Paper deep dive
X2Streaming-ASR: wait when uncertain, emit when ready for streaming ASR
Zhiwei Lin, Kaiqi Fu, Rime Wen, Zehan Liu, Shawn Qin, Roy Gan, Hao Wang, Qian Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 9/14/2026, 5:35:25 AM
Summary
The paper introduces X2Streaming-ASR, a streaming automatic speech recognition system that optimizes for low commit latency and high accuracy by adaptively deciding when to emit recognized text. Unlike systems with fixed delays, X2Streaming-ASR uses a three-stage training process involving supervised learning, commit-time probing, and Group Relative Policy Optimization (GRPO) to learn position-dependent wait/emit decisions. It achieves significantly lower latency (27-84 ms) compared to baselines (409-585 ms) while maintaining competitive Character Error Rates (CER) on datasets like AISHELL-1/2/3 and WenetSpeech.
Entities (14)
Relation Signals (11)
X2Streaming-ASR → achievesmetric → CER
confidence 100% · It achieves the best streaming CER among the evaluated systems on AISHELL-1 and AISHELL-3
X2Streaming-ASR → achievesmetric → Latency
confidence 100% · X2Streaming-ASR achieves a mean character-level commit latency of 27–84 ms
X2Streaming-ASR → evaluatedon → AISHELL-2
confidence 100% · Across AISHELL-1/2/3 and WenetSpeech, X2Streaming-ASR achieves a mean character-level commit latency of 27–84 ms
X2Streaming-ASR → evaluatedon → AISHELL-3
confidence 100% · It achieves the best streaming CER among the evaluated systems on AISHELL-1 and AISHELL-3
X2Streaming-ASR → evaluatedon → AISHELL-1
confidence 100% · Across AISHELL-1/2/3 and WenetSpeech, X2Streaming-ASR achieves a mean character-level commit latency of 27–84 ms
X2Streaming-ASR → evaluatedon → WenetSpeech
confidence 100% · Across AISHELL-1/2/3 and WenetSpeech, X2Streaming-ASR achieves a mean character-level commit latency of 27–84 ms
X2Streaming-ASR → outperforms → Zipformer
confidence 95% · X2Streaming-ASR has a mean character-level latency of 27–84 ms, approximately one-tenth of Zipformer... X2Streaming-ASR outperforms both Zipformer and Paraformer in both offline and streaming recognition on the AISHELL-1 and AISHELL-3 test sets.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Streaming automatic speech recognition (ASR) for real-time voice agents and full-duplex dialogue must provide accurate partial transcripts with low commit latency. Existing systems commonly use a fixed chunk size, look-ahead, or target delay, or encourage emissions near estimated acoustic boundaries. These approaches do not directly optimize how much additional context to use at each output position under a single-pass, hard-commit constraint. We propose X2Streaming-ASR, which decomposes streaming recognition into when to commit and what to commit. Its three-stage training procedure first establishes streaming recognition ability, then warm-starts the commit policy with automatically probed trajectories, and finally refines the policy using character-level, segment-assigned group-relative rewards for recognition accuracy and latency. Across AISHELL-1/2/3 and WenetSpeech, X2Streaming-ASR achieves a mean character-level commit latency of 27-84 ms relative to forced-aligned character endpoints, compared with 409-585 ms for the evaluated streaming baselines. It achieves the best streaming CER among the evaluated systems on AISHELL-1 and AISHELL-3 with substantially lower latency.
Tags
Links
- Source: https://arxiv.org/abs/2609.08672v1
- Canonical: https://arxiv.org/abs/2609.08672v1
Trouble viewing inline? Open PDF directly →
Full Text
29,582 characters extracted from source content.
