acceptodds
Under review as a conference paper at ICLR 2027

CueDuet: Distilling What to Say and How to Speak for Paralinguistic-Aware Spoken Dialogue

Abstract

Speech language models have made substantial progress in spoken interaction, yet adapting response content and vocal delivery to paralinguistic cues remains challenging. Teacher-guided adaptation can transfer cue-informed dialogue behavior, but its supervision mixes general response capabilities with the local adjustments needed for paralinguistic cues. Moreover, appropriate wording alone does not ensure that the generated speech conveys a suitable vocal style. We propose CueDuet, a dual-stream on-policy distillation framework that addresses these complementary needs through teacher supervision on student-generated responses. For the text stream, a single teacher evaluates the same response prefix from cue-aware and cue-agnostic perspectives; divergence between its predictions determines where to emphasize supervision for paralinguistic adaptation. For the speech stream, a speech teacher adapted on expressive text–speech pairs conditions on the student's reply, guiding speech generation toward faithful content and an appropriate style. The evaluation covers emotion- and age-sensitive dialogue across three benchmarks and two architecturally distinct speech language models, comparing offline distillation and standard on-policy distillation baselines. Across both backbones, CueDuet improves cue-responsive content and vocal delivery over offline and standard on-policy distillation, with particularly strong gains in cue-grounded safety. Text-loss analyses show that CueDuet consistently concentrates supervision on cue-responsive positions throughout training.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Speech Language Models, Paralinguistic-Aware Spoken Dialogue, On-policy Distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_S66YOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CueDuet combines cue-weighted text on-policy distillation with an adapted speech teacher. A single text teacher evaluates identical student prefixes with and without cue-specific conditioning; the divergence between these views determines normalized token weights for distillation toward the cue-aware target. A vocabulary-compatible speech teacher, adapted on expressive text-speech pairs, supervises delivery conditioned on the student's reply. Experiments use two different speech architectures and three evaluation suites, supplemented by allocation controls, speech-teacher analysis, human-recorded inputs, listening preferences and paired-cue evaluation.

Critical strengths

The strongest experimental feature is the matched Standard OPD baseline: it shares the warm-up checkpoint, adapted speech teacher, data and default losses, changing the text-position weighting. Shuffling weights within responses preserves their distribution while disrupting alignment, providing a meaningful control. The same-trajectory analysis further isolates loss allocation, and the appendix tests weighting under forward KL, reverse KL and JSD. The speech ablations hold response text fixed, separating vocal delivery from wording. The authors also disclose important limitations and unfavorable results, including the modified SD-Eval child inputs, adult-side selectivity trade-offs and the limited human-recorded advantage over some offline conditions/

Critical weaknesses

The task-level results need uncertainty estimates. The main table and most ablations provide point estimates without independent-training variability or paired confidence intervals.

Stronger child warnings are not equivalent to uniformly better age-selective assistance.Table 10 shows that Fun's child SAR increases from 91.75 to 97.25, while adult DAR drops from 76.25 to 70.75 and paired selectivity slightly decreases from 68.75 to 68.25. MiniCPM-o has a more favorable paired improvement, but also loses adult DAR. These findings are already acknowledged, and should be made more prominent in the main safety interpretation. Adult DAR is explicitly not adult safety accuracy, so additional AI-generated technical reference | Verify before use 5 appropriate-assistance and over-refusal metrics are needed. The original-synthetic adult-child pairs also do not isolate age alone.

Questions for authors
  • How stable are the primary gains across training seeds and paired evaluation resamples?
  • Can paired age selectivity and adult-side costs be presented alongside headline child SAR?
  • Do unfiltered human comparisons against Standard OPD confirm the speech benefit?
  • What happens when teacher-view divergence is tiny or the acoustic cue is ambiguous or incorrect?
Reviewer_bnD9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes CueDuet, a dual-stream on-policy distillation framework for paralinguistic-aware speech-to-speech dialogue. The method uses differences between cue-aware and cue-agnostic teacher predictions to emphasize cue-sensitive text tokens, while an adapted speech teacher guides appropriate vocal delivery. Experiments on two speech language models and three benchmarks demonstrate improvements in emotion-aware responses, age-sensitive dialogue, and cue-grounded safety over standard distillation methods.

