Paper deep dive
TASU2: Controllable CTC Simulation for Alignment and Low-Resource Adaptation of Speech LLMs
Jing Peng, Chenghao Wang, Yi Yang, Lirong Qian, Junjie Li, Yu Xi, Shuai Wang, Kai Yu
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%
Last extracted: 4/10/2026, 4:30:59 AM
Summary
TASU2 is a controllable CTC simulation framework for Speech LLM post-training that generates pseudo-CTC posterior distributions conditioned on transcripts and target Word Error Rate (WER) intervals. By simulating supervision that mimics acoustic decoding interfaces without requiring paired audio, TASU2 enables principled curriculum design and improves performance in low-resource domain adaptation while mitigating source-domain degradation.
Entities (5)
Relation Signals (3)
TASU2 → improves → Speech LLM
confidence 95% · TASU2 strengthens text-only alignment beyond TASU, improving recognition on the source domain
TASU2 → uses → CTC
confidence 95% · TASU2 generates pseudo CTC posterior distributions
TASU2 → outperforms → TTS-based augmentation
confidence 90% · TASU2 consistently outperforms strong baselines including text-only fine-tuning and TTS-based augmentation
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Speech LLM post-training increasingly relies on efficient cross-modal alignment and robust low-resource adaptation, yet collecting large-scale audio-text pairs remains costly. Text-only alignment methods such as TASU reduce this burden by simulating CTC posteriors from transcripts, but they provide limited control over uncertainty and error rate, making curriculum design largely heuristic. We propose \textbf{TASU2}, a controllable CTC simulation framework that simulates CTC posterior distributions under a specified WER range, producing text-derived supervision that better matches the acoustic decoding interface. This enables principled post-training curricula that smoothly vary supervision difficulty without TTS. Across multiple source-to-target adaptation settings, TASU2 improves in-domain and out-of-domain recognition over TASU, and consistently outperforms strong baselines including text-only fine-tuning and TTS-based augmentation, while mitigating source-domain performance degradation.
Tags
Links
- Source: https://arxiv.org/abs/2604.08384v1
- Canonical: https://arxiv.org/abs/2604.08384v1
Trouble viewing inline? Open PDF directly →
Full Text
26,034 characters extracted from source content.
Expand or collapse full text
TASU2: Controllable CTC Simulation for Alignment and Low-Resource Adaptation of Speech LLMs Jing Peng 1,∗ , Chenghao Wang 2,∗ , Yi Yang 1 , Lirong Qian 1 , Junjie Li 1 , Yu Xi 1 , Shuai Wang 3 , Kai Yu 1,∗ 1 X-LANCE Lab, Department of Computer Science and Engineering, Shanghai Jiao Tong University, China 1 MoE Key Lab of Artificial Intelligence, 1 Jiangsu Key Lab of Language Computing 2 AISpeech Ltd, Suzhou, China 3 Nanjing University, China jing.peng, kai.yu@sjtu.edu.cn, voldbaboon@gmail.com, shuaiwang@nju.edu.cn Abstract Speech LLM post-training increasingly relies on efficient cross- modal alignment and robust low-resource adaptation, yet col- lecting large-scale audio-text pairs remains costly. Text-only alignment methods such as TASU reduce this burden by simu- lating CTC posteriors from transcripts, but they provide lim- ited control over uncertainty and error rate, making curricu- lum design largely heuristic. We propose TASU2, a control- lable CTC simulation framework that simulates CTC poste- rior distributions under a specified WER range, producing text- derived supervision that better matches the acoustic decoding interface. This enables principled post-training curricula that smoothly vary supervision difficulty without TTS. Across mul- tiple source-to-target adaptation settings, TASU2 improves in- domain and out-of-domain recognition over TASU, and con- sistently outperforms strong baselines including text-only fine- tuning and TTS-based augmentation, while mitigating source- domain performance degradation. Index Terms: speech large language models, speech recogni- tion, domain adaptation 1. Introduction The rapid progress of large language models has accelerated Speech LLM research [1, 2]. However, strong Speech LLM performance often comes with heavy reliance on large-scale audio–text pairs and compute-intensive pipelines [3], making post-training, adaptation, and reproduction costly. Recent stud- ies therefore revisit lightweight alignment between speech and text representations, including TASU [4], LegoSLM [5], and AlignFormer [6]. Among them, TASU (Text-only Alignment for Speech Un- derstanding) is particularly appealing because it enables text- only post-training: it stochastically simulates CTC posteriors from transcripts, allowing training without paired audio while retaining real-audio inference. In practice, TASU can serve as an effective curriculum and improves recognition both on the source domain and under domain shift. Yet its simulation pro- vides limited contro l over posterior uncertainty and the result- ing error regime, making difficulty scheduling largely heuristic. In parallel, low-resource speech understanding remains a persistent challenge [7]. Many target domains lack sufficient paired audio, and straightforward audio-based fine-tuning can yield unstable gains and noticeable source-domain degrada- tion [8, 9]. Text-centric adaptation has been explored to reduce * These authors contributed equally. ** indicates the corresponding author. the need for paired audio, including parameter-efficient tun- ing [10] and text-only updates of the language component [11]. However, using plain text as the training signal still suffers from a mismatch to the acoustic decoding interface, and its improve- ments can be limited compared to stronger audio-based aug- mentation baselines such as TTS [12]. To address both limitations, we propose TASU2, a control- lable text-to-CTC simulation framework for Speech LLM post- training. Instead of relying on unconstrained stochastic simula- tion, TASU2 generates pseudo CTC posterior distributions un- der a specified target WER range, producing text-derived super- vision that better matches the acoustic decoding interface. This enables a principled curriculum that smoothly varies supervi- sion difficulty and error profiles without TTS or paired audio. Experiments show that TASU2 strengthens text-only align- ment beyond TASU, improving recognition on the source do- main and under cross-domain generalization without any audio- text training. More importantly, under low-resource transfer set- tings, TASU2 consistently outperforms strong baselines includ- ing text-only adaptation [11] and TTS-based augmentation [12], while better preserving source-domain performance. In summary, our contributions are: • We introduce a WER-conditioned text-derived post-training signal by simulating calibrated CTC posterior distributions from transcripts, bridging the gap between plain text supervi- sion and the acoustic decoding interface. • We develop a controllable text-to-CTC simulator that gener- ates posterior sequences under a specified WER range, en- abling explicit control over supervision difficulty and error profiles for curriculum design. • We demonstrate consistent gains over TASU on both source- domain and generalization evaluations without audio train- ing, and competitive improvements in low-resource transfer over text-only and TTS-based augmentation baselines. 