Expand or collapse full text
X2Streaming-ASR: Wait When Uncertain, Emit When Ready for Streaming ASR Zhiwei Lin Kaiqi Fu Rime Wen Zehan Liu Shawn Qin Roy Gan Hao Wang Qian Wang Abstract Streaming automatic speech recognition (ASR) for real-time voice agents and full-duplex dialogue must provide accurate partial transcripts with low commit latency. Existing systems commonly use a fixed chunk size, look-ahead, or target delay, or encourage emissions near estimated acoustic boundaries. These approaches do not directly optimize how much additional context to use at each output position under a single-pass, hard-commit constraint. We propose X2Streaming-ASR, which decomposes streaming recognition into when to commit and what to commit. Its three-stage training procedure first establishes streaming recognition ability, then warm-starts the commit policy with automatically probed trajectories, and finally refines the policy using character-level, segment-assigned group-relative rewards for recognition accuracy and latency. Across AISHELL-1/2/3 and WenetSpeech, X2Streaming-ASR achieves a mean character-level commit latency of 2727–84ms84\,ms relative to forced-aligned character endpoints, compared with 409409–585ms585\,ms for the evaluated streaming baselines. It achieves the best streaming CER among the evaluated systems on AISHELL-1 and AISHELL-3 with substantially lower latency. Index Terms: streaming automatic speech recognition, low-latency ASR, reinforcement learning †address: X Square Robot linzhiwei,wanghao@x2robot.com 1 Introduction Streaming speech recognition (ASR) requires the system to output corresponding text in real time during speech input, which is the foundational capability for real-time captioning and voice interaction. The main metrics for evaluating streaming ASR systems are recognition accuracy (CER/WER) and latency. This paper focuses on the character-level emission latency of relatively forced alignment: how long it takes from the end of the acoustic boundary of a character to the system outputting it, which measures how long each character waits after the acoustic boundary is reached. For cascaded voice agents, the traditional process requires waiting for offline transcription before starting downstream services. Some works [23] trigger downstream services on streaming identification prefixes, effectively reducing the overall system response time. However, incorrect recognition results will compromise subsequent services, and recognition that arrives too late will delay service startup. Furthermore, recent work on turn-taking introduces streaming ASR so that decisions can use semantic information, not just acoustic information[16, 4]. Therefore, the latency and accuracy of streaming ASR directly determine whether downstream services and turn-taking can be started as early and correctly as possible. Existing work still struggles to jointly optimize accuracy and emission latency, and falls into three categories. The first approach uses global configuration for streaming recognition, such as fixed chunk decoding and look-ahead or delay[17, 5, 1, 22, 7, 20]. Chunking specifies the decoding unit, while look-ahead and delay determine the amount of future context that is visible. Neither of these methods determines how much future information to wait for based on historical context and current acoustic information. The second approach shifts the emission time during training. FastEmit [19] encourages earlier output, whereas minLT [6, 13, 14] pulls emissions toward the alignment boundary. Unlike a global delay, they do not wait longer where upcoming context is needed for disambiguation. The third[9, 8, 11, 15] emits an unstable or partial hypothesis and later revises it. These designs miss the same fact: the amount of future context needed after acoustics boundary is position-dependent. A position-agnostic knob cannot implement this position-dependent waiting: setting it later delays every position, while setting it earlier exposes every position to the same risk. Treating “emit at the boundary” as the training objective forbids additional wait on hard characters. Revision-based methods reduce latency by correcting earlier output, and are therefore not single-pass hard commit. In this paper, we consider the setting where each character is emitted only once, with no second-pass correction. Streaming ASR needs to make two decisions: when to commit and what to commit. Supervision learning already handles the second well. The hard one is the first, which depends on acoustics and historical context, and