Paper deep dive
Look Less, Hear Better: Jointly Rewarded GRPO for Streaming ASR
Xiuwen Zheng
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/17/2026, 4:54:21 AM
Summary
This paper introduces AWED (Aligned Word Emission Delay), a metric for measuring true user-perceived latency in streaming Automatic Speech Recognition (ASR) systems using Delayed Streams Modeling (DSM). The authors propose a post-training method using Group Relative Policy Optimization (GRPO) with a joint reward for transcription accuracy and AWED latency. Applied to the Voxtral Realtime backbone, this approach significantly reduces Word Error Rate (WER) and median AWED across various structural delay budgets without architectural changes.
Entities (7)
Relation Signals (5)
Voxtral Realtime → uses → DSM
confidence 96% · Voxtral Realtime [1] builds on DSM with a causal streaming audio encoder
AWED → measures → user-perceived latency
confidence 95% · AWED... makes the true cost of a given tau visible and differentiates systems that share the same nominal delay.
DSM → uses → tau
confidence 95% · text and audio are placed on a shared time axis, with the transcript offset from the audio by a structural delay tau
GRPO → optimizes → AWED and WER
confidence 94% · post-train a DSM recognizer with GRPO under a reward that scores transcription accuracy and measured delay jointly.
Qwen3-ForceAligner → provides → word-level timestamps
confidence 90% · Word-level timestamps... are obtained by forced alignment using Qwen3-ForceAligner
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Streaming automatic speech recognition (ASR) must be judged jointly on what it transcribes and on how quickly it commits each word. Delayed streams modeling (DSM) has become the dominant paradigm for streaming large audio-language models, exposing a structural delay $\tau$ that bounds the decoder's lookahead. We show that $\tau$ is a poor proxy for user-perceived latency, and that the alignment-based supervision of DSM leaves latency on the table: the same forced-aligned transcript is used at every $\tau$, forcing the model to withhold words it could already commit. We introduce AWED, a word-level emission-delay metric defined relative to the acoustic end of each word, and post-train a DSM recognizer with GRPO under a reward that scores transcription accuracy and measured delay jointly. Trained at a single operating point ($\tau=6$ frames), our model dominates both its supervised fine-tuning initialization and the Voxtral Realtime backbone across all evaluated lookahead budgets: it cuts WER by 30.8\% relative at an 80\,ms structural delay, and by 5.7\% relative at 480\,ms while lowering median AWED from 1.17\,s to 1.04\,s. Latency-rewarded post-training thus advances the accuracy--latency Pareto frontier of streaming ASR without architectural change.
Tags
Links
- Source: https://arxiv.org/abs/2609.18333v1
- Canonical: https://arxiv.org/abs/2609.18333v1
PDF not stored locally. Use the link above to view on the source site.
Full Text
24,106 characters extracted from source content.
Expand or collapse full text
Look Less, Hear Better: Jointly Rewarded GRPO for Streaming ASR Xiuwen Zheng Abstract Streaming automatic speech recognition (ASR) must be judged jointly on what it transcribes and on how quickly it commits each word. Delayed streams modeling (DSM) has become the dominant paradigm for streaming large audio-language models, exposing a structural delay τ that bounds the decoder’s lookahead. We show that τ is a poor proxy for user-perceived latency, and that the alignment-based supervision of DSM leaves latency on the table: the same forced-aligned transcript is used at every τ, forcing the model to withhold words it could already commit. We introduce AWED, a word-level emission-delay metric defined relative to the acoustic end of each word, and post-train a DSM recognizer with GRPO under a reward that scores transcription accuracy and measured delay jointly. Trained at a single operating point (τ=6τ=6 frames), our model dominates both its supervised fine-tuning initialization and the Voxtral Realtime backbone across all evaluated lookahead budgets: it cuts WER by 30.8% relative at an 80 ms structural delay, and by 5.7% relative at 480 ms while lowering median