Emo-LiPO: Listwise Preference Optimization for Fine-Grained Emotion Intensity Control in LLM-based Text-to-Speech

1 Introduction

Large language model (LLM)-based text-to-speech (TTS) systems have recently enabled prompt-conditioned control over speaking style and emotion [Leng et al., 2023; Du et al., 2024a; Du et al., 2024b; Zhang et al., 2025b; Yang et al., 2025a; Zhou et al., 2026], offering substantially greater flexibility than traditional signal-driven approaches [Zhou et al., 2022b; Wang et al., 2023; Schnell and Garner, 2021; Im et al., 2022; Lei et al., 2022; Cho et al., 2024; Cho et al., 2025]. Despite this progress, fine-grained control over emotion intensity remains a fundamental challenge for promptconditioned LLM-based TTS systems [Guo et al., 2023b; Leng et al., 2023; Du et al., 2024a; Yang et al., 2025a].

∗Corresponding author.

A key difficulty lies in the semantic-acoustic gap between textual emotion descriptions and their realization in speech. While natural language prompts can effectively convey emotion categories (e.g., happy or sad), they often fail to reliably induce perceptually distinguishable variations in emotion strength [Guo et al., 2023b; Leng et al., 2023; Zhang et al., 2025b], as illustrated in Figure 1. As a result, existing models tend to realize emotion intensity implicitly at the acoustic level, leading to unstable or weakly differentiated intensity expressions, especially under high-intensity emotional prompts [Du et al., 2024a; Du et al., 2024b; Yang et al., 2025a]. This limitation significantly constrains the controllability and expressiveness of emotional TTS systems in real-world applications.

Meanwhile, recent advances in preference optimization have demonstrated strong potential for aligning generative models with subjective human judgments. Pair-wise and list-wise preference learning frameworks provide principled mechanisms for modeling relative comparisons, making them particularly suitable for perceptual generation tasks [Rafailov et al., 2023; Zhao et al., 2023; Liu et al., 2025a]. In the context of TTS, emerging preference-aligned methods have shown improved emotional expressiveness and perceptual quality by encouraging preferred outputs over less preferred ones [Zhang et al., 2024b; Gao et al., 2025; Liu et al., 2025b; Yang et al., 2025b; Li et al., 2025]. However, existing preference-based TTS approaches primarily focus on controlling emotion at the categorical level, without explicitly modeling the ordinal structure of emotion intensity. Consequently, relative intensity information specified in text prompts is not directly aligned with preference supervision during training.

In this work, we address the challenge of fine-grained emotion intensity control in LLM-based TTS by framing it as a learning-to-rank problem. We adopt a listwise preference optimization approach that explicitly aligns prompt-conditioned speech generation with relative emotion intensity conveyed by natural language. To facilitate systematic modeling and evaluation of emotion intensity, we additionally construct a multi-speaker dataset with explicit intensity variations and conduct extensive experiments to validate the effectiveness of our approach. Our main contributions can be summarized as follows:

• We propose Emo-LiPO 1 , a listwise preference optimization framework that formulates emotion intensity control in LLM-based TTS as a learning-to-rank problem, enabling fine-grained and text-grounded intensity modeling.
• We construct ESD-plus 2 , a multi-speaker emotional speech dataset with explicit emotion intensity variations to support fine-grained emotion intensity modeling and evaluation.
• Extensive experiments on ESD-plus demonstrate that EMO-LiPO outperforms supervised- and DPO-based baselines in emotion accuracy and intensity controllability, with especially pronounced gains at high intensity.

2 RelatedWork

2.1 Emotion Intensity Control in TTS

Early studies on emotion intensity control mainly operate in the acoustic domain, where emotion strength is explicitly modeled as a continuous variable derived from speech signals using relative ranking or learning-to-rank objectives [Zhou et al., 2022b; Wang et al., 2023]. Other approaches extract intensity representations at different temporal levels and inject them into TTS systems for fine-grained control [Schnell and Garner, 2021; Im et al., 2022; Lei et al., 2022]. More recent methods further disentangle emotion intensity from style using geometric representations or incorporate soft intensity guidance through diffusion-based models [Cho et al., 2024; Cho et al., 2025; Liang et al., 2025; Guo et al., 2023a; Wu et al., 2024]. While effective in controlled settings, these approaches typically rely on explicit acoustic features or specialized architectures and lack flexible, text-driven intensity control. In contrast, recent prompt-conditioned and LLMbased TTS systems leverage natural language descriptions for expressive speech generation, but primarily control discrete emotion categories or coarse prosodic attributes, leaving emotion intensity implicitly realized [Guo et al., 2023b; Leng et al., 2023; Zhang et al., 2025b; Du et al., 2024a; Du et al., 2024b; Yang et al., 2025a]. Our approach directly grounds emotion intensity in natural language and aligns it with preference supervision, enabling explicit and finegrained intensity control in LLM-based TTS.

