PaperlayerView exact revision on arXiv

Source-linked AI summary

Mega-ASR: Towards In-the-wild^2 Speech Recognition via Scaling up Real-world Acoustic Simulation

Zhifei Xie, Kaiyu Pang, Haobin Zhang, Deheng Ye, Xiaobin Hu, Shuicheng Yan, Chunyan Miao

arXiv:2605.19833v1cs.SDcs.AIcs.CLcs.MMeess.AS

TL;DR

Robust ASR remains limited in real-world environments where multiple acoustic effects combine. Mega-ASR pairs a large-scale compound-scenario dataset with progressive acoustic-to-semantic training, reporting over 30% relative WER reductions against prior state-of-the-art systems on challenging real-world conditions.

  • Problem

    Existing robust-ASR datasets mainly cover isolated acoustic conditions, while real-world environments combine effects and routinely push WER beyond 30%.

    Source: §3.1 · p. 3

  • Method

    Mega-ASR combines Voices-in-the-Wild-2M, covering seven acoustic phenomena and 54 compound scenarios, with progressive acoustic-to-semantic fine-tuning and dual-granularity WER-gated optimization.

    Source: §7 · p. 10

  • Results

    Over 30% relative WER reductions are reported against prior state-of-the-art systems on challenging real-world acoustic conditions.

    Source: §7 · p. 10

  • Takeaways & Limitations

    The findings support robust ASR-in-the-wild research on severe, compositional acoustic conditions.

    Source: §7 · p. 10

Abstract

from arXiv · show

Despite rapid advances in automatic speech recognition (ASR) and large audio-language models, robust recognition in real-world environments remains limited by an "acoustic robustness bottleneck": models often lose acoustic grounding and produce omissions or hallucinations under severe, compositional distortions. We propose Mega-ASR, a unified ASR-in-the-wild framework that combines scalable compound-data construction with progressive acoustic-to-semantic optimization. We introduce Voices-in-the-Wild-2M, covering 7 classic acoustic phenomena and 54 physically plausible compound scenarios, and train Mega-ASR with Acoustic-to-Semantic Progressive Supervised Fine-Tuning and Dual-Granularity WER-Gated Policy Optimization. Extensive experiments demonstrate that Mega-ASR achieves significant advantages over prior state-of-the-art systems on adverse-condition ASR benchmarks (45.69% vs. 54.01% on VOiCES R4-B-F, and 21.49% vs. 29.34% on NOIZEUS Sta-0). On complex compositional acoustic scenarios, Mega-ASR further delivers over 30% relative WER reduction against strong open- and closed-source baselines, establishing a scalable paradigm for robust ASR in-the-wild.

1 Introduction

MEGA-ASR targets the acoustic robustness bottleneck with scalable compound-scenario data and progressive acoustic-to-semantic training. The framework combines supervised fine-tuning with a WER-gated reinforcement-learning reward for complex acoustic environments.

1 Introduction

  • Real-world ASR can reach 70% WER with omissions or hallucinations, while prior approaches leave gaps in scenario coverage, compositional robustness, and training-data difficulty.Existing training data emphasize 4%–10% WER, unlike challenging settings above 30%.Source: §1 · p. 2
  • VOICES-IN-THE-WILD-2M uses spectral-manipulation simulation to create data spanning 7 canonical meta-scenarios and 54 physically checked compound scenarios.The construction calibrates difficulty and filters samples above 70% WER to support training stability.Source: §1 · p. 2
  • A2S-SFT progressively trains MEGA-ASR-BASE to extract acoustic information under heavy perturbation and recover intended semantics at medium-to-high WER.The resulting base model provides foundational capabilities for the reward signal used in subsequent reinforcement learning.Source: §1 · p. 2
  • MEGA-ASR substantially outperforms prior state-of-the-art systems on adverse-condition and compositional real-world benchmarks.The introduction presents these experiments as evidence for the framework's performance across both types of conditions.Source: §1 · p. 2
  • DG-WGPO combines token-level refinement and sentence-level semantic reconstruction rewards, dynamically weighting them through a WER-gated mirrored fusion strategy.It addresses the shift above 30% WER from word-level errors toward semantic errors, hallucinations, and dropped sentences, where WER-based rewards alone are ineffective.Source: §1 · p. 2

2 Related Work

Prior ASR models perform strongly on standard benchmarks, but existing research leaves severe, compositional acoustic conditions underexplored. Benchmarks and augmentation methods do not yet provide a systematic model of real acoustic worlds.