AWED from 1.17 s to 1.04 s. Latency-rewarded post-training thus advances the accuracy–latency Pareto frontier of streaming ASR without architectural change. Index Terms: streaming speech recognition, speech language models, reinforcement learning, emission latency †address: University of Illinois Urbana-Champaign 1 Introduction Streaming automatic speech recognition (ASR) underpins live captioning, voice assistants, meeting transcription and simultaneous interpretation. In all of these settings the quality a user perceives depends not only on whether words are transcribed correctly, but also on when they appear. Streaming ASR is therefore inherently a two-objective problem, and progress is properly judged by the accuracy–latency Pareto frontier rather than by word error rate (WER) alone. Large audio-language models (LALMs) have recently achieved highly competitive performance on ASR benchmarks [1], and delayed streams modeling (DSM) [2] has emerged as the dominant recipe for making them streaming: text and audio are placed on a shared time axis, with the transcript offset from the audio by a structural delay τ that bounds the lookahead available to the decoder. Voxtral Realtime [1] builds on DSM with a causal streaming audio encoder and conditions a single model on τ sampled from 1 to 30 frames (80–2400 ms), attaining state-of-the-art streaming accuracy across a broad range of lookahead budgets. In parallel, reinforcement learning (RL) post-training with group relative policy optimization (GRPO) [3] has proven effective for LLMs and, more recently, for speech tasks [4] — but always under accuracy-only rewards. Two gaps remain. First, latency in DSM is declared rather than measured. The structural delay τ is a nominal bound on lookahead, not the time a user waits for a word: at τ=1τ=1 the median word-level emission delay of Voxtral Realtime is 0.77 s, nearly an order of magnitude above the nominal 80 ms. The discrepancy arises because a word is committed only after its aligned end time plus τ, after any hesitation of the decoder, and after enough subsequent tokens have been emitted to disambiguate it. Existing latency measures — token emission delay for RNN-T, average lagging and its differentiable variants from simultaneous translation, or first-token latency — were not designed for the alignment-anchored output stream of DSM and do not expose this gap. What is not measured cannot be optimized. Second, alignment-supervised training pins emission to a fixed schedule. Under DSM the training signal instructs the model to emit each word exactly τ frames after its forced-aligned end time, and the same alignment is reused for every τ. When τ is large the decoder holds far more acoustic evidence than its decision requires, yet the target forbids it from committing earlier; the schedule is likewise uniform across words, ignoring that many words are unambiguous well before their acoustic offset. This slack is out of reach for supervised fine-tuning (SFT), whose objective is to imitate one predetermined schedule, and equally out of reach for a hand-designed alignment heuristic, which would require knowing the optimal per-word emission time in advance. Emission timing is discrete and non-differentiable — precisely the regime where RL applies. We bridge both gaps. We introduce AWED (aligned word emission delay), a word-level latency metric defined relative to the acoustic end of each word and reported over its full distribution, which makes the true cost of a given τ visible and differentiates systems that share the same nominal delay. We then post-train a DSM streaming recognizer with GRPO under a reward that scores transcription accuracy and measured emission delay jointly, gated by a match rate that withholds latency credit from rollouts whose WER is too high for early emission to be meaningful. Trained at a single operating point (τ=6τ=6), the resulting policy dominates both its SFT initialization and the Voxtral Realtime backbone across the entire latency spectrum, reducing WER by 30.8% relative at the tightest 80 ms budget at no latency cost. Our contributions are as follows: 1. AWED, a word-level emission-delay metric tailored to DSM, which separates true output latency from the nominal structural delay. 