Paper deep dive
DuplexCascade: Full-Duplex Speech-to-Speech Dialogue with VAD-Free Cascaded ASR-LLM-TTS Pipeline and Micro-Turn Optimization
Jianing Yang, Yusuke Fujita, Yui Sudo
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 97%
Last extracted: 3/13/2026, 12:59:47 AM
Summary
DuplexCascade is a VAD-free cascaded streaming pipeline for full-duplex speech-to-speech dialogue. It replaces traditional VAD-based segmentation with a chunk-wise micro-turn interaction model, utilizing conversational special tokens to control LLM behavior and turn-taking. By applying LoRA fine-tuning to a text-based LLM backbone, it achieves state-of-the-art full-duplex performance while preserving conversational intelligence.
Entities (5)
Relation Signals (3)
DuplexCascade → evaluatedon → Full-DuplexBench
confidence 100% · We evaluate DuplexCascade on Full-Duplex-Bench [14] against other state-of-the-art models.
DuplexCascade → usestechnique → LoRA
confidence 100% · perform lightweight LoRA adaptation [13] for 5k steps
DuplexCascade → usesmodel → Qwen2-7B-Instruct
confidence 95% · Following Freeze-Omni, we adopt Qwen2-7B-Instruct as the text LLM backbone
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Spoken dialog systems with cascaded ASR-LLM-TTS modules retain strong LLM intelligence, but VAD segmentation often forces half-duplex turns and brittle control. On the other hand, VAD-free end-to-end model support full-duplex interaction but is hard to maintain conversational intelligence. In this paper, we present DuplexCascade, a VAD-free cascaded streaming pipeline for full-duplex speech-to-speech dialogue. Our key idea is to convert conventional utterance-wise long turns into chunk-wise micro-turn interactions, enabling rapid bidirectional exchange while preserving the strengths of a capable text LLM. To reliably coordinate turn-taking and response timing, we introduce a set of conversational special control tokens that steer the LLM's behavior under streaming constraints. On Full-DuplexBench and VoiceBench, DuplexCascade delivers state-of-the-art full-duplex turn-taking and strong conversational intelligence among open-source speech-to-speech dialogue systems.
Tags
Links
- Source: https://arxiv.org/abs/2603.09180v1
- Canonical: https://arxiv.org/abs/2603.09180v1
Trouble viewing inline? Open PDF directly →
Full Text
25,630 characters extracted from source content.
Expand or collapse full text
DuplexCascade: Full-Duplex Speech-to-Speech Dialogue with VAD-Free Cascaded ASR–LLM–TTS Pipeline and Micro-Turn Optimization Jianing Yang 1,2 , Yusuke Fujita 1 , Yui Sudo 1 1 SB Intuitions Corp., Japan, 2 The University of Tokyo, Japan baleyang@g.ecc.u-tokyo.ac.jp, yusuke.fujita@sbintuitions.co.jp, yui.sudo@sbintuitions.co.jp Abstract Spoken dialog systems with cascaded ASR–LLM–TTS mod- ules retain strong LLM’s intelligence, but VAD segmentation often forces half-duplex turns and brittle control. On the other hand, VAD-free end-to-end model supports full-duplex inter- action but is hard to maintain conversational intelligence. In this paper, we present DuplexCascade, a VAD-free cascaded streaming pipeline for full-duplex speech-to-speech dialogue. Our key idea is to convert conventional utterance-wise long turns into chunk-wise micro-turn interactions, enabling rapid bidirectional exchange while preserving the strengths of a capa- ble text LLM. To reliably coordinate turn-taking and response timing, we introduce a set of conversational special control to- kens that steer the LLM’s behavior under streaming constraints. On Full-Duplex-Bench and VoiceBench, DuplexCascade deliv- ers state-of-the-art full-duplex turn-taking and strong conver- sational intelligence among open-source speech-to-speech dia- logue systems. Index Terms: Full-Duplex Dialogue; Speech-To-Speech Sys- tems; Cascaded Architectures; Turn-Taking Control 1. Introduction Recent advances in large language models (LLMs) have signif- icantly improved the quality and usability of spoken dialogue