2. Text-only Alignment: From TASU to TASU2 CTC introduces a blank symbol and marginalizes over align- ments, collapsing frame-level posteriors into compact label se- quences via blank removal and repetition merging [13, 14]. This compact representation has inspired several speech-LLM alignment methods, like AlignFormer and LegoSLM. However, these methods still rely on paired audio–text data and exhibit performance drops under limited supervision. TASU eliminates the need for paired data by simulating CTC posteriors from text alone.During inference, Label- arXiv:2604.08384v1 [eess.AS] 9 Apr 2026 One-hot Vector Real CTC LPEncoder Decoder werLP Concat Q K, V Pseudo CTCCE Loss Simulator LP b. The Structure of Simulator and Training details LibriSpeech Corpus Augmented + Original Dataset Low-WER Subset (WER < 5%) Interval 1 wer: 0–6% Interval 2 wer: 10–40% Interval 3 wer: 50%+ c. Simulator Training Data Construction 3-Level Data Augmentation Evenly divided into 3 WER Intervals Simulator One-hot Vector 5% wer 20% wer 50% wer Pseudo CTC a. The function of Simulator Hello, this is Daniel speaking. Figure 1: An Overview of TASU2. Synchronous Decoding (LSD) [15, 16] compresses real audio- derived CTC posteriors P∈ R T×V by removing frames where the blank probability exceeds a threshold τ : P ′ t = ( ∅,if P t (< blank >) > τ, P t , otherwise, (1) and then merges consecutive identical frames via averaging: P ′ t = 1 |S j | X t∈S j P ′ t , j = 1,...,J.(2) For text-only training, CTC Posterior Simulation (CPS) gener- ates pseudo-posteriors ̃ S from token sequences by applying ran- dom label smoothing ( ̃ p = αδ y + (1− α) 1 V 1), deletions, and insertions (blanks or duplicates). A projector trained on these simulated posteriors maps them into a frozen LLM, enabling zero-shot speech recognition and, when used as a curriculum pre-training stage, improved domain generalization. Despite its effectiveness, TASU’s simulation remains un- controlled and may not fully capture real acoustic-phonetic confusions. This raises two questions: controllability and fidelity, which we focus on in TASU2. 3. TASU2: WER-Controllable Text-to-CTC Posterior Simulation As motivated in related work, a key question in text-only align- ment is whether a simulator can generate CTC-like posteriors that are both (i) close to real acoustic posteriors and (i) con- trollable to support principled curricula and low-resource adaptation. We address this with TASU2, which learns a con- trollable text-to-CTC posterior simulator that outputs pseudo CTC posterior sequences conditioned on a transcript and a dis- crete WER control code. Fig. 1 overviews the pipeline. 3.1. Task and Notation Let the transcript be a token sequence y = (y 1 ,...,y U ) from a tokenizer with vocabulary size V (including the CTC blank). TASU2 learns a simulator that maps (y,c) to a posterior- frame sequence ˆ P = ( ˆ p 1 ,..., ˆ p T ), where ˆ p t ∈ [0, 1] V and P V v=1 ˆ p t,v = 1. The control code c ∈ 1,...,K indicates a target WER interval (e.g., 1,2,3 or low/medium/high). 3.2. Training Signal: Distribution-level Supervision As described in detail in Algorithm 1, for each utterance, a teacher ASR system provides a real CTC posterior sequence P = (p 1 ,...,p T ⋆ ). Since T ⋆ varies, we pad to a fixed hori- zon T train with a validity mask m ∈ 0, 1 T train . We train the simulator with posterior-level cross-entropy: L sim (θ) = E " − 1 P t m t T train X t=1 m t V X v=1 p t,v log ˆ p t,v # .(3) This distribution-matching objective encourages CTC-like structure like blank dominance and token confusability, which is more faithful to acoustic decoding than text-only perturba- tions. 3.3. Architecture Design of Simulator We instantiate the simulator as a lightweight Transformer encoder–decoder (Fig. 1(b)). The transcript is embedded and combined with a learned embedding for the WER code c, which conditions generation. The decoder autoregressively outputs one posterior frame per step. During training we use teacher forcing; at inference time, the simulator runs autoregressively conditioned only on (y,c). 