2. A two-stage recipe that reproduces the Voxtral Realtime SFT pipeline and post-trains it with GRPO under a joint accuracy–latency reward. 3. A consistent advance of the accuracy–latency Pareto frontier of streaming ASR, generalizing from a single training τ to all evaluated lookahead budgets. 2 Method 2.1 Delayed Streams Modeling and Supervised Baseline DSM places the audio and text streams on a shared time axis discretized at 1/Δ=12.51/ =12.5 Hz (Δ=80 =80 ms). At each frame t the decoder emits one token at∈∪[P]a_t ∪\ [P]\, where the pad symbol [P] denotes wait; emission timing is therefore learned end-to-end rather than governed by an external policy. The text stream is offset from the audio stream by a structural delay of τ frames, which bounds the lookahead available when any token is produced. Following [1], τ is injected into the decoder through AdaRMSNorm conditioning and sampled per utterance during training, so that a single model serves every operating point in the range at inference time. Given a transcript w1,…,wKw_1,…,w_K with forced-aligned end times e1,…,eKe_1,…,e_K, the supervised target is constructed by placing the tokens of wkw_k contiguously from frame tk=⌈ek/Δ⌉+τ,t_k\;=\; e_k/ +τ, (1) and filling every remaining frame with [P]. Equation (1) is the object of interest for the rest of this paper: it prescribes a single emission schedule, shifted rigidly by τ, for all words and all operating points. We additionally append τpad _pad pad frames after eKe_K before the end-of-sequence symbol; anchoring EOS directly to the final word causes the model to terminate prematurely and inflates deletion errors. This stage reproduces the DSM recipe on our own aligned corpora. It is not a contribution in itself, but it supplies both a like-for-like reference point and the initialization πref _ref for the post-training stage below. 2.2 Aligned Word Emission Delay Latency in DSM is conventionally reported as τ, which is a nominal bound on lookahead rather than the time a user waits for a word. We therefore measure emission delay directly, at word level, relative to the acoustic end of each word. Let a hypothesis consist of words w^1,…,w^J w_1,…, w_J, and let f(j)f(j) be the frame at which the final token of w^j w_j is emitted. We align reference and hypothesis by word-level edit distance and retain the set ℳM of matched pairs (k,j)(k,j). For each such pair, AWED(k,j)=f(j)Δ−ek,AWED(k,j)\;=\;f(j)\, \;-\;e_k, (2) i.e. the audio duration consumed at the moment the word is committed, minus the time at which the word actually ended in the signal. Delays are averaged within an utterance and aggregated over a corpus by quantile; we report the median (p50) as the primary figure and p90/p95 to expose the tail. Restricting (2) to ℳM is not a convenience: a deleted reference word has no emission time and an inserted hypothesis word has no reference time, so emission delay is undefined outside the matched set. This property is what makes an unconstrained latency objective exploitable, and motivates the gate introduced in Sec. 2.3. Anchoring to eke_k decouples the measurement from word duration and speaking rate, which distinguishes AWED from existing measures. Token emission delay for RNN-T is defined on that model’s alignment lattice; average lagging and its differentiable variants from simultaneous translation quantify read/write lag against input length rather than against an acoustic landmark; first-token latency observes only the onset of the stream. None of the three exposes the gap we are concerned with: at τ=1τ=1, a nominal delay of 8080 ms corresponds to a median AWED of 0.770.77 s. 2.3 Latency-Rewarded GRPO Under DSM the state at frame t is the conditioning delay together with the consumed audio and the tokens emitted so far, and the action is the token emitted next — including [P] and EOS. What to emit and when to emit it thus inhabit the same discrete action space, and emission timing can be optimized without any additional head or auxiliary decision model. Because that timing is discrete and the delay in (2) is a non-differentiable function of the sampled trajectory, we optimize it by policy gradient rather than by a modified supervised target. For each utterance we draw a group of G rollouts from