systems, bringing speech-to-speech assistants closer to natural, low-friction interaction [1]. In practice, most deployed systems adopt a cascaded design in which automatic speech recognition (ASR) transcribes user speech [2], an LLM performs dialogue reasoning in the text domain, and text-to-speech (TTS) synthe- sizes the system response [3, 4]. This modular architecture re- mains attractive because it is straightforward to engineer and iterate on, and it inherits the strong instruction-following and reasoning capabilities of modern text LLMs. Despite their strong conversational intelligence, cascaded systems typically depend on an external voice activity detec- tor (VAD) to segment user speech into turns. This reliance often forces a half-duplex “listen-then-speak” interaction style and makes turn control brittle under pauses, overlaps, and noise, leading to unnatural turn-taking behaviors such as miss- ing backchannels or failing to handle mid-utterance interrup- tions [5–8]. In contrast, recent end-to-end (E2E) speech–text models aim to remove explicit VAD-based segmentation and directly support concurrent listening and speaking, enabling more nat- ural full-duplex turn control [9–11]. However, this VAD-free duplex capability often comes with a practical downside: com- pared with strong text LLMs, conversational intelligence can degrade due to the difficulty of jointly learning robust cross- modal representations and dialogue policies [12]. In this work, we propose DuplexCascade 1 , a cascaded streaming pipeline that achieves full-duplex speech-to-speech dialogue while preserving the intelligence benefits of an LLM- centered design. The key idea is to transform conventional utterance-wise long turns into chunk-wise micro-turn interac- tions, enabling more natural full-duplex dialogue. We further introduce a carefully designed set of conversational special to- kens to explicitly regulate LLM behaviors critical for duplex interaction, enabling VAD-free turn-taking control and mak- ing the overall policy more stable and controllable. With only 50k multi-turn text dialogues, we perform lightweight LoRA adaptation [13] for 5k steps, achieving strong full-duplex per- formance on Full-Duplex-Bench [14]. Moreover, the text-only adaptation avoids cross-modal alignment issues, yielding strong conversational intelligence performace on VoiceBench [12]. 2. Related Works 2.1. VAD-Based Turn Taking Control Most recent speech-to-speech dialogue systems in practice fol- low a cascaded design with ASR, LLM, and TTS compo- nents, and rely on an external VAD to decide turn bound- aries [6, 15, 16]. However, VAD-based endpointing is funda- mentally limited because it has difficulty leveraging conversa- tional semantics and intent, which are often crucial for making precise turn-taking decisions [7, 17]. As a result, cascaded sys- tems often suffer from unstable turn-taking behaviors: they may interrupt the user when a pause is semantically meaningful, or remain silent when an immediate response or backchannel is ex- pected [8, 18, 19]. Recent cascade-style systems such as Freeze- Omni [5] inherit this difficulty, and VAD errors frequently be- come the dominant failure mode for turn-taking quality. 2.2. VAD-Free Time-Division-Multiplexing Approach To- ward Full-Duplex Interaction Beyond VAD-driven cascades, recent work has explored mak- ing LLM-centered systems behave in a full-duplex manner by restructuring interaction into short, alternating segments with- out using external VAD [20–22]. MiniCPM-Duplex [20] in- troduces the concept of Time-Division-Multiplexing encoding- decoding, which divides an interaction into time slices and pro- cesses each input slice immediately to produce the correspond- ing output slice. While this formulation enables pseudo full- duplex interaction after fine-tuning, it does not explicitly model or control turn-taking decisions; consequently, performance can degrade in realistic dialogue settings. 