3.4. WER-conditioned Data Construction To obtain controllable supervision, we need the same tran- script paired with teacher posteriors spanning different error regimes. Following Fig. 1(c), we sample only a one-seventh portion of LibriSpeech and generate multiple augmented vari- ants (noise/reverb). For each variant, we run a teacher ASR model to obtain a greedy hypothesis ̃ y and CTC posteriors P. We compute WER( ̃ y,y) and map it to one of K WER intervals to form the control code c. Using discrete intervals reduces sen- sitivity to measurement noise and alleviates skewed WER dis- tributions, yielding a more stable control signal. And we train the simulator on this data. 3.5. Simulator Fidelity Analysis We further assess whether the simulator indeed reproduces acoustic-like CTC posteriors and whether the WER control be- haves as intended. We consider two complementary aspects: (i) Posterior similarity to teacher CTC (unconditioned). We compare TASU2 against the TASU baseline simulation (without Algorithm 1: TASU2: WER-Conditioned Text-to- CTC Simulation Input: Transcript y, vocab size V (incl. blank), WER binsI k K k=1 , teacher ASRT , augmentorA, horizon T train Output: Simulator parameters θ 1 (1) Build multi-WER supervision 2 foreach utterance (x,y) do 3Generate augmented waveformsx (j) ←A(x) 4foreach x (j) do 5Obtain teacher posteriors P (j) ←T CTC (x (j) ) 6Obtain greedy hypothesis ̃ y (j) ←T greedy (x (j) ) 7Compute WER (j) = WER( ̃ y (j) ,y) and assign c (j) ← bin(WER (j) ;I k ) 8Store tuple (y,c (j) ,P (j) ) 9 (2) Train conditional simulator 10 while not converged do 11Sample a tuple (y,c,P) 12Pad P to length T train and build mask m 13Predict ˆ P← p θ (·| y,c) with teacher forcing 14Update θ by minimizingL sim in Eq. (3) Table 1: Posterior similarity metrics (lower is better except Acc). MethodCE↓KL↓Acc↑ProbDiff↓ TASU-style baseline2.51492.33720.82200.0813 TASU2 (AR simulator) 1.2296 1.0497 0.87820.0615 WER conditioning) using distribution-level metrics between teacher posteriors P and simulated posteriors ˆ P: CE = 1 N N X t=1 − V X v=1 p t,v log ˆp t,v , KL = 1 N N X t=1 V X v=1 p t,v log p t,v ˆp t,v , Acc = 1 N N X t=1 I arg max v p t,v = arg max v ˆp t,v , ProbDiff = 1 N N X t=1 max v p t,v − max v ˆp t,v , (4) where N is the number of valid frames. As shown in Table 1, TASU2 yields substantially lower CE/KL and improved argmax agreement, indicating closer matching to real CTC posteriors. (i) WER controllability (conditioned). To verify controllabil- ity, we inject WER-bin codes and decode the simulated poste- riors with a fixed decoder, then measure the realized WER for each bin. Fig. 2 reports the empirical WER distributions across bins, showing that TASU2 reliably separates error regimes and tracks the target intervals. 4. Experiments To validate the effectiveness and controllability of TASU2, we evaluate (i) text-to-CTC alignment quality and cross-domain bin 1 (0-6%) bin 2 (10-40%) bin 3 (50-150%) 0.005 0.01 0.02 0.06 0.1 0.14 Probability Density WER Distribution bin 1 bin 2 bin 3 Figure 2: WER controllability under WER-bin conditioning. generalization, and (i) low-resource domain adaptation. 4.1. Model Architecture Our simulator is a Transformer encoder–decoder with 6 encoder and 6 decoder layers and hidden size 512. It takes a tran- script and a WER-bin ID as input and autoregressively gener- ates pseudo CTC posterior frames. For the Speech LLM, we follow TASU [4] and use SenseVoice-Small [17] as the speech encoder and Qwen2.5-1.5B [18] as the LLM, connected by a Linear–SiLU–Linear projector. 4.2. Datasets We pre-train the CTC simulator on one-seventh of LibriSpeech (960h) [19] with augmentation, to learn a general mapping from linguistic units to acoustic-like posterior distributions. For Stage 2 post-training, we synthesize pseudo CTC supervision from text-only corpora in LibriSpeech and Medical [20]. We evaluate on LibriSpeech (test-clean/other), Medical (8h), TED- LIUM 3 [21], SlideSpeech [22], and CoVoST2 En→Zh [23]. 4.3. Training and Evaluation Setup WER-conditioned simulation.We discretize the realized teacher WER into three bins: 0–6% (low), 10–40% (medium), and 50–150% (high). Unless otherwise stated, we use bin 1 for conditioning in the main experiments. Two-stage post-training. Stage 1 trains on simulated supervi- sion with learning rate 5× 10 −5 for 5 epochs. Stage 2 adapts the model using simulated text from the target domain (Medical or SlideSpeech), following the same optimization protocol. Optimization. We use AdamW with DeepSpeed ZeRO-2 on 8×Ascend 910B NPUs. LoRA (rank=16, α=32) is used only for the two-stage adaptation setting in Table 4. 