ASR Foundation Models and Robust Speech Recognition

  • ASR foundation models achieve strong standard-benchmark results, but clean or mildly noisy performance does not establish robustness to simultaneous deployment degradations.Real-world speech can combine noise, far-field propagation, reverberation, obstruction, device distortion, and transmission dropout.Source: §2 · pp. 2–3

Datasets and Simulation for In-the-wild ASR

  • Adverse-condition benchmarks often focus on isolated factors or mild regimes, while augmentation methods remain local perturbations rather than systematic models of real acoustic worlds.Real environments such as classrooms, corridors, and vehicles combine noise, far-field attenuation, echo, occlusion, and device distortion.Source: §2 · p. 3

3 VOICES-IN-THE-WILD-2M

VOICES-IN-THE-WILD-2M addresses the gap between datasets of isolated acoustic conditions and real-world environments with compound effects. It uses spectrogram-level simulation to build challenging training data and is accompanied by a real-recording evaluation benchmark.

3.1 Overview

  • VOICES-IN-THE-WILD-2M combines seven calibrated atomic acoustic effects into 54 agent-validated scenarios and 2.4M clips; Qwen3-ASR [Shi et al., 2026] averages 35% WER.Existing datasets mostly cover isolated conditions with mild WER of 4%–10%, while real-world mixtures routinely exceed 30% WER.Source: §3.1 · p. 3

3.2 Realistic Simulation of Compound Acoustic Environments

  • The scalable pipeline simulates seven effects on spectrograms, composes two to five physically plausible effects, adopts linear severity sampling, and removes clips above 70% WER.Effects are iteratively tuned against real-world recordings; synthesis uses clean speech and broad collections of real-world background and speech material.Source: §3.2 · pp. 3–4, Figure 3 · p. 4

3.3 Voices-in-the-wild-Bench: A Real-Recording Evaluation Benchmark

  • The released Voices-in-the-wild-Bench contains 5,000 English/Mandarin clips covering seven atomic phenomena, including 3,500 synthetic clips and 1,500 real recordings.The real-world recordings were collected from internet sources and 16 human participants.Source: §3.3 · p. 4

4 Mega-ASR

Mega-ASR combines Acoustic-to-Semantic Progressive Supervised Fine-Tuning with Dual-Granularity WER-Gated Policy Optimization to improve recognition under complex acoustic conditions. Environment-aware routing activates its noise-robust weights when the input requires them.

4.1 Acoustic-to-Semantic Progressive Supervised Fine-Tuning

  • Mega-ASR-Base starts from Qwen3-ASR and progressively trains the encoder and aligner on WER<30%, <50%, then <70% data, tunes the LLM on WER<70% samples, and jointly fine-tunes all components.The sequence builds acoustic perception incrementally, activates semantic recovery under unreliable acoustic evidence, then aligns the encoder, aligner, and LLM end to end.Source: §4 · p. 4, §4.1 · p. 5

4.2 Dual-Granularity WER-Gated Policy Optimization

  • At WER<=30%, errors are predominantly word-level confusions, but above that threshold they shift to sentence-level hallucinations and omissions; standard WER conflates the regimes and saturates under heavy degradation.DG-WGPO addresses this by retaining a static WER-plus-repetition-penalty signal and adding WER-gated fine- and coarse-grained rewards.Source: §4.2 · p. 5

4.2.1 Static Rule-Based Rewards

  • Static rewards anchor learning to the evaluation metric and hard-gate rollouts containing excessive repeated n-grams, filtering degenerate outputs.The static signal combines WER reward with the anti-repetition gate.Source: §4.2.1 · p. 5

4.2.2 Dual-Granularity Dynamic Reward

  • DG-WGPO pairs token-level refinement, which discounts soft substitutions, with sentence-level reconstruction, which preserves sequence backbone and penalizes truncation, then WER-gates their relative weights.Insertions and deletions are treated as hard errors; the sentence reward uses longest-common-subsequence agreement and a length term.Source: §4.2.2 · p. 6

4.3 Environment-Aware Routing for Plug-and-Play Inference

  • A LoRA-tuned binary classifier routes clean and Voices-in-the-Wild inputs to either the original backbone or Mega-ASR’s noise-robust weights, preserving clean-domain performance.The routing is intended to retain complementary clean-speech, hotword, and streaming capabilities while activating Mega-ASR only when the acoustic environment demands it.Source: §4.3 · p. 6

5 Experiments

MEGA-ASR is evaluated on standard, adverse-condition, and compound-condition speech, with results showing strong clean-speech performance and improved robustness under acoustic perturbations. Ablations examine semantic gains, training components, reward design, and hyperparameter sensitivity.