1 https://sbintuitions.github.io/ DuplexCascadeDemo arXiv:2603.09180v1 [cs.CL] 10 Mar 2026 StreamingASR Buffer: send a sequence of words received during the latest ∆푡 seconds Streaming TTS ∆푡=0.6푠 <User> What is the capital <EOS> ∆푡=0.6푠 ... <User> of Japan <EOS> (Micro-turn history) <System> <System Backchannel> <EOS> <System> <User finish speaking> It is Tokyo <EOS> <System> <System Backchannel> <EOS> <System> <User finish speaking> It is Tokyo <EOS> System Audio LLM finetuned with Conversational-style Micro-Turn Sequence User Audio Japan of capitalthe is What uh-huh It is Tokyo Conversational Special Tokens: <System Backchannel> <User is speaking> <User finish speaking> <User isinterrupting> ... Figure 1: Overview of DuplexCascade. User audio is transcribed by a streaming ASR and periodically flushed into text micro-turns (∆t = 0.6 s). The LLM consumes the dialogue history and the latest micro-turn to generate the next system micro-turn together with conversational special tokens (e.g., wait, respond, or backchannel). The generated text is then synthesized by a streaming TTS to produce system audio, enabling full-duplex interaction. 3. DuplexCascade 3.1. Overview Fig. 1 illustrates the overall proposed streaming pipeline. Un- like conventional cascaded systems, the proposed pipeline does not rely on VAD. Instead, user audio is continuously fed into a streaming ASR module to produce partial text outputs in real time. The partial text outputs are periodically, e.g. every 0.6 s, aggregated into a “micro-turn” and sent to an LLM. The proposed LLM is designed to handle a micro-turn se- quence, where user and system micro-turns are interleaved us- ing a dedicated end-of-turn token <EOS>. Given the micro-turn history, the LLM predicts the next system micro-turn and then sends it to a streaming TTS module, which synthesizes system audio incrementally. Importantly, the micro-turn does not nec- essarily contain speech content, but can contain special tokens that control the conversational flow. 3.2. Conversational Special Tokens To make turn-taking decisions controllable under streaming constraints, we introduce conversational special tokens. 3.2.1. User’s special token. • <no voice>: insert when the buffer is flushed but no text has been recognized, to represent user silence for the current ∆t interval. 3.2.2. System’s special tokens. • <user is speaking>: output when the user is still speaking, to represent the system should stay silent; the sys- tem micro-turn ends immediately with <EOS>. • <user finish speaking>: output when the user has finished speaking, to represent the system should start re- sponding; this token is followed by the system utterance con- tent. • <user is interrupting>: output when the user in- terrupts while the system is speaking, to represent the system should stop generation; the system micro-turn ends immedi- ately with <EOS>. • <user backchannel>: output when the user produces a backchannel while the system is speaking, to represent the system should ignore it and continue its current utterance. • <user is thinking>: output when the user is silent af- ter the system finishes response and is likely thinking, to rep- resent the system should take no action; the system micro- turn ends immediately with <EOS>. • <system backchannel>:output when the system should emit a short backchannel during user speech; once the LLM outputs <system backchannel>, we immedi- ately play a pre-synthesised backchannel audio clip (sampled at random) and end the micro-turn with <EOS>. 3.3. Dynamic Construction of Duplex Training Data What is the capital of Japan?It is Tokyo. What isthe capital of Japan?It isTokyo. What is <S> the capital of Japan?