5. Evaluation And Analysis We evaluate TASU2 from three perspectives. Table 2 probes whether simulated posteriors improve CTC-style alignment and cross-domain generalization without audio training. Ta- ble 3 provides a lightweight multi-task sanity check beyond ASR. Our key result is Table 4, where TASU2 achieves strong low-resource target gains while largely preserving source- domain performance, outperforming both text-only adaptation and TTS-based augmentation. Table 2: Alignment and generalization of simulated CTC training. All models are trained on LibriSpeech and evaluated on LibriSpeech, SlideSpeech, and TED-LIUM; WER (%)↓ is re- ported. TASU2 conditions: ● None (no WER conditioning); ❍ WER-binned (coarse 3-level WER bins; we use bin 1 here). System Train DataLibri Slide TED Text(Audio, Text) clean/other TASULibri–4.57 / 9.90 24.07 19.36 TASULibri+Slide–4.21 / 10.31 18.70 13.23 SLAM-CTC–Libri3.13 / 8.59 18.59 14.61 TASU2 ● Libri–4.63 / 9.82 16.41 14.02 TASU2 ❍ Libri–3.41 / 8.15 17.67 14.49 TASU2 ❍ Libri+Slide–3.94 / 8.25 17.31 13.93 5.1. Alignment and Generalization We evaluate simulated CTC training on both in-domain and out- of-domain test sets to probe alignment and generalization of TASU2. We start from ASR and further extend the evaluation to a multi-task speech understanding setting. ASR performance. As illustrated in Table 2, relative to the original TASU random simulation, TASU2 consistently im- proves in-domain recognition while yielding much stronger cross-domain transfer.In particular, TASU2 reduces WER markedly on SlideSpeech and TED-LIUM3, and in several cases becomes competitive with, or even surpasses, the strong SLAM-CTC audio baseline, despite using text-only training inputs. Moreover, when we expand the text side by adding SlideSpeech transcripts (Libri+Slide), TASU2 further improves on the target and TED sets, indicating that text expansion and posterior simulation are complementary: additional target-style text benefits transfer, while TASU2 provides the CTC-like su- pervision signal needed to make such text effective. To understand the role of controllability, we compare two simulator variants for generating pseudo posteriors: uncondi- tioned simulation (None) and WER-binned simulation (coarse 3-level bins; we use bin 1 here). The ablation reveals a clear trade-off. The unconditioned simulator tends to improve out- of-domain recognition (better Slide/TED WER), but it incurs noticeable regression on the source domain (worse Libri). In contrast, WER-binned conditioning yields a better balance be- tween alignment and transfer: it preserves strong generalization while substantially reducing source-domain degradation. These results support our design choice that lightweight, discrete er- ror control is useful not only for curriculum design, but also for stabilizing domain shift behavior in simulated-CTC training. Table 3: TASU2 performance under multi-task evaluation. TASU2 uses WER-conditioned simulation with discretized WER bins; here we use only bin 1 for conditioning. Model Train audio duration(h) LibriSpeech↓ clean/other (WER↓) CoVoST2↑ En2Zh (BLEU↑) TASU06.47 / 10.3533.35 TASU203.68 / 8.2533.08 SLAM1.8k3.30 / 7.2437.34 Step-Audio>1000k2.36 / 6.32– Qwen2.5-Omni >1000k2.37 / 4.2141.40 Multi-task performance. Table 3 provides a lightweight multi- Table 4: Two-stage domain adaptation (source→ target). All systems are first trained on the source domain (LibriSpeech) and then adapted to the target domain (Medical). We report WER%↓. LoRA is used for all systems for better performance. SystemTrain dataLibriSpeech Medical Stage 1Stage 2clean/othertest SLAM-CTCAudio –2.43 / 6.0717.55 Raw text [11]2.56 / 6.6213.62 TTS audio2.72 / 6.8012.79 Raw audio2.70 / 6.7612.35 TASU2 ❍ Text Sim-CTC–2.94 / 7.1615.34 Text Sim-CTC Text Sim-CTC2.96 / 7.23 12.12 task sanity check. Since TASU2 simulates CTC posteriors with WER-conditioned control, its supervision is most directly aligned with token recognition and decoding, so gains beyond ASR are not necessarily expected. Still, TASU2 substantially improves over TASU on LibriSpeech without using any train- ing audio, and remains competitive on CoVoST2, suggesting that CTC-style simulated supervision can transfer modestly be- yond pure ASR under a zero-audio regime. 