1 Code is available at hlt-cuhksz/Emo-LiPO.

2 Dataset is available at hlt-cuhksz/ESD-plus.

2.2 Preference Optimization for Generation Tasks

Preference optimization aims to align pre-trained large language models (LLMs) with human preferences through gradient-based optimization during training. Existing approaches can be broadly categorized into point-wise, pairwise, and list-wise methods. Point-wise methods optimize models using feedback defined on individual samples, typically in the form of scalar rewards or binary desirability signals, as exemplified by Proximal Policy Optimization (PPO) [Schulman et al., 2017], Reward Maximization (ReMax) [Li et al., 2024], and Kahneman-Tversky Optimization (KTO) [Ethayarajh et al., 2024]. Pair-wise methods, such as Direct Preference Optimization (DPO) [Rafailov et al., 2023; Zhang et al., 2025a] and SLiC-HF [Zhao et al., 2023], learn from relative preferences between two candidate outputs by contrasting a preferred response against a less preferred one, shaping the model's generation distribution accordingly [Zhang et al., 2024a]. List-wise methods further generalize preference alignment as a ranking problem, directly optimizing list-structured preference signals within a learning-to-rank framework [Song et al., 2024; Liu et al., 2025a]. These methods motivate us by providing a principled approach to learning from subjective human judgments in natural language processing, making them especially well suited for perceptual generation tasks such as emotional speech synthesis.

2.3 Preference Alignment for TTS

Recent studies have begun to incorporate preference alignment into TTS systems to better match human perceptual judgments of speech quality, expressiveness, and emotional appropriateness. SpeechAlign [Zhang et al., 2024b] is an early attempt in this direction, aligning codec-based speech language models by learning preferences between golden and synthetic codec tokens to mitigate distribution mismatches between training and inference. Building upon this paradigm, several works apply preference optimization to emotional speech synthesis. Emo-DPO [Gao et al., 2025] and ARDMDPO [Liu et al., 2025b] adopt pair-wise preference objectives to encourage emotionally preferred outputs over less preferred ones, while RLAIF-SPA [Yang et al., 2025b] and EMORL-TTS [Li et al., 2025] employ reinforcement learning with task-specific rewards to improve emotional expressiveness and controllability. These approaches demonstrate the effectiveness of preference-based optimization for enhancing perceptual speech quality. However, existing preferencealigned TTS methods mainly focus on selecting better samples rather than explicitly modeling the ordinal structure of emotion intensity. As a result, emotion strength is typically optimized implicitly at the acoustic level, and relative intensity specified in text prompts is not directly aligned with preference signals.

3 Method

3.1 ProblemFormulation

Task Definition. We investigate fine-grained control of emotion intensity in LLM-based TTS systems. Given a text transcript and an emotion specification, the TTS model generates speech conditioned on both linguistic content and emotional attributes. Natural language prompts specify an emotion category and, for non-neutral emotions, a relative intensity level. Our goal is to train an LLM-based TTS system to generate speech that accurately follows these specifications.

Formally, let C = Cemo ∪ {neutral} denote the set of emotion categories, where Cemo = {happy, sad, angry, . . .} and neutral corresponds to emotionally flat speech. For each c ∈ Cemo, we define an ordered set of intensity levels L = {l1, l2, . . . , lK}, where li < lj for i < j indicates strictly increasing emotional intensity (e.g., slightly, moderately, extremely). Crucially, the weakest level l1 must remain perceptually distinct from neutral speech, preserving a clear boundary between emotional and non-emotional outputs.