fitting emit times with a supervised target works poorly. On the data side, the first time a character can be committed is not always its acoustic boundary, so aligned tendt_end is only a proxy for the emit target, not a unique gold time. Teacher-probed “earliest correct” times are also imperfect: they inherit the teacher’s language priors and the probe protocol. On the training side, the cross-entropy loss only fits these labeled times; latency and accuracy are not in the objective. Waiting slightly longer to recognize the character correctly, or emitting correctly earlier than the label, may help the real decision, but both count as deviations from this loss and are penalized. Reinforcement learning closes both gaps. First, it does not need pre-labeled recognizable points: multiple decision trajectories are sampled from the same audio, so the same character is committed at different times. The consequences of waiting and of committing early appear directly in the reward. Second, it optimizes the true objective: the reward combines the error count and latency, so the beneficial deviations that supervised learning would penalize are rewarded here. We therefore propose X2Streaming-ASR, trained in three stages so that the model waits only where leftover context is still required. Stage 1 trains a streaming ASR that later serves as both a probe and the initialization. Stage 2 uses this model to probe a plausible commit time for each character and applies a supervised warm-start; this preserves recognition quality and yields an initial commit policy, but the probed times are not treated as ground truth. Stage 3 refines the commit policy with Group Relative Policy Optimization (GRPO). The reward combines errors and latency, with errors taking priority, so the model waits at ambiguities and commits as soon as the evidence is sufficient. In summary, our contributions are as follows: • We propose X2Streaming-ASR, which replaces global fixed delay with position-dependent commit decisions for adaptive character commitment. • We introduce a three-stage training framework: supervised learning first establishes the ASR capability, commit-time probing provides a warm start for the commit policy, and character-level GRPO with commit-segment credit assignment directly optimizes recognition accuracy and emission latency. • Across five test sets, X2Streaming-ASR achieves 27–84 ms mean latency, reducing mean latency by 79–95% compared with streaming baselines, while achieving the best CER on AISHELL-1 and AISHELL-3. 2 Method Figure 1: The X2Streaming-ASR model architecture alternates between listen and decode states. It’s important to note that while the language model is decoding the previous audio segment, the causal audio encoder is simultaneously encoding the current audio; these two processes operate asynchronously. 2.1 Architecture As shown in Fig 1, X2Streaming-ASR is built on Voxtral Realtime, which consists of a causal audio encoder, an adapter, and a decoder-only language model. The difference is that Voxtral Realtime uses a global delay τ to conditionalize the language model, and adds text and audio tokens at the same sequence position. X2Streaming-ASR instead allows the model to adaptively decide whether to wait or commit, no longer using τ as a conditional input. Given a 16kHz waveform, we extract a log-Mel spectrogram and map it to continuous audio tokens 1:La_1:L in the text embedding space. The audio tokens have a frame rate of 12.5Hz, corresponding to 80ms of audio, which are interleaved with other tokens. During streaming recognition, the language model alternates between listening and Decoding states. While Listening, the language model takes the historical interleaved sequence z(t)z^(t) and the current audio token ata_t as input, where z(t)z^(t) consists a1:ta_1:t, previously recognized text tokens and special tokens e. The language model then makes a binary decision: ct∼pθ(c|z(t),at),c∈w,e.c_t p_θ(c|z^(t),a_t), c∈\w,e\. (1) Here w means ”wait”, which is not written back; e means ”emit”. If ct=wc_t=w, the language model stays in listening and reads next audio token at+1a_t+1. If ct=ec_t=e, the language model will write e back to the input as the starting point for recognition, stop reading new audio token, and perform autoregressive generates text token yky_k utils ⟨Eos⟩ token, where yk∼pθ(y|z(t),at,e).y_k p_θ(y|z^(t),a_t,e). (2) Then X2Streaming-ASR return to listen state and ⟨Eos⟩ token is not written back into input. 