5.2. Domain Adaptation Section 5.1 indicates that TASU2 achieves stronger alignment and generalization. In this section, we turn to its domain adapta- tion capability, with a particular focus on low-resource settings. We study two-stage domain adaptation from a source domain to a low-resource target domain. LibriSpeech serves as the source- domain pre-training data (Stage 1), while the medical set sim- ulates a low-resource target domain for adaptation (Stage 2). We report WER on both LibriSpeech and the medical test set to jointly measure target improvement and source retention. Table 4 highlights a clear trade-off in conventional adapta- tion: target-domain gains often come with reduced source reten- tion. For SLAM-CTC, adapting with target-domain audio (raw or TTS) improves the Medical WER to 12.35/12.79, but also shifts LibriSpeech from 2.43/6.07 to around 2.70–2.72/6.76– 6.80, indicating a non-trivial source-domain drop. In contrast, TASU2 yields a more favorable low-resource transfer profile. Using only text-derived simulated CTC su- pervision for Stage 2, TASU2 achieves the best Medical WER of 12.12, outperforming text-only adaptation (13.62) and TTS augmentation (12.79), and even slightly surpassing raw- audio adaptation (12.35). Meanwhile, source-domain WER remains nearly unchanged from 2.94/7.16 to 2.96/7.23 (only +0.02/+0.07), demonstrating strong source retention. These re- sults suggest that WER-controlled posterior simulation provides an effective and stable alternative to target-audio-heavy fine- tuning when paired data is scarce. 6. Conclusion We presented TASU2, a WER-controllable text-to-CTC sim- ulator trained with distribution-level supervision. By generat- ing pseudo posteriors that better match acoustic CTC behavior, TASU2 provides stronger alignment signals for speech foun- dation models and Speech LLMs. Across evaluations, it im- proves robustness and cross-domain generalization, and enables effective source-to-target adaptation. In low-resource targets, TASU2 can outperform TTS-based augmentation, offering a practical alternative when paired audio is scarce. 7. Generative AI Use Disclosure During the preparation of this work, we used generative AI tools for assistance. The AI tools were only employed for improving the presentation, readability, and formatting of the manuscript, as well as for auxiliary support in code development and verifi- cation. They were not used to generate any substantial content, core ideas, experimental design, analysis, or conclusions of the paper. 8. References [1] J. Peng, Y. Wang, B. Li, Y. Guo, H. Wang, Y. Fang, Y. Xi, H. Li, X. Li, K. Zhang, S. Wang, and K. Yu, “A survey on speech large language models for understanding,” IEEE Journal of Selected Topics in Signal Processing, p. 1–32, 2025. [Online]. Available: http://dx.doi.org/10.1109/JSTSP.2025.3640535 [2] S. Arora, K.-W. Chang, C.-M. Chien, Y. Peng, H. Wu, Y. Adi, E. Dupoux, H.-Y. Lee, K. Livescu, and S. Watanabe, “On the landscape of spoken language models: A comprehensive survey,” 2025. [Online]. Available: https://arxiv.org/abs/2504.08528 [3] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large- scale weak supervision,” 2022. [Online]. Available:https: //arxiv.org/abs/2212.04356 [4] J. Peng, Y. Yang, X. Li, Y. Xi, Q. Tang, Y. Fang, J. Li, and K. Yu, “Tasu: Text-only alignment for speech understanding,” 2026. [Online]. Available: https://arxiv.org/abs/2511.03310 [5] R. Ma, T. Chen, K. Audhkhasi, and B. Ramabhadran, “Legoslm: Connecting llm with speech encoder using ctc posteriors,” arXiv preprint arXiv:2505.11352, 2025. [6] R. Fan, B. Ren, Y. Hu, R. Zhao, S. Liu, and J. Li, “Alignformer: Modality matching can achieve better zero-shot instruction- following speech-llm,” IEEE Journal of Selected Topics in Signal Processing, p. 1–10, 2025. [7] J. Zhao and W.