the current policy at a fixed conditioning delay and score rollout i by ri=(1−WERi)−λ[mi≥m0]d~i,r_i\;=\; (1-WER_i )\;-\;λ\,1\! [m_i≥ m_0 ] d_i, (3) where mim_i is the fraction of reference words matched by rollout i, d~i d_i is the group-normalized emission delay d~i=(di−mini′di′)/(maxi′di′−mini′di′+ϵ) d_i=(d_i- _i d_i )/( _i d_i - _i d_i +ε) with did_i the mean AWED of rollout i, and λ trades accuracy against latency. Normalizing within the group places both terms on a common O(1)O(1) scale, so that λ is interpretable and stable across corpora of differing speaking rate. The indicator withholds latency credit from rollouts whose match rate falls below m0m_0: as noted in Sec. 2.2, delay is measured only on matched words, so a rollout that emits little or emits poorly would otherwise be rewarded for appearing fast. The gate is inspired by the quality-thresholded reward of HPO [5]. Rewards are converted to group-relative advantages, A^i=(ri−mean())/std() A_i=(r_i-mean(r))/std(r), which removes the need for a learned critic, and the policy is updated with the clipped GRPO objective (θ)=[1G∑i1|oi|∑tmin(ρi,tA^i,clip(ρi,t,1−ε,1+ε)A^i)]−βKL[πθ∥πref], splitJ(θ)=E [ 1G _i 1|o_i| _t (& _i,t A_i,\;clip( _i,t,1- ,1+ ) A_i ) ]\\ -\;β\,D_KL& [ _θ\,\|\, _ref ], split (4) with ρi,t=πθ(oi,t∣si,t)/πθold(oi,t∣si,t) _i,t= _θ(o_i,t s_i,t)/ _ _old(o_i,t s_i,t). We train at a single conditioning delay. Nothing in (3) references τ, so the policy is not rewarded for waiting a prescribed number of frames but for committing once the evidence suffices; since the delay condition is shared across operating points through the same AdaRMSNorm pathway, the learned behaviour is expected to transfer across the range. Sec. 4 confirms that it does. 3 Experimental Setup 3.1 Data Preparation This work uses English ASR data exclusively. We aggregate twelve publicly available corpora: YODAS [6], LibriHeavy [7], VoxPopuli [8], GigaSpeech [9], Fisher [10], Common Voice [11], LibriSpeech [12], National Speech Corpus [13], Speech Accessibility Project [14], Switchboard [15], Europarl [16], and CALLHOME [17], amounting to approximately 175k hours of raw audio. Only utterances between 1 s and 30 s in duration are retained for training. Word-level timestamps, which our training objective requires, are obtained by forced alignment using Qwen3-ForceAligner [18]. Evaluation follows the HuggingFace Open ASR Leaderboard [19], on its eight English test sets: AMI, Earnings22, GigaSpeech, LibriSpeech test-clean, LibriSpeech test-other, SPGISpeech, TED-LIUM, and VoxPopuli. Together these cover read speech, meetings, earnings calls, parliamentary and lecture speech, so that no single domain dominates the aggregate. For each set we compute WER by pooling errors and reference words over all of its utterances, and report the unweighted mean of the eight resulting values; the same hypotheses are used to compute AWED as defined in Sec. 2.2. Text normalization is performed with an in-house normalizer rather than the leaderboard default, and is applied identically to every system compared in this paper, so all reported numbers are internally consistent but not directly comparable to published leaderboard entries. 3.2 Implementation Details Our experiments build upon the Voxtral-Mini-4B-Realtime11 1 https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602 checkpoint using its Hugging Face implementation. Only the language decoder is updated via LoRA (r=α=256r=α=256), while the causal audio encoder remains frozen. The training procedure comprises two sequential stages. Stage 1 performs supervised fine-tuning for one epoch. Following Voxtral-Realtime [1], the structural streaming delay τ is uniformly sampled from 1 to 30 frames (8080–2400ms2400\,ms, where 1 frame=80ms1 frame=80\,ms) to cover a broad spectrum of lookahead configurations. To prevent premature truncation, six padding tokens are appended post-transcript, positioning the EOS token 6 frames (480ms480\,ms) beyond the reference ending. Stage 2 applies GRPO to the Stage 1 checkpoint with τ fixed at 6 frames, the optimal operating point identified in [1]. 