2.2 Training Strategy To enable X2Streaming-ASR to both determine when to commit and what to commit, we design a three-stage training strategy. The first stage trains a streaming recognition model: given incremental audio, it only recognizes the content where the acoustics have ended. For example, if the input contains acoustic information of two and a half characters, only the first two characters will be recognized. It serves as the initial weights and labeled model for the second stage of training. Specifically, we use Qwen3-ForcedAligner[10] to force alignment of the reference text, obtain the start and end times of each character, and then randomly cut the aligned character sequence into continuous blocks with length L∈1,..,6L∈\1,..,6\. The commit time temit_emit of each block is the end time of the last character tendt_end of block plus Δt∈0ms, 80ms, 160ms t∈\0\,ms,\,80\,ms,\,160\,ms\. If there is a next character, the cut point does not exceed the first half of its duration dnextd_next: temit=min(tend+Δt,tstartnext+12dnext).t_emit= (t_end+ t,\;t^next_start+ 12\,d_next ). (3) In the first stage, we mask the loss at the w and e positions and only train the model’s ASR capability. We then label each reference character’s emit time from left to right with the first-stage model. Probing starts at the end time of the first character. At time t, the model greedily decodes the incremental audio. Let y1:ky_1:k be the longest reference prefix that matches the decoding result. These k characters share temit=t_emit=t. For the next mismatched reference character yk+1y_k+1, if t<tend(k+1)t<t^(k+1)_end, the next probe starts at tend(k+1)t^(k+1)_end; if t has already passed tend(k+1)t^(k+1)_end, probing continues at t+80mst+80ms and never moves backward. If the exploration ends but the submission time of the complete reference text cannot be obtained, discard the sentence. An utterance is discarded if the probe ends but the commit time of the complete reference text can not be obtained, or if any character has temit−tend>640mst_emit-t_end>640ms. Labels with such a long wait after the acoustic endpoint teach the model to keep choosing w, and we observed that including them to train can stop the model from triggering e. Stage 2 initializes from the first-stage weights and trains recognition and init commit policy jointly by next-token supervision on these probed temit_emit labels. During the first and second phases of training, we mixed in offline full-text recognition at a ratio of 0.20 to enable X2Streaming-ASR to have offline recognition capabilities. The third stage refines commit policy with group-relative policy optimization (GRPO). We sample a group of wait–emit trajectories τkk=1K\ _k\_k=1^K at the sentence level. The action at each frame is ct∈w,ec_t∈\w,e\, drawn from π(⋅∣z(t),at)π(\,· z^(t),a_t). If ct=ec_t=e, the model decodes greedily until ⟨Eos⟩ . A commit segment is an e together with the consecutive w that precede it, so a trajectory naturally splits into several commit segments. However, the sampling unit should not be confused with the credit unit. If sentence-level reward are broadcast to every frame-level decision, it is unclear to determine which step caused the inter-sentence discrepancy. When accuracy is prioritized, a more accurate trajectory, though slower, yields a higher sentence-level reward. So broadcasting would reinforce useful waits and superfluous waits equally. The policy would then learn to wait throughout the utterance, rather than to wait only where it must and to emit where it can. Therefore we keep the group-relative comparison on the same reference character and send the advantage back to the corresponding commit segment. Concretely, we align each hypothesis to the forced-aligned reference y1:Jy_1:J and score character yjy_j on trajectory k by sk,j=−λdk,j,correct alignment,−1,substitution or deletion.s_k,j= cases-λ\,d_k,j,&correct alignment,\\ -1,&substitution or deletion. cases (4) The latency is normalized as dk,j=min(max(ℓk,j,0),H)H+1∈[0,1),d_k,j= ( ( _k,j,0),H )H+1∈[0,1), (5) where ℓk,j=temit(k,j)−tend(j) _k,j=t_emit^(k,j)-t_end^(j), λ is a latency weight, and H is a truncation cap. Then subtracting the group mean on the same character gives Ak,j=sk,j−s¯j,s¯j=1K∑k=1Ksk,j.A_k,j=s_k,j- s_j, s_j= 1K _k=1^Ks_k,j. (6) Each Ak,jA_k,j is then accumulated onto the commit segment that produced or missed that character. Since the insertion has no reference character, the error penalty is recorded on the commit segment that produced it, and zero-sum correction is performed within the group. Let the m-th commit segment of trajectory k be Sk,mS_k,m , the set of reference