<S>It isTokyo. <S> <S><S> User’s Turn System’s Turn <S> <S> User’s Turn with special token System’s Turn with special token Split utterance-wise long-turns into chunk-wise micro-turns. Insert corresponding micro-turnswith special token. Figure 2: Proposed dynamic construction pipeline for duplex training sequences from text-only dialogues. Table 1: Full-Duplex-Bench results. The best-performing duplex dialogue model is shown in bold, and the second best is underlined. § means the results are infered from offical checkpoints. † means the results directly obtained from the paper. DimensionPause HandlingBackchannelSmooth Turn TakingUser InterruptionAveraged Turn-Taking Accuracy Data Metric Synthetic TOR (↓) Candor TOR (↓) TOR (↓) ICC Freq (↑) JSD (↓) Candor TOR (↑) Latency (↓) TOR (↑) Synthetic GPT-4o(↑) Latency (↓) dGSLM † 0.9340.9350.6910.0150.9340.9750.3520.9170.2012.5310.466 Moshi † 0.9850.9801.0000.0010.9570.9410.2651.0000.7650.2570.395 Freeze-Omni † 0.6420.4810.6360.0010.9970.3360.9530.8673.6151.4090.489 PersonaPlex † 0.3580.431 0.2730.0420.6620.9080.1700.9504.2900.2400.759 MiniCPM-Duplex § 0.6130.7310.5820.0030.9600.9160.6101.01.140.4470.598 Gemini Live † 0.2550.3100.0910.0120.8960.6551.3010.8913.3761.1830.778 DuplexCascade0.0580.2220.2180.0090.9490.8321.7240.9554.0161.2250.858 DuplexCascade-β0.343 0.4580.3090.0340.8110.8990.5670.9504.0110.8500.748 Real full-duplex dialogue corpora with turn-taking anno- tations are scarce. We therefore randomly sample 50k dia- logues from UltraChat [23] and construct duplex-style train- ing sequences dynamically by simulating 6 key interaction phe- nomena. 3.3.1. Convert text dialogue into micro-turn segmentation As illustrated in Fig. 2, we split both user and system utterance- wise long-turns into chunk-wise micro-turns. After each user micro-turn, we insert a system micro-turn that contains only <user is speaking> to indicate that the user continues. After each system micro-turn, we insert a user micro-turn <no voice> to represent that the user is silent. For the first system micro-turn of each original system turn, we prepend <user finish speaking> before the content to prompt the sys- tem to take the turn. During training, we apply next-token prediction LoRA fine-tuning [13] only on system micro-turns, which helps preserve the backbone LLM’s conversational intel- ligence. 3.3.2. Simulating duplex interaction phenomena To better approximate real dialogue conditions, we simulate the following interaction phenomena. Randomized micro-turn length. To emulate the variability of streaming ASR text emission, we sample the user micro-turn length uniformly from 1–7 tokens randomly. For training sta- bility, we fix the system micro-turn length to 10 tokens. Natural pauses. After each user micro-turn that are split from the original long turn, with probability 0.10 we randomly in- sert 1–5 additional silent user micro-turns represented by <no voice>. For these inserted silent micro-turns, the correspond- ing system micro-turns are supervised to output <user is speaking>, indicating that the system should refrain from taking the turn. User interruption. For each system long turn, with probabil- ity 0.30 we simulate an interruption by letting the user start the next question at a randomly selected system micro-turn bound- ary. Upon receiving the first interrupting user micro-turn, the system is supervised to output <user is interrupting> and abandon the remainder of the interrupted response; during the subsequent user continuation, system micro-turns are super- vised to output <user is speaking>. User backchannels. While the system is speaking, each system micro-turn independently has a probability of 0.01 of replacing the next user <no voice> with a short textual backchannel (e.g., “yes”, “okay”). In this case, the system is supervised to output <user backchannel> and continue generating the ongoing response. System backchannels.Because text-only dialogues rarely contain explicit system backchannels, we post-process user turns with Qwen2-72B-Instruct 2 to insert marker <BC/> after sentence-level units where a backchannel is appropriate. Dur- ing training, these markers are used to supervise the system to output the token <system backchannel>. User thinking. After the system finishes its response, we randomly insert 1–20 additional silent user micro-turns repre- sented by <no voice>. For these inserted silent micro-turns, the corresponding system micro-turns are supervised to output <user is thinking>, indicating that the system should wait while the user is processing the response. 