-Q. Zhang, “Improving automatic speech recogni- tion performance for low-resource languages with self-supervised models,” IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, p. 1227–1241, 2022. [8] Y. Takashima, S. Horiguchi, S. Watanabe, P. Garc ́ ıa, and Y. Kawaguchi, “Updating only encoders prevents catastrophic forgetting of end-to-end asr models,” 2022. [Online]. Available: https://arxiv.org/abs/2207.00216 [9] S. Burdisso, E. Villatoro-Tello, A. Carofilis, S. Kumar, K. Hacioglu, S. Madikeri, P. Rangappa, M. K. E, P. Motlicek, S. Venkatesan, and A. Stolcke, “Text-only adaptation in llm- based asr through text denoising,” 2026. [Online]. Available: https://arxiv.org/abs/2601.20900 [10] F.-T. Liao, Y.-C. Chan, Y.-C. Chen, C.-J. Hsu, and D.-s. Shiu, “Zero-shot domain-sensitive speech recognition with prompt- conditioning fine-tuning,” in 2023 IEEE Automatic Speech Recog- nition and Understanding Workshop (ASRU).IEEE, 2023, p. 1–8. [11] Y. Fang, J. Peng, X. Li, Y. Xi, C. Zhang, G. Zhong, and K. Yu, “Low-resource domain adaptation for speech llms via text-only fine-tuning,” arXiv preprint arXiv:2506.05671, 2025. [12] E. Casanova,C. Shulby,A. Korolev,A. C. Junior, A. da Silva Soares, S. Alu ́ ısio, and M. A. Ponti, “Asr data augmentation in low-resource settings using cross-lingual multi-speaker tts and cross-lingual voice conversion,” 2023. [Online]. Available: https://arxiv.org/abs/2204.00618 [13] M. Jung, O. Kwon, S. Seo, and S. Seo, “Blank collapse: Compressing ctc emission for the faster decoding,” 2023. [Online]. Available: https://arxiv.org/abs/2210.17017 [14] K. Deng, S. Cao, Y. Zhang, and L. Ma, “Improving hybrid ctc/attention end-to-end speech recognition with pretrained acoustic and language model,” 2021. [Online]. Available: https://arxiv.org/abs/2112.07254 [15] Z. Chen, W. Deng, T. Xu, and K. Yu, “Phone synchronous decod- ing with ctc lattice.” in Interspeech, 2016, p. 1923–1927. [16] K. Deng and P. C. Woodland, “Label-synchronous neural transducer for adaptable online e2e speech recognition,” 2023. [Online]. Available: https://arxiv.org/abs/2311.11353 [17] K. An, Q. Chen, C. Deng, Z. Du, C. Gao, Z. Gao, Y. Gu, T. He, H. Hu, K. Hu, S. Ji, Y. Li, Z. Li, H. Lu, H. Luo, X. Lv, B. Ma, Z. Ma, C. Ni, C. Song, J. Shi, X. Shi, H. Wang, W. Wang, Y. Wang, Z. Xiao, Z. Yan, Y. Yang, B. Zhang, Q. Zhang, S. Zhang, N. Zhao, and S. Zheng, “Funaudiollm: Voice understanding and generation foundation models for natural interaction between humans and llms,” 2024. [Online]. Available: https://arxiv.org/abs/2407.04051 [18] Qwen, :, A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Tang, T. Xia, X. Ren, X. Ren, Y. Fan, Y. Su, Y. Zhang, Y. Wan, Y. Liu, Z. Cui, Z. Zhang, and Z. Qiu, “Qwen2.5 technical report,” 2025. [Online]. Available: https://arxiv.org/abs/2412.15115 [19] V. Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: An asr corpus based on public domain audio books,” in 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, p. 5206–5210. [20] FigureEightInc.,“Medicalspeech,transcription, andintent,”KaggleDataset,2019.[Online].Avail- able:https://w.kaggle.com/datasets/paultimothymooney/ medical-speech-transcription-and-intent [21] F. Hernandez, V. Nguyen, S. Ghannay, N. Tomashenko, and Y. Est ` eve, TED-LIUM 3: Twice as Much Data and Corpus Repartition for Experiments on Speaker Adaptation.Springer International Publishing, 2018, p. 198–208. [Online]. Available: http://dx.doi.org/10.1007/978-3-319-99579-3 21 [22] H. Wang, F. Yu, X. Shi, Y. Wang, S. Zhang, and M. Li, “Slidespeech: A large-scale slide-enriched audio-visual corpus,” 2023. [Online]. Available: https://arxiv.org/abs/2309.05396 [23] C. Wang et al., “CoVoST 2 and massively multilingual speech-to- text translation,” in Proc. Interspeech, 2021, p. 2247–2251.