characters aligned to it be k,mJ_k,m, and the insertion penalty be Ak,minsA^ins_k,m. Then the advantage of Sk,mS_k,m is A^k,m=∑j∈k,mAk,j+Ak,mins. A_k,m= _j _k,mA_k,j+A^ins_k,m. (7) Every decision in the segment shares A^k,m A_k,m. Let the commit segment that contains decision step t be m(t)m(t). Each decision step t inherits the advantage A^k,m(t) A_k,m(t) from Sk,m(t)S_k,m(t). We then optimize the current policy πθ _θ against the sampling policy πold _old. The importance ratio ρt=πθ(ct∣z(t))πold(ct∣z(t)) _t= _θ(c_t z^(t)) _old(c_t z^(t)) (8) measures the relative change in the probability of the action between the policy. At decision step t we maximize ℒ=[min(ρtA^k,m(t),clip(ρt,1−ε,1+ε)A^k,m(t))],L=E [ ( _t A_k,m(t),\;clip( _t,1- ,1+ )\, A_k,m(t) ) ], (9) which strengthens or suppresses the corresponding w/ew/e actions according to the segment advantage, while keeping ρt _t near 11 so that a single update cannot move too far. 3 Experiments Table 1: Corpus CER (%) and character-level mean and P95 latency (ms). X2Streaming-ASR and Zipformer use the same checkpoint for streaming and offline recognition; Paraformer uses a separate offline model. Uni-ASR streaming is the 1000ms1000\,ms chunk, beam-33 result from the paper; MoCha-ASR numbers are from the paper. Neither reports leftover-wait latency, so those entries are –. Best / second-best among streaming systems are bold / underlined. Best / second-best among offline systems are ⋆ / †. X2Streaming-ASR Zipformer Paraformer Uni-ASR MoCha-ASR Streaming Off. Streaming Off. Streaming Off. Streaming Off. Streaming Off. Test set CER Mean lat. P95 CER CER Mean lat. P95 CER CER Mean lat. P95 CER CER CER CER CER AISHELL-1 1.96 27 240 0.88⋆ 1.97 472 800 1.39† 3.06 585 880 2.29 2.15 1.44 5.1 4.9 AISHELL-2 4.93 54 240 3.52 4.16 450 800 3.24 3.91 568 840 3.02† 3.25 2.60⋆ 5.5 5.0 AISHELL-3 2.68 44 240 1.61⋆ 2.94 464 720 2.17† 3.60 576 849 2.76 – – – – WenetSpeech Meeting 10.73 61 320 7.87 7.87 435 800 6.24⋆ 10.05 582 880 6.98 8.04 6.32† – – WenetSpeech Net 9.92 84 320 8.57 8.67 409 720 7.19 8.54 563 880 6.65† 6.44 5.78⋆ – – Table 2: AISHELL-1 listen-policy ablation. CER (%) and character-level latency (ms). Model Mode CER Mean lat. P95 Proposed Streaming 1.96 27 240 Stage 2 Streaming 5.83 11 80 Stage 2 Forced tendt_end 7.58 0.7 0 Stage 3 + content KL Streaming 2.58 33 160 Stage 3 + content KL Forced tendt_end 7.56 0.7 0 Sentence-level reward Streaming 1.38 383 1120 Table 3: AISHELL-1 under forced post-boundary delay, natural wait/emit, and offline decoding with same checkpoint. A single emit may cover later characters, which then do not receive their own Δt t. Inference CER(%) Mean lat. P95 Forced tendt_end 6.90 0.6 80 Forced tendt_end+80ms+80\,ms 3.40 88 160 Forced tendt_end+160ms+160\,ms 3.84 147 240 Streaming 1.96 27 240 3.1 Setup We train Stage 1 on AISHELL-1[2], AISHELL-2[3], AISHELL-3[12], AliMeeting[18], and WenetSpeech[21]. Stage 2 needs commit-time labels. Annotating the full WenetSpeech training set is too costly, so we probe only AISHELL-1/2/3 and the WenetSpeech L and S subsets. Stage 3 uses the same data, so that supervised and RL listen policies are compared on the same labeled pool. Stage 3 rewards need only the reference characters and their tendt_end, and do not require extra emit-time labels. Evaluation is on the AISHELL-1/2/3 test sets and the WenetSpeech meeting and net tests. For each test utterance, Qwen3-ForceAligner gives character end times tendt_end; leftover wait is ℓ=temit−tend =t_emit-t_end. We report mean latency and P95 relative to the acoustic boundary, and CER for recognition. We used Zipformer and Paraformer-online as open-sourced streaming baselines, evaluated under their public streaming configurations. MoCha-ASR [14] and Uni-ASR [15] are not open-sourced; we quote CER from their papers and cannot measure leftover-wait latency. Uni-ASR streaming uses the paper’s 1000ms1000\,ms chunk result. In Stage 3, K=8K=8 trajectories are sampled per utterance at temperature 1.41.4, with λ=0.3λ=0.3, H=2000msH=2000\,ms, and ε=0.2 =0.2. Stages 1 and 2 train all parameters. Stage 3 freezes the encoder and adapter and fine-tunes the language model with LoRA. 3.2 Comparison with Baseline Table 1 compares X2Streaming-ASR with different models. Across all test sets, X2Streaming-ASR has a mean character-level latency of 2727–84ms84\,ms, approximately one-tenth of Zipformer and Paraformer-online. The results on P95 indicate that the latency of X2Streaming-ASR is acceptable even in a few extreme cases, while the extreme latency of these two baselines is approaching 