5.1 Experimental setup

  • The experiments initialize from Qwen3-ASR-1.7B and train on VOICES-IN-THE-WILD-2M for both SFT and RL, evaluating standard, adverse-condition, and compound-condition ASR.Evaluation includes clean and multilingual benchmarks, CHiME-4, VOiCES and NOIZEUS, and Voices-in-the-Wild-Bench; standard results are reported with and without dynamic routing LoRA.Source: §5.1 · p. 7

5.2 Main results

  • 6.70 average WER gives MEGA-ASR the best overall result on CHiME-4, VOiCES, and NOIZEUS, ahead of Qwen3-ASR (7.93), Whisper-Large-v3 (10.72), and Qwen2.5-Omni (15.14).On NOIZEUS 0dB, WER is 19.80 versus 23.97 for Qwen3-ASR and 55.78 for Gemini-3-Flash.Source: §5.2 · p. 8
  • 2.73/4.57 WER on mixed degradations beats Whisper-Large-v3 (8.91/14.79) and Gemini-3-Flash (7.99/9.62) on Voices-in-the-Wild-Bench.The reported relative reductions are 65.8%/69.1% over Whisper-Large-v3 and 65.8% over Gemini-3-Flash.Source: §5.2 · p. 8, Table 4 · p. 8
  • With routing, MEGA-ASR improves LibriSpeech WER from 1.78/3.57 to 1.63/3.37 and achieves 3.86/3.17 on Fleurs zh/en.The paper also reports consistent gains on WenetSpeech-meeting and VoxPopuli and says performance remains competitive on clean and multilingual benchmarks.Source: §5.2 · p. 8

5.3 Analysis

  • Missed-content falls from 14.2 to 5.9 versus Qwen3-ASR, indicating semantic-level gains beyond WER reduction, including fewer hallucinations and dropped utterances.The evaluation uses an LLM-as-judge and averages results over Voices and Noizeus.Source: §5.3 · p. 8, Table 7 · p. 9
  • 7.35/7.64 WER for full MEGA-ASR is a 1.59/1.81 reduction over Qwen3-ASR; removing A2S-SFT stages or DG-WGPO components degrades performance.Removing Rstruc causes the largest DG-WGPO degradation (7.54/7.85); vanilla DAPO with Rwer alone beats vanilla GRPO by 0.11/0.13 WER.Source: Table 5 · p. 8, §5.3 · p. 9
  • 19.57s per step for the rule-based reward is 3.2× faster than the LLM-judge variant at 62.23s, while their WER differences remain within roughly 0.1 on Voices and Noizeus and 0.11 on Voi-R.The paper adopts the rule-based reward as default given comparable WER and lower time cost.Source: §5.3 · p. 9, Table 6 · p. 9
  • 7.4 to 9.1 V.N.R. WER when αdyn rises to 0.8 shows greater sensitivity to αdyn than αs; the selected weights are (αdyn, αs)=(0.6, 0.4).At αdyn=0.4, V.F. rises by 0.4 and V.F.R. by 0.3; τ=0.3 is most balanced, while τ=0.5 degrades results.Source: §5.3 · p. 9, Table 8 · p. 9

6 Case study

The case study compares MEGA-ASR with QWEN3-ASR and GEMINI-3-PRO across far-field reconstruction, hallucination, and entity recovery, and reports precise recovery by MEGA-ASR where competitors fail or produce incorrect fluent text.