4. Experiments 4.1. Implementation Details We implement the proposed cascaded pipeline with stream- ing ASR and streaming TTS based on DSM-ASR and DSM- TTS [24].Following Freeze-Omni, we adopt Qwen2-7B- Instruct 3 as the text LLM backbone and perform lightweight adaptation on our dynamically constructed duplex training set using LoRA [13] on the query and value projections (LoRA rank r=16, α=32). The embedding parameters of the conver- sational special tokens are initialized from a zero-mean Gaus- sian distribution. During adaptation, we fully fine-tune the original to- ken embedding matrix, the embeddings of the newly intro- duced special tokens, and the token prediction head, while the remaining backbone parameters are updated through LoRA. To address class imbalance among conversational spe- cial tokens, we use a weighted loss with the following weights: <user is speaking> = 1, <user finish talking> = 10, <user is interrupting> = 5, <user backchannel> = 2, <user is thinking> = 1, and <system backchannel> = 3. We train two variants: DuplexCascade (without system backchannels) and DuplexCascade-β (with system backchan- nels), to evaluate whether training LLM with <system backchannel> affects overall model performance. Duplex- Cascade does not include the <system backchannel> to- ken and is trained without any system-backchannel supervision, whereas DuplexCascade-β enables this token and is trained with the corresponding simulation described in Sec. 3.3. Optimization uses AdamW [25] with learning rate 1×10 −5 2 https://huggingface.co/Qwen/Qwen2-72B-Instruct 3 https://huggingface.co/Qwen/Qwen2-7B-Instruct Table 2: VoiceBench results. The best-performing duplex dialogue model is shown in bold, and the second best is underlined. § means the results are infered from offical checkpoints. † means the results directly obtained from the paper. ModelAlpacaEval CommonEval WildVoice SD-QA MMSU OBQA BBH IFEval AdvBench Overall DSM-ASR+ Qwen2-7B-Instruct 4.573.873.7346.4258.5272.0958.1051.2997.1269.66 Freeze-Omni † 4.033.463.1553.4528.1430.9850.7023.4097.3055.20 Moshi † 2.011.601.3015.6424.0425.9347.4010.1244.2329.51 PersonaPlex § 2.692.251.9518.7724.9224.4049.4011.958.0730.59 DuplexCascade4.403.643.5645.5652.8656.0459.8043.3899.0465.41 DuplexCascade-β4.383.703.5346.3552.8458.4659.0044.4599.0365.81 and a linear warmup schedule for 500 steps. We set the batch size to 32, fine-tune for 5k steps, and use a maximum sequence length of 4096 tokens. Each model was trained using 8 NVIDIA H100 GPUs for only 5 hours. 4.2. Full-Duplex-Bench Results We evaluate DuplexCascade on Full-Duplex-Bench [14] against other state-of-the-art models. The benchmark reports Take- Over Rate (TOR) for each condition; however, TOR has mixed directions across subsets, making it inconvenient to summarize overall turn-taking quality. To provide an aggregated view, we introduce Averaged Turn-Taking Accuracy. For subsets where lower TOR is bet- ter (Pause Handling and Backchannel), we define accuracy as 1−TOR. For subsets where higher TOR is better (Smooth Turn Taking and User Interruption), we define accuracy as TOR. Av- eraged Turn-Taking Accuracy is the unweighted mean of these accuracies across all TOR-reported subsets. Table 1 shows that DuplexCascade achieves the best Aver- aged Turn-Taking Accuracy among the evaluated open systems, indicating consistently strong turn-taking decisions across di- verse duplex scenarios. In contrast, Freeze-Omni, which uses VAD-driven approaches that rely on an external endpointing head, exhibits noticeably weaker turn-taking robustness, high- lighting the benefit of explicitly modeling duplex control deci- sions via LLM tokens rather than inferring them from external endpointing head. And DuplexCascade-β achieved the second-best results on backchannel-related metrics (ICC Freq and JSD), while com- petitive on overall turn-taking accuracy. The results suggest that our proposed method can control system’s reaction styles by text-only LLM training. 