1 second. The recognition results show a positive correlation between offline and streaming recognition performance: offline results represent the upper limit of the model’s recognition ability under sufficient context, while streaming recognition performance degrades due to the lack of complete audio. It’s important to note that X2Streaming-ASR and Zipformer use the same checkpoint for both streaming and offline recognition, while the offline and streaming results of Paraformer are derived from the parameter of the corresponding version. MoCha-ASR and Uni-ASR are likewise quoted from the corresponding paper versions; Uni-ASR streaming uses the 1000ms1000\,ms chunk setting. Specifically, X2Streaming-ASR outperforms both Zipformer and Paraformer in both offline and streaming recognition on the AISHELL-1 and AISHELL-3 test sets. On the AISHELL-2 test set and the WenetSpeech net test set, Uni-ASR performs best, followed by Paraformer and Zipformer. On the WenetSpeech meeting test set, Zipformer performs best, followed by Uni-ASR and Paraformer. This is mainly because the X2Streaming-ASR’s upper limit of recognition ability is weaker than the baseline, further leading to similar streaming results. Nevertheless, X2Streaming-ASR still maintains an order-of-magnitude latency advantage. 3.3 Global Post-Boundary Delay Figure 2: AISHELL-1 CER vs. mean character-level latency. The polyline is Forced emit at tend+Δt_end+ t on the Stage-3 checkpoint. The dashed line is the same checkpoint with full-utterance decoding. To verify whether X2Streaming-ASR has learned a global wait strategy, we compared the impact of different latency levels on WER on the AISHELL-1 test set with the same Stage-3 checkpoint. Specifically, for reference characters that have not yet been recognized, we forced the model to begin greedy decoding when temit=tend+Δt_emit=t_end+ t, where Δt∈0,80,160ms t∈\0,80,160\ms; note that when Δt=160ms t=160ms, temit_emit may overwrite the acoustic information of the next character, thus decoding multiple characters at once. Table 3 shows that committing immediately at the end of the aligned word is the fastest (5.5ms5.5\,ms), but the CER is as high as 6.906.90. Increasing the global wait to 80ms80\,ms reduced the CER to 3.403.40, and both the latency and WER were worse than X2Streaming-ASR’s natural online inference. The results in Fig 2 also indicate that X2Streaming-ASR does not choose a compromise point on a fixed delay curve, but waits when uncertain and makes a decision when certain. A further increase to 160ms160\,ms does not help (3.843.84) and raises mean latency to 147ms147\,ms. We attribute this mainly to the longer wait makes it easier to bring up the next character, allowing it to be decoded even in the absence of subsequent acoustic information. To verify the impact of supervised learning on the submission decision, we evaluated the performance of the Stage 2 model on the AISHELL-1 test set. As shown in Table 2 and Fig 2, the results of Stage 2 fall near the poly line, close to the results of force tendt_end, indicating that supervised learning tends to be global early. To verify the benefits of commit policy, we trained a Stage 3 model with content-based KL divergence, anchoring the model’s recognition ability to near that of the Stage 2. The results in rows 3 and 5 of Table 2 show that adding content-based KL divergence prevents the model’s recognition ability from drifting. The results in rows 1, 2 and 4 indicate that optimizing commit policy is the main source of CER and latency optimization. Finally, to compare different GRPO strategies, we implemented a reinforcement learning version with sentence-level rewards. The results show that using sentence-level rewards for optimization leads to global waiting, sacrificing waiting latency for improved recognition performance. 4 Conclusion This paper proposes X2Streaming-ASR, which decomposes streaming speech recognition into when to commit and what to submit. We use multi-stage training to optimize recognition capability and commit policy separately, enabling the model to wait when uncertain and commit when certain. Experiments show that the average latency is reduced from hundreds of milliseconds to tens of milliseconds. In the future, we will try our method in more languages and consider more real-world scenarios to improve the model’s robustness and recognition performance. References [1] K. An, H. Zheng, Z. Ou, H. Xiang, K. Ding, and G. Wan (2022) CUSIDE: chunking, simulating future context and decoding for streaming asr. arXiv preprint