROSS FOUR LALMS

  • Both paired-drift measures have lower point estimates with ICAP-Gate than ungated inference in all 16 full-split model–condition evaluations.The evaluations include environmental-sound and natural-speech interference; paired metrics reveal changes that aggregate Accuracy may not show.Source: Table 1 · p. 6, §4.2 · p. 6

4.3 ICAP-GATE PROVIDES COMPETITIVE STABILITY WITHIN THE SINGLE-GENERATION REGIME

  • On MMLU-SIB, prompting has 8.87 percentage points higher IR and 11.76 percentage points higher Answer Flip than ICAP-Gate; Self-Consistency has lower drift estimates in all four matched comparisons.Self-Consistency uses eight generations per query, whereas ICAP-Gate uses one.Source: Figure 4 · p. 7, §4.3 · p. 7

4.4 PATHWAY CONTROL PRESERVES EVALUATED ASR CAPABILITY

  • The selected ICAP-Gate pathway and attenuation scale are architecture-adapted, with full-split confirmation scopes reported for each configuration.The configurations are summarized in Table 4.Source: Table 4 · p. 8
  • Fixed pathway suppression increases WER across all four architectures, whereas ICAP-Gate matches ungated WER on the evaluated ASR task for every model.The ASR evaluation uses 400 LibriSpeech dev-clean utterances per model.Source: §4.4 · pp. 7–8
  • Instruction-scoped routing preserves explicit audio requests and avoids false positives on question-content hard negatives, unlike full-query matching.The routing audit and ASR diagnostics are reported in appendices.Source: §4.4 · p. 8

4.5 TRANSCRIPT AUDIT SUPPORTS TASK-IRRELEVANT SPEECH CONSTRUCTION

  • In the MMLU-SIB transcript audit, none of 28 manually reviewed IDF-overlap candidates was potentially relevant or answer-bearing.The audit supports the task-irrelevant construction of the natural-speech condition.Source: §4.5 · p. 8

5 ANALYSIS AND IMPLICATIONS

The analysis presents selective modality influence control as a way to move from robustness evaluation to task-conditioned pathway control. Across heterogeneous audio stacks, actionable sites recur in late audio pathways, while model-specific pathways and attenuation parameters implement the shared rule.

5.1 FROM ROBUSTNESS TO SELECTIVE MODALITY INFLUENCE

  • The MMLU-SIB audit recovered all 14,042 transcripts and found none of 28 manually reviewed IDF-overlap candidates potentially relevant or answer-bearing.The audit supports the construction of speech interference as task-irrelevant.Source: Table 5 · p. 8

5.2 CROSS-MODEL EVIDENCE SUPPORTS A SHARED COMPUTATIONAL ROLE

  • Across Qwen2.5-Omni, Phi-4-MM, and Voxtral-Mini-3B, actionable sites recur in late audio pathways despite distinct audio stacks; both paired-drift measures have lower point estimates than ungated inference in all 16 evaluations.The shared computational role does not imply a shared layer identity.Source: §5.2 · pp. 8–9, Figure 5 · p. 9
  • The shared task-conditioned control rule uses model-specific pathways and attenuation parameters, with modality demand specified at the instruction level.The evaluated interference conditions are environmental sounds and natural speech.Source: §5.2 · p. 9

6 CONCLUSION

The conclusion presents ICAP-Gate as mechanism-guided, task-conditioned control of intervention-sensitive late audio pathways. The accompanying protocol describes paired evaluation and a single-generation online procedure that attenuates the pathway for text-reasoning instructions while preserving it for explicit audio requests.

6 CONCLUSION

  • Across four LALMs, two reasoning benchmarks, and environmental-sound and natural-speech interference, ICAP-Gate has lower Influence Rate and Answer Flip point estimates than ungated inference in all 16 full-split evaluations.It also has lower paired-drift point estimates than mitigation prompting in four settings and competitive stability relative to eight-sample Self-Consistency while using one generation per query.Source: §6 · p. 9

A REPRODUCIBILITY AND EVALUATION PROTOCOL

A.2 EVALUATION SCOPE

  • The protocol evaluates multiple-choice ARC and MMLU under FSD and randomly paired speech interference, with LibriSpeech ASR as the audio-relevant safety task.Each model uses an architecture-adapted late audio pathway and model-specific attenuation scale.Source: §A.2 · p. 13, §A.4 · p. 13