4.3. VoiceBench Results A practical spoken dialogue system must also preserve the rea- soning and instruction-following capability of the underlying text LLM. We therefore evaluate conversational intelligence us- ing VoiceBench [12]. In addition to other state-of-the-art duplex models, to assess how our method affects the LLM’s conversa- tional intelligence, we also implement a naive pipeline as de- fined in VoiceBench [12], which combines DSM-ASR [24] and Qwen2-7B-Instruct. As shown in Table 2, our models substantially outper- form prior duplex systems across nearly all VoiceBench di- mensions, suggesting that LoRA fine-tuning with text (instead of speech) helps preserve conversational intelligence by avoid- ing cross-modality alignment issues. Compared with the naive DSM-ASR+Qwen2-7B-Instuct pipeline, DuplexCascade and DuplexCascade-β achieve competitive scores with a moderate 0.30.60.91.21.51.8 t 0.78 0.80 0.82 0.84 0.86 0.88 0.90 0.92 0.94 Averaged Turn-Taking Accuracy Accuracy & Latency vs. t Turn-Taking Accuracy Turn Taking Latency 1.50 1.75 2.00 2.25 2.50 2.75 3.00 3.25 3.50 Smooth Turn Taking Latency Figure 3: Effect of the micro-turn duration ∆t on Averaged Turn-Taking Accuracy and turn-taking latency on Full-Duplex- Bench. gap, suggesting that our text-only adaptation largely retains the backbone LLM’s capabilities despite operating under stream- ing micro-turn constraints. Notably, DuplexCascade-β demon- strates that our method enables the LLM to produce backchan- nels while preserving its core capabilities. 4.4. Micro-Turn ∆t Analysis We study how the micro-turn duration ∆t affects turn-taking accuracy and response latency. Specifically, we evaluate ∆t ∈ 0.3, 0.6, 0.9, 1.2, 1.5, 1.8 s on Full-Duplex-Bench, and re- port the resulting Averaged Turn-Taking Accuracy and Smooth Turn-Taking Latency in Fig. 3. We observe that Averaged Turn- Taking Accuracy improves as ∆t increases up to 1.2 s, after which it degrades. In contrast, Smooth Turn-Taking Latency increases monotonically with larger ∆t, since the system must wait longer before each buffer flush. These results suggest that, under our simulation setting, ∆t=1.2 s provides the strongest turn-taking performance, but at the cost of higher latency. We therefore choose ∆t=0.6 s as a practical trade-off between turn- taking accuracy and latency. 5. Conclusion In this work, we presented DuplexCascade, a practical cas- caded pipeline for full-duplex speech-to-speech dialogue. By adapting a strong text LLM with only a small amount of text- only dialogues via lightweight LoRA fine-tuning, DuplexCas- cade achieves VAD-free and robust turn-taking ability while largely preserving the conversational intelligence of the under- lying LLM. These results highlight that full-duplex interaction can be enabled within a modular cascade without sacrificing the strengths of modern text LLMs. 6. Generative AI Use Disclosure Generative AI tools were used for language editing and improv- ing the phrasing of this manuscript. 7. References [1] S. Ji, Y. Chen, M. Fang, J. Zuo, J. Lu, H. Wang, Z. Jiang, L. Zhou, S. Liu, X. Cheng, X. Yang, Z. Wang, Q. Yang, J. Li, Y. Jiang, J. He, Y. Chu, J. Xu, and Z. Zhao, “WavChat: A Survey of Spoken Dialogue Models,” 2024. [2] Y. Shangguan, J. Li, L. Qiao, R. Alvarez, and I. McGraw, “Ana- lyzing the Quality and Stability of a Streaming End-to-End On- Device Speech Recognizer,” in Proc. INTERSPEECH, 2020. [3] T. Saeki, S. Takamichi, and H. Saruwatari, “Incremental text-to- speech synthesis using pseudo lookahead with large pretrained language model,” IEEE Signal Processing Letters, vol. 28, p. 857–861, 2021. [4] N. Torgashov, G. E. Henter, and G. Skantze, “VoXtream: Full-Stream Text-to-Speech with Extremely Low Latency,” arXiv:2509.15969, 2025. [5] X. Wang, Y. Li, C. Fu, , Y. Zhang, Y. Shen, L. Xie, K. Li, X. Sun, and L. Ma, “Freeze-Omni: A Smart and Low Latency Speech-to- speech Dialogue Model with Frozen LLM,” ICML, 2025. [6] M. Shannon, G. Simko, S.