6 Case study

  • The case study reports that MEGA-ASR’s precise recovery pattern persists under severe noise and in entity-dense utterances.The comparisons cover far-field reconstruction, content hallucination, and entity recovery.Source: §6 · p. 9, Figure 6 · p. 10
  • The examples show that degraded signals and rare entities can cause subtle recognition errors hidden by outputs that appear fluent.GEMINI-3-PRO produces a fluent but incorrect far-field hypothesis, while QWEN3-ASR returns an empty transcription.Source: §6 · p. 9, Figure 6 · p. 10
  • At Peak -5.2 dB, MEGA-ASR achieves WER 0.0% versus QWEN3-ASR’s 100.0% and GEMINI-3-PRO’s 86.1% on far-field reconstruction.QWEN3-ASR returns an empty transcription, while GEMINI-3-PRO fabricates content unrelated to the source.Source: §6 · p. 9, Figure 6 · p. 10

7 Conclusion · A Qualitative Case Studies

MEGA-ASR combines large-scale compound acoustic data with progressive acoustic-to-semantic training and policy optimization to improve robust ASR. Qualitative examples show it mitigating catastrophic sentence-level failures under severe acoustic degradation.

7 Conclusion

  • Over 30% relative WER reductions are reported in challenging real-world acoustic conditions, alongside significant improvements over prior state-of-the-art systems.The conclusion presents MEGA-ASR as a scalable paradigm for robust ASR in-the-wild.Source: §7 · p. 10
  • VOICES-IN-THE-WILD-2M covers 7 classic acoustic phenomena and 54 physically plausible compound scenarios.The dataset is central to the proposed framework for modeling compound acoustic environments at scale.Source: §7 · p. 10
  • Acoustic-to-Semantic Progressive Supervised Fine-Tuning and Dual-Granularity WER-Gated Policy Optimization support robust perceptual recovery and semantic reconstruction.Source: §7 · p. 10

A Qualitative Case Studies

  • The qualitative examples cover five failure patterns and clarify error types reduced by MEGA-ASR rather than replacing quantitative evaluation.The patterns are off-audio hallucination, empty-output collapse, dropout-induced semantic drift, noisy semantic drift, and entity-level recovery on standard noisy benchmarks.Source: A Qualitative Case Studies · p. 15
  • Under severe degradation, baseline errors include cross-lingual hallucinations, empty outputs, semantic drift, and missing key entities rather than only local substitutions.The examples support a shift from ordinary word-level errors to sentence-level semantic failures.Source: A Qualitative Case Studies · p. 15
  • MEGA-ASR often turns catastrophic baseline failures into correct or near-correct transcriptions, recovering the reference utterances’ semantic backbone despite possible minor lexical differences.Examples include off-audio hallucination in a compound case, empty-output collapse under recording coloration, semantic drift under dropout and noise, and changed entities and relations on CHiME-4.Source: A Qualitative Case Studies · p. 15

B Additional Robust Benchmark Results … C.4 Construction of Compound Acoustic Scenarios

The robust-benchmark comparison finds that both adaptation variants improve on Qwen3-ASR-1.7B, while VOICES-IN-THE-WILD-2M is built by composing controllable acoustic effects into realistic compound scenarios. The construction separates incompatible acoustic geometries from portable degradations and preserves coherent processing chains.

B Additional Robust Benchmark Results

  • Both enhanced variants improve the three-benchmark average over Qwen3-ASR-1.7B; Merged-v2 leads on VOiCES and NOIZEUS, while quality-routed 3-LoRA leads on CHiME-4.The benchmarks span real and simulated noisy speech, controlled additive noise at different SNR levels, and far-field room acoustics.Source: B Additional Robust Benchmark Results · p. 16

C.1 Hierarchical Simulation Pipeline

  • VOICES-IN-THE-WILD-2M builds scenarios hierarchically from primitive effects to atomic effects and then to compounds, with fixed processing order to avoid implausible chains.The hierarchy aims to model co-occurring real-world degradation sources while applying low-level mechanisms consistently.Source: §C.1 · pp. 17, 19

C.2 Primitive Acoustic Effects

  • Eight modular, parameterized primitives represent signal-level degradations, including noise, echo, reverberation, distortion, resampling, filtering, loudness change, and frame-level stutter.Their interpretable parameters control degradation strength and are tied to the global severity variable used in synthesis.Source: §C.1 · p. 18, §C.2 · p. 19

C.3 Construction of Seven Atomic Acoustic Effects

  • Seven atomic effects combine ordered primitive chains, making one degradation dominant while retaining realistic secondary artifacts.Examples include noise with loudness normalization, far-field reverberation-filtering-attenuation, and obstructed speech with filtering, reverberation, and attenuation.Source: §C.3 · pp. 20–22
  • Electronic distortion isolates clipping and harmonic artifacts, while transmission dropout uses frame replacement or silence to create temporal discontinuities without changing audio length.The dropout design avoids extra filtering so the challenge remains temporal rather than spectral.Source: §C.3 · pp. 22–23