A.3 PAIRED INFERENCE CONDITIONS

  • Clean inference provides the paired reference for audio-induced change, while correctness is assessed independently against the gold answer.Records are aligned by sample identity and task metadata before paired metrics are computed.Source: §A.3 · p. 13

A.5 ANSWER EXTRACTION AND RESPONSE HANDLING

  • Responses are parsed to extract a normalized multiple-choice answer, and first-answer-line truncation is used as a robustness check rather than a second metric definition.The full generated response is recorded under task-specific prompting and response settings.Source: §A.5 · pp. 13–14

A.6 METRICS

  • Influence Rate operates on correctness states, whereas Answer Flip operates on answer identity; NetC2W and NetAnswer summarize directional transitions relative to the clean answer.Metrics are computed from the same aligned records.Source: §A.6 · p. 14

A.7 INSTRUCTION-CONDITIONED ROUTING

  • Instruction-conditioned routing applies a model-specific low scale to text-reasoning instructions and s = 1.0 to explicit ASR or transcription requests.The production rule parses only the instruction, preventing question-content terms from triggering the open pathway.Source: §A.7 · p. 14

A.8 ONLINE ICAP-GATE INFERENCE PROCEDURE

  • Online ICAP-Gate uses one instruction-routing decision and temporary pathway modulation, then generates once without a clean-reference pass or auxiliary relevance model.The pathway hook is removed immediately after generation.Source: §A.8 · pp. 14–15

A.9 STATISTICAL ANALYSIS

The appendix describes paired uncertainty estimation and data-integrity checks, then states boundaries on what the reported intervals and evaluation scope support. Its baseline results show that irrelevant audio changes decisions across models and interference families, with varying drift magnitude.

A.9 STATISTICAL ANALYSIS

  • Primary reasoning comparisons are paired by example, with paired bootstrap intervals and selected sign tests for repair-versus-damage symmetry among discordant cases.The bootstrap uses aligned records; sign tests are used for selected transition analyses.Source: §A.9 · p. 15
  • Reported confidence intervals quantify uncertainty over evaluated examples, not variation across checkpoints, inference backends, or independent stochastic decoding seeds.This distinction is especially relevant for external models with task- and scale-dependent behavior.Source: §A.9 · p. 15
  • The routing audit uses nominal Wilson intervals for matched, repeated-template scenarios, so its intervals are not population-generalization intervals.The audit contains matched scenario families and repeated templates.Source: §A.9 · p. 15

A.10 PAIRED BOOTSTRAP INPUT AUDIT

  • The paired-bootstrap audit reports zero query, sample-identifier, gold-answer, and MMLU-subject mismatches, with a maximum recomputation discrepancy of 0.009 percentage points.Subject-stratified MMLU resampling agrees directionally with the IID bootstrap.Source: §A.10 · p. 15

A.11 PROTOCOL SCOPE

  • The protocol covers ARC and MMLU with environmental sounds and randomly paired five-second speech; semantically similar, conversational, multi-speaker, and adversarial speech remain future evaluation settings.LibriSpeech ASR is the audio-relevant safety task.Source: §A.11 · p. 15

B BASELINE REPRODUCTION

  • The baseline table reports clean and ungated accuracy, Influence Rate, and Answer Flip for all sixteen full-split model–task combinations.The metrics are measured relative to clean inference for the drift outcomes.Source: Table 8 · p. 16, §B.2 · p. 16
  • Representative Qwen transition examples illustrate the analysis categories but are not an additional statistical evaluation.Source: Table 9 · p. 16

B.2 CLEAN AND UNGATED FULL-SPLIT COMPARISONS

  • Irrelevant audio changes decisions across every model and both interference families, with larger drift in several Phi-4-MM and Voxtral conditions.The variation motivates model-specific pathway selection.Source: §B.2 · p. 16

B.3 REPRESENTATIVE TRANSITION EXAMPLES

Paired evaluation makes audio-induced reasoning changes visible beyond aggregate Accuracy, while targeted interventions probe where those changes arise. The section also compares ICAP-Gate with prompting and Self-Consistency.