4 Results and analysis 4.1 Main Results Table 1: Performance across structural delay settings (τ∈1,3,6,12τ∈\1,3,6,12\ frames) over eight benchmarks. Latency metrics (p50,p90,p95p_50,p_90,p_95) are reported in seconds via AWED; WER is reported as mean Micro-WER (%). Delay 80 ms (τ=1τ=1) Delay 240 ms (τ=3τ=3) Delay 480 ms (τ=6τ=6) Delay 960 ms (τ=12τ=12) Model WER p50p_50 p90p_90 p95p_95 WER p50p_50 p90p_90 p95p_95 WER p50p_50 p90p_90 p95p_95 WER p50p_50 p90p_90 p95p_95 Exp 0 (Voxtral Realtime) 12.67 0.77 1.06 1.32 9.57 0.93 1.10 1.23 8.31 1.17 1.33 1.41 7.77 1.65 1.80 1.87 Exp 1d (Stage 1 SFT) 10.91 0.75 0.89 0.95 9.14 0.91 1.04 1.11 8.24 1.15 1.28 1.35 7.75 1.63 1.76 1.85 Exp 4c (Stage 2 GRPO) 8.76 0.74 0.89 0.95 8.18 0.86 1.00 1.07 7.84 1.04 1.18 1.26 7.66 1.53 1.69 1.76 Figure 1: Accuracy-latency Pareto frontiers. Y-axis: mean Micro-WER (%); X-axis: output delay (AWED p50p_50 in seconds). Dashed lines indicate p90p_90 and p95p_95 long-tail delays. Exp 4c (Stage 2 GRPO) consistently dominates the frontiers. To evaluate the accuracy-latency trade-off, Table 1 and Fig. 1 present the quantitative performance and corresponding Pareto frontiers across four structural delay settings (τ∈1,3,6,12τ∈\1,3,6,12\ frames, where 1 frame=80ms1 frame=80\,ms). Here, recognition accuracy is reported as the mean Micro-WER across eight inference benchmarks, while latency is measured via AWED (p50p_50 serves as the primary metric, with p90p_90 and p95p_95 representing long-tail delays as indicated by dashed lines in the Pareto plots). Compared to the pretrained Voxtral-Realtime backbone (Exp 0), Stage 1 SFT (Exp 1d) achieves comparable overall performance while offering pronounced gains under tight delay constraints. At an 80ms80\,ms structural delay (τ=1τ=1), Stage 1 SFT lowers WER from 12.67%12.67\% to 10.91%10.91\% (13.9%13.9\% relative reduction) at similar latency AWED p50AWED p_50. Moreover, it substantially mitigates long-tail emission lag, reducing p90p_90 from 1.06s1.06\,s to 0.89s0.89\,s and p95p_95 from 1.32s1.32\,s to 0.95s0.95\,s, which stabilizes token emissions and provides a well-behaved checkpoint for Stage 2. Building upon the Stage 1 checkpoint, Stage 2 GRPO post-training—using our grounded hyperparameter choices of λ=0.01λ=0.01 and 5k5k steps (Exp 4c, Section 4.2)—yields substantial performance gains across the entire latency spectrum. Under the default setting (480ms480\,ms delay budget), Exp 4c brings WER down to 7.84%7.84\% (vs. 8.31%8.31\% in Exp 0) while reducing AWED p50p_50 latency to 1.04s1.04\,s (down from 1.17s1.17\,s). Crucially, despite being conditioned on a fixed τ=6τ=6 during training, these optimizations generalize effectively across the entire latency spectrum. In particular, under the 80ms80\,ms setting (τ=1τ=1), Exp 4c pushes WER down to 8.76%8.76\%—achieving relative error reductions of 19.7%19.7\% and 30.8%30.8\% relative to Exp 1d and Exp 0, respectively—while maintaining an AWED p50p_50 latency comparable to Exp 1d. 4.2 Ablation Study (a) Impact of GRPO training steps (λ=0.025λ=0.025). (b) Impact of latency reward weight λ. Figure 2: Pareto frontiers illustrating the trade-off between recognition error (Micro-Average WER across 8 benchmarks) and streaming delay (AWED p50p_50). The ablations evaluate: (a) policy convergence over GRPO training steps, and (b) model sensitivity to the latency penalty coefficient λ. This section presents systematic ablations over training duration and latency penalty weight to validate our optimal GRPO recipe (λ=0.01λ=0.01, 5k5k steps). First, regarding training duration, we evaluate checkpoints across 1k1k to 3k3k steps under λ=0.025λ=0.025. The 3k3k-step setup markedly outperforms 1k1k steps and achieves comparable trade-offs to 2k2k steps across the Pareto curve (fig. 2(a)). Extended training to 3k3k steps yields additional subtle WER gains at higher latency budgets, confirming that 3k3k–5k5k steps are sufficient for full convergence without over-fitting. Second, concerning the latency reward weight (λ), while higher penalties (λ∈0.025,0.05,0.075λ∈\0.025,0.05,0.075\) yield near-identical trade-off profiles, a milder penalty (λ=0.01λ=0.01) demonstrates distinct advantages (fig. 2(b)). Although it lags slightly at the tightest 80ms80\,ms delay budget, λ=0.01λ=0.01 unlocks noticeable accuracy gains as the delay budget expands. This indicates that a smaller weight avoids over-constraining the model’s capacity, preserving higher accuracy for inference-time delay tuning. 