-Y. Chang, and C. Parada, “Improved End-of-Query Detection for Streaming Speech Recognition,” in Proc. Interspeech, 2017, p. 1909–1913. [7] D. Liang, H. Su, T. Singh, J. Mahadeokar, S. Puri, J. Zhu, E. Thomaz, and M. Seltzer, “Dynamic speech endpoint detection with regression targets,” in Proc. ICASSP. IEEE, 2023, p. 1–5. [8] G. Skantze, “Turn-taking in conversational systems and human- robot interaction: a review,” Computer Speech & Language, vol. 67, p. 101178, 2021. [9] A. D ́ efossez, L. Mazar ́ e, M. Orsini, A. Royer, P. P ́ erez, H. J ́ egou, E. Grave, and N. Zeghidour, “Moshi: a speech-text foundation model for real-time dialogue,” Tech. Rep., 2024. [10] R. Roy, J. Raiman, S.-g. Lee, T.-D. Ene, R. Kirby, S. Kim, J. Kim, and B. Catanzaro, “PersonaPlex: Voice and Role Control for Full Duplex Conversational Speech Models,” 2026. [11] T. A. Nguyen, E. Kharitonov, J. Copet, Y. Adi, W.-N. Hsu, A. Elkahky, P. Tomasello, R. Algayres, B. Sagot, A. Mohamed et al., “Generative spoken dialogue language modeling,” Trans- actions of the Association for Computational Linguistics, vol. 11, p. 250–266, 2023. [12] Y. Chen, X. Yue, C. Zhang, X. Gao, R. T. Tan, and H. Li, “VoiceBench:Benchmarking LLM-Based Voice Assistants,” arXiv preprint arXiv:2410.17196, 2024. [13] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-Rank Adaptation of Large Language Models,” in Proc. ICLR, 2022. [14] G.-T. Lin, J. Lian, T. Li, Q. Wang, G. Anumanchipalli, A. H. Liu, and H.-y. Lee, “Full-duplex-bench: A benchmark to evaluate full- duplex spoken dialogue models on turn-taking capabilities,” arXiv preprint arXiv:2503.04721, 2025. [15] R. Maas, A. Rastrow, C. Ma, G. Lan, K. Goehner, G. Tiwari, S. Joseph, and B. Hoffmeister, “Combining Acoustic Embed- dings and Decoding Features for End-of-Utterance Detection in Real-Time Far-Field Speech Recognition Systems,” in Proc. IEEE ICASSP, 2018, p. 5544–5548. [16] J. H. Ko, J. Fromm, M. Philipose, I. Tashev, and S. Zarar, “Lim- iting Numerical Precision of Neural Networks to Achieve Real- Time Voice Activity Detection,” in Proc. IEEE ICASSP, 2018, p. 2236–2240. [17] A. Raux and M. Eskenazi, “Optimizing Endpointing Thresholds using Dialogue Features in a Spoken Dialogue System,” in Proc. SIGDIAL, 2008, p. 1–10. [18] A. Maier, J. Hough, and D. Schlangen, “Towards Deep End-of- Turn Prediction for Situated Spoken Dialogue Systems,” in Proc. Interspeech, 2017, p. 1676–1680. [19] G. Castillo-L ́ opez et al., “A Survey of Recent Advances on Turn- taking Modeling in Conversational Systems,” in Proc. IWSDS, 2025, p. 254–271. [20] X. Zhang, Y. Chen, S. Hu, X. Han, Z. Xu, Y. Xu, W. Zhao, M. Sun, and Z. Liu, “Beyond the Turn-Based Game: Enabling Real-Time Conversations with Duplex Models,” in Proc. EMNLP, 2024, p. 11 543–11 557. [21] B. Veluri, B. N. Peloquin, B. Yu, H. Gong, and S. Gollakota, “Be- yond Turn-Based Interfaces: Synchronous LLMs as Full-Duplex Dialogue Agents,” in Proc. EMNLP, 2024, p. 21 390–21 402. [22] Q. Zhang, L. Cheng, C. Deng, Q. Chen, W. Wang, S. Zheng, J. Liu, H. Yu, C. Tan, Z. Du, and S. Zhang, “OmniFlatten: An End-to-end GPT Model for Seamless Voice Conversation,” in Proc. ACL, 2025, p. 14 570–14 580. [23] N. Ding, Y. Chen, B. Xu, Y. Qin, S. Hu, Z. Liu, M. Sun, and B. Zhou, “Enhancing chat language models by scaling high- quality instructional conversations,” in Proc. EMNLP, 2023, p. 3029–3051. [24] N. Zeghidour,E. Kharitonov,M. Orsini,V. Volhejn, G. de Marmiesse, E. Grave, P. P ́ erez, L. Mazar ́ e, and A. D ́ efossez, “Streaming Sequence-to-Sequence Learning with Delayed Streams Modeling,” Tech. Rep., 2025. [25] I. Loshchilov and F. Hutter, “Decoupled Weight Decay Regular- ization,” in Proc. ICLR, 2017.