B.3 REPRESENTATIVE TRANSITION EXAMPLES

  • Table 9 gives qualitative Qwen transition examples selected by category, not an additional statistical evaluation.The baseline protocol also adds paired IR and Answer Flip measures, environmental-sound and speech interference, and sample-aligned records for intervention analysis.Source: §B.3 · p. 16

B.4 MATCHED MITIGATION COMPARISON

  • Prompting has higher drift point estimates than ICAP-Gate in all four matched settings; Self-Consistency has lower estimates in all four, with a resolved advantage on ARC-SIB only.For Self-Consistency, the other three comparisons show no clear difference.Source: §B.4 · p. 17

C.1 DIAGNOSTIC TRACES AND THE LOCALIZATION QUESTION

  • Audio conditions alter internal trajectories and answer-option logits before the extracted answer, but these traces alone do not locate a unique intervention site.Source: §C.1 · p. 17
  • Single residual-state patching does not consistently reproduce the full behavior, motivating module-level ablation and output scaling to find a stable, effective control point.Source: §C.1 · p. 17

C.2 OFFLINE PATHWAY CALIBRATION

The selected Qwen pathway and attenuation scale were calibrated offline before locked full-split evaluation. Activation statistics and interventions inform the mechanistic interpretation, alongside a separate Phi-4-MM architecture check.

C.2 OFFLINE PATHWAY CALIBRATION

  • The Qwen pathway and attenuation scale were selected offline using paired-drift and transition-recovery criteria, then fixed before full-split evaluation.No full-split result was used to select them.Source: §C.2 · p. 18

C.3 ACTIVATION STATISTICS AND THE DIRECTIONAL INTERPRETATION

  • C2W and control cases have similar late-site activation RMS but sharply different clean gold-answer margins, consistent with directional or fusion-related influence rather than generic magnitude increase.The ratios are close to one at late25 and late30.Source: §C.3 · p. 18, Table 12 · p. 19

C.4 MECHANISTIC INTERPRETATION AND SCOPE

  • Drift traces, interventions, and activation statistics identify audio tower.layers.25 as an actionable Qwen2.5-Omni site under the evaluated protocol.The study uses Qwen2.5-Omni-7B for higher-resolution localization, with model-specific late pathways and suppression strengths for the other evaluated models.Source: §C.4 · p. 19

D PHI-4-MM EXTERNAL VALIDATION

  • Phi-4-MM tests cross-architecture transfer by targeting its distinct late speech/audio encoder path, audio embed.encoder.encoders.23.This selected-pathway analysis also examines task- and scale-dependent responses.Source: D PHI-4-MM EXTERNAL VALIDATION · p. 19

D.1 FULL ARC AND MMLU RESULTS

Phi-4-MM shows lower IR and Answer Flip point estimates under gating across its four full-split conditions. The section also describes scale-dependent responses, transition significance testing, and the Voxtral validation scope.

D.1 FULL ARC AND MMLU RESULTS

  • IR and Answer Flip have lower point estimates than ungated inference in all four full-split Phi-4-MM conditions.The transition summaries quantify associated correctness changes.Source: §D.1 · p. 19

D.2 SCALE-DEPENDENT MMLU RESPONSE

  • The two evaluated Phi-4-MM scales produce different transition profiles across MMLU conditions, supporting selection of scale together with pathway rather than treating one scale as universal.Table 14 compares layer 23 across 14,042 examples; positive Net Overall values denote more repairs than damages.Source: §D.2 · pp. 19–20, Table 14 · p. 20

D.3 CROSS-MODEL TRANSITION SIGNIFICANCE

  • Table 15 reports two-sided exact sign-test p-values over discordant paired transitions for the listed model-condition pairs.These are not bootstrap intervals.Source: §D.3 · p. 20, Table 15 · p. 20

E VOXTRAL-MINI-3B CROSS-MODEL VALIDATION

E.1 FULL-SPLIT TRANSFER RESULTS

Full-split results describe model-specific stabilization across environmental sound and speech interference. The section also audits instruction routing and evaluates the ASR trade-off between fixed suppression and instruction-conditioned preservation.