5 Conclusion References [1] A. H. Liu, A. Ehrenberg, A. Lo, C. Sun, G. Lample, J. Delignon, K. R. Chandu, P. von Platen, P. R. Muddireddy, R. Arora, et al. (2026) Voxtral realtime. arXiv preprint arXiv:2602.11298. Cited by: §1, §2.1, §3.2. [2] N. Zeghidour, E. Kharitonov, M. Orsini, V. Volhejn, G. de Marmiesse, E. Grave, P. Pérez, L. Mazaré, and A. Défossez (2025) Streaming sequence-to-sequence learning with delayed streams modeling. arXiv preprint arXiv:2509.08753. Cited by: §1. [3] Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. K. Li, Y. Wu, and D. Guo (2024) DeepSeekMath: pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300. Cited by: §1. [4] P. G. Shivakumar, Y. Gu, A. Gandhe, and I. Bulyko (2025) Group relative policy optimization for speech recognition. arXiv preprint arXiv:2509.01939. Cited by: §1. [5] S. Ouyang, S. Ding, O. Hrinchuk, V. Lavrukhin, B. Yan, B. Ginsburg, and L. Li (2026) Hierarchical policy optimization for simultaneous translation of unbounded speech. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 1772–1787. Cited by: §2.3. [6] X. Li, S. Takamichi, T. Saeki, W. Chen, S. Shiota, and S. Watanabe (2023) Yodas: youtube-oriented dataset for audio and speech. In 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), p. 1–8. Cited by: §3.1. [7] W. Kang, X. Yang, Z. Yao, F. Kuang, Y. Yang, L. Guo, L. Lin, and D. Povey (2024) Libriheavy: a 50,000 hours asr corpus with punctuation casing and context. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 10991–10995. Cited by: §3.1. [8] L. Lugosch, T. Likhomanenko, G. Synnaeve, and R. Collobert (2022) Pseudo-labeling for massively multilingual speech recognition. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 7687–7691. Cited by: §3.1. [9] G. Chen, S. Chai, G. Wang, J. Du, W. Zhang, C. Weng, D. Su, D. Povey, J. Trmal, J. Zhang, et al. (2021) Gigaspeech: an evolving, multi-domain asr corpus with 10,000 hours of transcribed audio. arXiv preprint arXiv:2106.06909. Cited by: §3.1. [10] C. Cieri, D. Miller, and K. Walker (2004) The fisher corpus: a resource for the next generations of speech-to-text.. In LREC, Vol. 4, p. 69–71. Cited by: §3.1. [11] R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Henretty, R. Morais, L. Saunders, F. Tyers, and G. Weber (2020) Common voice: a massively-multilingual speech corpus. In Proceedings of the twelfth language resources and evaluation conference, p. 4218–4222. Cited by: §3.1. [12] V. Panayotov, G. Chen, D. Povey, and S. Khudanpur (2015) Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), p. 5206–5210. Cited by: §3.1. [13] J. X. Koh, A. Mislan, K. Khoo, B. Ang, W. Ang, C. Ng, and Y. Tan (2019) Building the singapore english national speech corpus. In Proc. Interspeech 2019, p. 321–325. Cited by: §3.1. [14] M. Hasegawa-Johnson, X. Zheng, H. Kim, C. Mendes, M. Dickinson, E. Hege, C. Zwilling, M. M. Channell, L. Mattie, H. Hodges, et al. (2024) Community-supported shared infrastructure in support of speech accessibility. Journal of Speech, Language, and Hearing Research 67 (11), p. 4162–4175. Cited by: §3.1. [15] J. J. Godfrey, E. C. Holliman, and J. McDaniel (1992) SWITCHBOARD: telephone speech corpus for research and development. In [Proceedings] ICASSP-92: 1992 IEEE International Conference on Acoustics, Speech, and Signal Processing, Vol. 1, p. 517–520. Cited by: §3.1. [16] P. Koehn (2005) Europarl: a parallel corpus for statistical machine translation. In Proceedings of machine translation summit x: papers, p. 79–86. Cited by: §3.1. [17] A. Canavan, D. Graff, and G. Zipperlen (1997) CALLHOME American English speech LDC97S42. Linguistic Data Consortium, Philadelphia, PA, USA. Note: Web Download External Links: Document Cited by: §3.1. [18] X. Shi, X. Wang, Z. Guo, Y. Wang, P. Zhang, X. Zhang, Z. Guo, H. Hao, Y. Xi, B. Yang, et al. (2026) Qwen3-asr technical report. arXiv preprint arXiv:2601.21337. Cited by: §3.1. [19] V. Srivastav, S. Zheng, E. Bezzam, E. L. Bihan, N. R. Koluguri, P. Żelasko, S. Majumdar, A. Moumen, and S. Gandhi (2025) Open asr leaderboard: towards reproducible and transparent multilingual and long-form speech recognition evaluation. arXiv preprint arXiv:2510.06961. Cited by: §3.1.