Critical strengths

1.The experiments cover multiple benchmarks and backbones, with useful ablation studies.

2.The method shows clear improvements in paralinguistic awareness

Critical weaknesses

1.The speech teacher relies only on response text, potentially limiting adaptation when identical responses require different vocal styles.

2.The current method leads to a slight decline in intelligence. I am curious whether incorporating intelligence-related data into the training process could help maintain or even improve performance,and how, in that case, might the model's paralinguistic capabilities change?

3.Current human evaluation methods are more akin to those used for TTS systems, as the response text is fixed; I don’t believe this offers much meaningful insight for empathetic spoken dialogue.

  1. Methodological innovation is primarily concentrated on the text side. Compared to Standard OPD, CueDuet employs the same Speech Teacher and Speech Loss; the core distinction lies solely in the dynamic weighting of text tokens.

  2. The training simplicity of the original OPD approach is compromised. The method still necessitates the additional synthesis of approximately 93,000 text–speech pairs, the fine-tuning of a Speech Teacher, and an SFT warm-up phase for the student model.

Questions for authors

see above

Reviewer_kW6mOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CueDuet, a dual-stream on-policy distillation framework for paralinguistic-aware spoken dialogue that combines cue-weighted text distillation with style-aligned speech distillation. Experimental results demonstrate that the proposed method effectively enhances the cue-responsive text generation, cue-grounded safety, and vocal expression of spoken language models, proving it to be an effective OPD architecture with promising broad applicability.

The key strength lies in the dual-stream OPD design for paralinguistic-aware dialogue. It is convincingly demonstrated that this design enhances vocal expressiveness while preserving response quality, rather than merely applying OPD to SLMs as a straightforward extension.

While Table 9 evaluates the general capability retention of the CueDuet framework, the paper lacks component-level ablations on general capabilities for the text-stream cue weighting and speech-stream OPD modules individually. Specifically, Table 2(a) primarily demonstrates the improvements of Cue-weighted OPD on SpeechRel and SAR, but does not substantiate gains in factual correctness or general reasoning. Furthermore, Table 2(b) relies on fixed response texts, which only confirms that speech-stream OPD does not degrade speech synthesis accuracy. Given that the model is trained end-to-end, it remains unclear whether the speech loss compromises the accuracy of autonomously generated textual responses. Consequently, it is currently unverified whether Cue-weighted OPD genuinely improves response correctness, nor can it be ruled out that the speech-

Critical strengths

The key strength lies in the dual-stream OPD design for paralinguistic-aware dialogue. It is convincingly demonstrated that this design enhances vocal expressiveness while preserving response quality, rather than merely applying OPD to SLMs as a straightforward extension.

Critical weaknesses

While Table 9 evaluates the general capability retention of the CueDuet framework, the paper lacks component-level ablations on general capabilities for the text-stream cue weighting and speech-stream OPD modules individually. Specifically, Table 2(a) primarily demonstrates the improvements of Cue-weighted OPD on SpeechRel and SAR, but does not substantiate gains in factual correctness or general reasoning. Furthermore, Table 2(b) relies on fixed response texts, which only confirms that speech-stream OPD does not degrade speech synthesis accuracy. Given that the model is trained end-to-end, it remains unclear whether the speech loss compromises the accuracy of autonomously generated textual responses. Consequently, it is currently unverified whether Cue-weighted OPD genuinely improves response correctness, nor can it be ruled out that the speech-stream OPD achieves enhanced vocal expressiveness at the expense of textual response quality.

Questions for authors
  1. The paper lacks ablation studies on general capabilities. Could the authors provide empirical evidence on whether text-stream cue weighting genuinely improves response correctness? Furthermore, does speech-stream OPD degrade the model's general capabilities? Component-level ablations addressing these two aspects would significantly strengthen the claims.
  2. The paper does not compare the proposed method with reinforcement learning approaches, such as ParaBridge mentioned in the related work. While CueDuet demonstrates advantages over conventional SFT and OPD, could the authors clarify whether it can serve as a viable alternative to RL-based methods like GRPO for the post-training of future SLMs? Specifically, is the relationship between the proposed framework and RL-based methods competitive or complementary?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.