C.4 Construction of Compound Acoustic Scenarios

  • The 54 scenarios comprise 7 single-effect, 18 two-effect, 13 three-effect, and 16 higher-order categories, combining one anchor at most with portable modifiers.Far-field, echo&reverb, and obstructed define distinct acoustic geometries; recording coloration, electronic distortion, noise, and transmission dropout can accompany different anchors.Source: §C.4 · p. 23
  • Compound signal chains preserve each selected atomic effect’s internal order and repeated operators while removing cross-effect duplicate primitives, except additive noise.Allowing repeated noise supports real environments with multiple independent noise sources.Source: §C.4 · p. 24

C.5 Severity Sampling and Difficulty Calibration

The section defines a globally coherent severity variable for simulated samples and compares four mappings from uniform draws to difficulty. Pilot experiments select linear sampling as the default because it balances coverage, learnability, and difficulty diversity.

C.5 Severity Sampling and Difficulty Calibration

  • A single latent draw x is mapped to global severity m and shared across a sample’s active effects, keeping degradation difficulty internally coherent.This prevents arbitrary combinations of very strong and negligible effects within the same sample.Source: §C.5 · pp. 24–26
  • The four mappings produce distinct difficulty profiles: sqrt-forward favors hard samples, sqrt-backward favors easy samples, gaussian-mid emphasizes intermediate difficulty, and linear sampling is even.Figure 8 visualizes the mappings, all of which map the same uniform draw into the severity range [0, 1].Source: §C.5 · p. 25
  • Severity m instantiates heterogeneous effect parameters through ranges and monotonicity flags, while categorical options use m to select from ordered candidates.Depending on parameter semantics, larger m can mean stronger reverberation or distortion, lower cutoff frequency or target loudness, or higher stutter probability.Source: §C.5 · p. 25

D Router Implementation and Training Details · D.1 Motivation

Mega-ASR adds an environment-aware router to choose between the original Qwen3-ASR backbone and robust LoRA-enhanced weights, preserving the backbone’s complementary capabilities while applying robustness selectively.

D.1 Motivation

  • The original Qwen3-ASR backbone can better preserve clean-speech recognition, hotword recognition, and streaming-style inference, so robust weights are not optimal for every input.This motivates selecting between the base backbone and Mega-ASR’s robust LoRA-enhanced weights based on the acoustic environment.Source: §D.1 · p. 27
  • The router selects the base Qwen3-ASR model for clean inputs and the MEGA-ASR LoRA branch for degraded inputs without generating transcripts or changing ASR decoding.It makes Mega-ASR a plug-and-play robustness module rather than a full replacement of the original ASR system.Source: §D.1 · p. 27

D.2 Router Model Architecture

The router is a lightweight audio-quality classifier that predicts whether an input is clean or degraded. It processes log-Mel features through a compact Transformer-based architecture to limit routing overhead.

D.2 Router Model Architecture

  • The router predicts a binary clean-or-degraded label from log-Mel acoustic features using a single-layer Transformer to minimize routing overhead.Source: §D.2 · p. 27
  • An 80-dimensional log-Mel sequence passes through a downsampling convolutional frontend, sinusoidal positional encoding, one Transformer encoder layer, attention pooling, and a linear binary head.Source: §D.2 · p. 27

D.3 Router Training Data … D.6 Inference Overhead

The appendix describes training a binary router to distinguish clean from degraded speech and using its decision to switch Qwen3-ASR between base and robust LoRA weights. On CHiME-4, routed inference takes 371 seconds versus 374 seconds for direct inference, a relative difference of −0.8%.