E.1 FULL-SPLIT TRANSFER RESULTS

  • The Voxtral full-split matrix reports model-specific stability patterns across environmental sound and speech interference, interpreting Accuracy separately from drift.The table reports IR and Answer Flip as ungated-to-gated comparisons.Source: §E.1 · p. 20, Table 16 · p. 21

F INSTRUCTION-ROUTING HARD-CASE AUDIT

  • The routing audit tests whether instruction-only cues preserve audio pathways for requests and avoid misleading lexical triggers, using 240 manually curated cases across matched scenarios.The deployed policy inspects the instruction before the first blank line; a full-query condition also exposes question and answer-choice content.Source: F INSTRUCTION-ROUTING HARD-CASE AUDIT · p. 21, §F.1 · p. 21

F.2 ROUTING COVERAGE

  • Instruction-scoped routing prevents question-content keywords from changing decisions on the tested hard negatives.The audit table separates scope-level counts from category-level TP/TN/FP/FN counts.Source: §F.2 · p. 21, Table 17 · p. 21

F.3 IMPLICATION

  • The routing evidence is limited to evaluated instruction categories and is not a general semantic audio-relevance evaluation.Source: §F.3 · p. 22

G.1 PROTOCOL

  • The ASR evaluation compares ungated, fixed-suppression, and ICAP-Gate inference on 400 LibriSpeech dev-clean utterances per model.ICAP-Gate uses the low scale for non-audio instructions and restores the pathway to s = 1.0 for explicit speech-recognition or transcription requests.Source: §G.1 · p. 22

G.3 FIXED SUPPRESSION AND INSTRUCTION-CONDITIONED PRESERVATION

  • Fixed suppression raises WER in every architecture, whereas ICAP-Gate matches ungated WER by routing ASR instructions to the open pathway.Table 18 reports point estimates for all four models and paired bootstrap intervals for matched utterance-level summaries.Source: §G.3 · p. 22, §G.2 · p. 22, Table 18 · p. 22
  • Fixed-gate WER increases range from 16.28 to 123.76 points, with comparisons made within each model because ungated WER varies by model and inference pipeline.Source: §G.3 · p. 22

G.4 PAIRED UNCERTAINTY AND ROUTING AUDIT

  • All 400 ASR instructions per model are marked audio-relevant and assigned the high scale; broader audio-understanding safety remains outside this evaluation.For the two models with matched utterance summaries, ICAP-Gate's paired WER difference is exactly zero in these audited runs.Source: §G.4 · p. 23, §G.5 · p. 23

H SPEECH INTERFERENCE BENCHMARK CONSTRUCTION AND VALIDITY

SIB pairs text questions with independently selected natural speech from LibriSpeech, and its transcript audit supports treating the speech as irrelevant. The evaluation remains limited to random single-speaker speech and excludes several other interference settings.

H SPEECH INTERFERENCE BENCHMARK CONSTRUCTION AND VALIDITY

H.1 SOURCE CORPUS

  • The speech pool is LibriSpeech dev-clean, with 40 speakers, 97 chapters, and approximately 5.39 hours of source speech.Retained clips are 16-kHz mono segments of fixed five-second duration.Source: §H.1 · p. 23

H.2 PAIRING AND REUSE STATISTICS

  • ARC-SIB contains 1,172 unique paired utterances, while MMLU-SIB contains 14,042 pairs sampled with replacement from 2,694 utterances.MMLU-SIB utterances have mean reuse 5.212, median 5, 95th-percentile reuse 9, and maximum reuse 14; 1,170 utterances are shared across tasks, and 99.83% of ARC-SIB utterances also occur in MMLU-SIB.Source: §H.2 · p. 23, Table 20 · p. 24

H.3 CONSTRUCTION PROTOCOL

  • Construction uses fixed five-second clips, randomly cropping long utterances and zero-padding short ones; ARC-SIB samples without replacement and MMLU-SIB with replacement.Recorded defaults are seed zero and no RMS normalization; unrecorded overrides and original crop-offset distributions are excluded from reported claims.Source: §H.3 · p. 24

H.4 TRANSCRIPT-LEVEL VALIDITY AUDIT

  • The transcript audit flags 28 of 14,042 MMLU-SIB pairs (0.20%), but none of the reviewed candidates is judged answer-bearing or potentially answer-relevant.Review classifies 22 flags as lexical false positives and six as weakly topical; Table 20 summarizes the audit, not the routing rule.Source: §H.4 · p. 24, Table 20 · p. 24

H.5 SPEECH-PAIRING SUMMARY

  • The audit supports random single-speaker speech as an irrelevant-speech condition, while conversational, multi-speaker, semantically conflicting, language-controlled, and signal-to-noise-controlled interference remain unevaluated.The pairing is traceable to the defined LibriSpeech pool, and recorded source indices support quantitative analyses across both conditions.Source: §H.4 · p. 24, §H.5 · p. 24
Loading 2610.11196v1…