arXiv:2203.16758. Cited by: §1. [2] H. Bu, J. Du, X. Na, B. Wu, and H. Zheng (2017) Aishell-1: an open-source mandarin speech corpus and a speech recognition baseline. In 2017 20th conference of the oriental chapter of the international coordinating committee on speech databases and speech I/O systems and assessment (O-COCOSDA), p. 1–5. Cited by: §3.1. [3] J. Du, X. Na, X. Liu, and H. Bu (2018) Aishell-2: transforming mandarin asr research into industrial scale. arXiv preprint arXiv:1808.10583. Cited by: §3.1. [4] K. Fu, R. Wen, A. Lin, S. Qin, R. Gan, H. Wang, and Q. Wang (2026) X2-turn: frame-synchronous dual-head modeling for joint streaming asr and turn state prediction. arXiv preprint arXiv:2608.10878. Cited by: §1. [5] Z. Gao, S. Zhang, I. McLoughlin, and Z. Yan (2022) Paraformer: fast and accurate parallel transformer for non-autoregressive end-to-end speech recognition. arXiv preprint arXiv:2206.08317. Cited by: §1. [6] H. Inaguma, Y. Gaur, L. Lu, J. Li, and Y. Gong (2020) Minimum latency training strategies for streaming sequence-to-sequence asr. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 6064–6068. Cited by: §1. [7] 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. [8] D. Liu, G. Spanakis, and J. Niehues (2020) Low-latency sequence-to-sequence speech recognition and translation by partial hypothesis selection. arXiv preprint arXiv:2005.11185. Cited by: §1. [9] D. Macháček, R. Dabre, and O. Bojar (2023) Turning whisper into real-time transcription system. In Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics: System Demonstrations, p. 17–24. Cited by: §1. [10] B. Mu, X. Shi, X. Wang, H. Liu, J. Xu, and L. Xie (2026) LLM-forcedaligner: a non-autoregressive and accurate llm-based forced aligner for multilingual and long-form speech. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 25627–25638. Cited by: §2.2. [11] 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: §1. [12] Y. Shi, H. Bu, X. Xu, S. Zhang, and M. Li (2020) Aishell-3: a multi-speaker mandarin tts corpus and the baselines. arXiv preprint arXiv:2010.11567. Cited by: §3.1. [13] Y. Shinohara and S. Watanabe (2022) Minimum latency training of sequence transducers for streaming end-to-end speech recognition. arXiv preprint arXiv:2211.02333. Cited by: §1. [14] G. Wan, W. Zhang, J. Zhang, S. Xiong, J. Gao, and Z. Ye (2026) Streaming speech recognition with decoder-only large language models and latency optimization. In ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 16367–16371. Cited by: §1, §3.1. [15] Y. Xia, J. Tang, J. Hou, G. Xu, and H. Yao (2026) Uni-asr: unified llm-based architecture for non-streaming and streaming automatic speech recognition. arXiv preprint arXiv:2603.11123. Cited by: §1, §3.1. [16] R. Yan, W. Chen, Z. Liu, Z. Ma, H. Lin, H. Wen, H. Xie, J. Wu, Y. Liang, Y. Zhao, et al. (2026) Soulx-duplug: plug-and-play streaming state prediction module for realtime full-duplex speech conversation. arXiv preprint arXiv:2603.14877. Cited by: §1. [17] Z. Yao, L. Guo, X. Yang, W. Kang, F. Kuang, Y. Yang, Z. Jin, L. Lin, and D. Povey (2024) Zipformer: a faster and better encoder for automatic speech recognition. In International Conference on Learning Representations, Vol. 2024, p. 44440–44455. Cited by: §1. [18] F. Yu, S. Zhang, Y. Fu, L. Xie, S. Zheng, Z. Du, W. Huang, P. Guo, Z. Yan, B. Ma, X. Xu, and H. Bu (2022) M2MeT: the ICASSP 2022 multi-channel multi-party meeting transcription challenge. In Proc. ICASSP, Cited by: §3.1. [19] J. Yu, C. Chiu, B. Li, S. Chang, T. N. Sainath, Y. He, A. Narayanan, W. Han, A. Gulati, Y. Wu, et al. (2021) Fastemit: low-latency streaming asr with sequence-level emission regularization. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 6004–6008. Cited by: §1. [20] 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. [21] B. Zhang, H. Lv, P. Guo, Q. Shao, C. Yang, L. Xie, X. Xu, H. Bu, X. Chen, C. Zeng, et al. (2022) Wenetspeech: a 10000+ hours multi-domain mandarin corpus for speech recognition. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 6182–6186. Cited by: §3.1. [22] W. Zhao, Z. Li, C. Yu, and Z. Ou (2024) Cuside-t: chunking, simulating future and decoding for transducer based streaming asr. In 2024 IEEE 14th International Symposium on Chinese Spoken Language Processing (ISCSLP), p. 11–15. Cited by: §1. [23] W. Zou, Y. Miao, Z. Ma, J. Xu, J. Gao, J. Hao, R. He, and J. Xu (2026) LTS-voiceagent: a listen-think-speak framework for efficient streaming voice interaction via semantic triggering and incremental reasoning. arXiv preprint arXiv:2601.19952. Cited by: §1.