D.3 Router Training Data

  • Router training combines clean samples from LibriSpeech, AISHELL-1, CommonVoice22, and WenetSpeech with degraded samples from VOICES-IN-THE-WILD-2M.Clean speech is labeled 0 and degraded speech 1.Source: §D.3 · p. 27

D.4 Training Objective and Optimization

  • Training assigns clean and degraded speech binary labels, optimizes cross-entropy on 16 kHz mono log-Mel inputs, and uses gain/noise augmentation with AdamW and mixed precision.Source: §D.3 · p. 27, §D.4 · p. 28

D.3 Router Training Data

  • The final router dataset contains 552,651 clean and 674,107 degraded samples, split into 1,104,084 training, 61,337 validation, and 61,337 test samples.Source: §D.3 · p. 27
  • The router achieves over 99.5% binary classification accuracy on the held-out development set.The passage attributes this to the reliable acoustic distinction between clean and degraded inputs detectable by a lightweight model.Source: §D.3 · p. 27

D.5 Integration with Qwen3-ASR and LoRA Delta Switching

  • A single Qwen3-ASR-1.7B instance switches precomputed LoRA deltas: at γ = 0.5, pdirty ≥γ selects MEGA-ASR, otherwise the base branch decodes.Switching adds or subtracts the precomputed delta tensors rather than reloading the full model.Source: §D.5 · p. 28

D.6 Inference Overhead

  • 371 seconds versus 374 seconds for direct Qwen3-ASR inference on CHiME-4 yields a relative difference of −0.8%, with no measurable inference overhead in practice.The passage describes the difference as within normal runtime fluctuation.Source: Table 21 · p. 29, §D.6 · p. 29
  • The small router runs once without reloading the ASR model; batch inference groups utterances by routing decision and decodes each group separately to reduce unnecessary switching.Source: §D.6 · p. 29

E Training and Implementation Details … F Additional Related works

The section details staged A2S-SFT and multimodal DG-WGPO training, then situates the work among robust ASR methods, audio-language models, and speech benchmarks. It describes a WER-graded acoustic warm start followed by semantic and joint adaptation, and reward-gated policy optimization for transcription robustness.

E.1 A2S-SFT Hyperparameters

  • All A2S-SFT phases start from Qwen3-ASR-1.7B and use LoRA; the final joint phase sets encoder-aligner learning rates to 5.0 × 10−7 and the LLM rate to 1.0 × 10−6.The smaller acoustic-side rate is intended to preserve Phase I representations while allowing the LLM to adapt to noisy transcriptions.Source: §E.1 · pp. 30–31
  • The final A2S-SFT schedule warms up encoder-aligner adaptation on WER < 30%, < 50%, then < 70% samples, adapts the LLM, and finishes with joint alignment to reduce instability and language-prior reliance.Only Phase I uses the progressive WER curriculum; Phases II and III use the full targeted split.Source: §E.1 · pp. 30, 32

E.2 DG-WGPO Hyperparameters

  • DG-WGPO adapts Qwen3-ASR as a multimodal DAPO policy, initializes from A2S-SFT, and trains on WER-aware splits enriched for medium- and high-WER examples.The adaptation preserves audio preprocessing and prompt construction while exposing acoustic features and text tokens to policy training.Source: §E.2 · p. 32
  • The main DG-WGPO run uses DAPO with GRPO-style group-relative advantages and LoRA updates across encoder, aligner, and LLM; each prompt samples K = 12 transcriptions.Full-scope updates target both acoustic grounding and semantic reconstruction.Source: §E.2 · pp. 32–33
  • DG-WGPO uses T = 0.50 and a WER-gated dynamic reward (τ = 0.5, αs = 0.4, αdyn = 0.6), emphasizing token refinement below the gate and sentence-level recovery above it.Temperature probing balances candidate potential, preservation of base transcription ability, and valid output rate.Source: §E.2 · pp. 34–35, Table 29 · p. 35
  • Reward diagnostics compare references, initial predictions, sampled hypotheses, and reward rankings; checkpoint selection also monitors validation WER, empty outputs, repetitions, overlong outputs, and medium- and high-WER behavior.The checks help detect rewards that favor incomplete, off-audio, or semantically incorrect transcriptions.Source: §E.2 · p. 35

F Additional Related works

  • Traditional robust ASR uses enhancement, normalization, adaptation, multi-condition training, augmentation, and pretraining, while LibriSpeech Panayotov et al. [2015], CHiME Watanabe et al. [2016], and Speech Robust Bench Shah et al. [2025] mainly assess transcription via WER.The section notes that WER-focused evaluation is less suited to reasoning or instruction-following ability.Source: F Additional Related works · pp. 35–36
  • Large Audio Language Models support audio-conditioned instruction following and reasoning but can hallucinate plausible, audio-ungrounded transcriptions; reasoning-based ASR has also been explored (Gong et al. [2024]; Zhifei et al. [2025]).Source: F Additional Related works · p. 36

Related papers

Loading 2605.19833v1…