We define the model input as x = (t, Pc,l), consisting of a text transcript t and an emotion prompt Pc,l specifying an emotion category c and intensity level l. The TTS model generates speech S = πθ(x), which should satisfy: (i) Content fidelity, meaning the speech faithfully conveys the linguistic content of the transcript t; (ii) Category correctness, meaning the generated speech follows emotion c exactly; and (iii) Intensity ordering for c ∈ Cemo, such that speech generated with higher intensity levels is perceived as strictly stronger than that generated with lower levels.

Learning-to-Rank Formulation via LiPO (Emo-LiPO). To meet these requirements, we adopt the Listwise Preference Optimization (LiPO) framework [Liu et al., 2025a] and formulate fine-grained emotion intensity control as a Learningto-Rank (LTR) problem. In this formulation, all candidate speech samples compared under the same prompt are generated from the same text transcript. LiPO learns from listwise preference data defined over multiple candidates generated for the same prompt, making it well suited for modeling both emotion category correctness and global intensity ordering.

In contrast, Supervised Fine-Tuning (SFT) relies on independent prompt-target pairs, DSFT =

{(x = (t, Pc,l), y = Sc,l)}, which provide no supervision on relative emotional intensity across samples. Direct Preference Optimization (DPO) extends SFT by introducing pairwise preferences, DDPO = {(x = (t, Pc,l), y = Sc,l, ¯ Sc,l))}, where ¯ Sc,l denotes a negative sample. While DPO captures local preference information, it remains limited to binary comparisons. In emotional TTS, existing DPO-based methods such as Emo-DPO [Gao et al., 2025] similarly rely on pairwise emotional preferences 3 . As a result, these approaches are inherently constrained to local comparisons and cannot enforce a global ordering of emotional intensity.

Notably, Emo-LiPO adopts a listwise formulation, DLiPO = {(x = (t, Pc,l), y = (Tc,l, ψc,l))}, where Tc,l is a list of speech samples generated by the TTS model, and ψc,l ∈ [0, 1]|Tc,l| is a vector of real-valued preference scores associated with the corresponding samples, with larger values indicating stronger preference. This listwise supervision enables Emo-LiPO to explicitly model fine-grained emotional intensity rankings across multiple candidates.

3.2 Emo-LiPO Framework

We present the Emo-LiPO framework by specifying the construction of listwise preference data and the corresponding optimization objective under the LiPO formulation.

Rule-Based Preference Construction. For a prompt Pc,l, we construct a listwise preference set Tc,l of K + 2 speech candidates with the same text transcript t using a rule-based ranking strategy. The list consists of: (i) a target sample Sc,l with the exact emotion category c and intensity level l; (ii) K - 1 same-emotion samples with the same category c but different intensity levels l′ ∈ L \ {l}; (iii) one neutral speech sample; and (iv) one negative sample with a randomly selected non-target emotion category. Under this rule-based strategy, the K-1 same-emotion samples are ordered according to their absolute intensity distance |l′ - l|, with samples closer to the target intensity being preferred; ties are resolved by random ordering. The resulting preference list follows the order

3 In Emo-DPO, each training instance is defined as x = (t, Pc), y = (Sc, S¯ c).

where lclosest and lfarthest denote intensity levels of the same emotion category ordered by increasing distance from l.

Based on the rule-based ranking described above, we assign a real-valued preference label vector ψc,l ∈ [0, 1]|Tc,l| to the speech samples in Tc,l. Specifically, we adopt an indexbased preference scoring scheme, where the preference label of each sample is determined by its position in the ordered list Tc,l. Let i denote the index of a sample in the list, with smaller indices indicating stronger preference. The preference label of the i-th sample is defined as

Multi-Stage Optimization. As illustrated in Figure 2, following Emo-DPO [Gao et al., 2025], we optimize the TTS model in a multi-stage manner, starting with SFT to learn basic instruction following, and subsequently applying LiPO for fine-grained emotion intensity control. The LLM-based TTS model formulates text-to-speech as a conditional autoregressive generation task [Du et al., 2024a], in which speech tokens are generated sequentially conditioned on the joint input x = (t, Pc,l), together with a speaker embedding.

Specifically, in the SFT stage, we perform supervised finetuning on a paired prompt-speech dataset DSFT to initialize the backbone TTS model πbase:

where logπbase(S | x) is computed using teacher forcing with token-level cross-entropy over the autoregressively generated speech tokens. The resulting model is denoted as the reference TTS policy πref in the subsequent stage.

In the LiPO stage, initialized from πref , we optimize the model using the listwise preference dataset DLiPO as defined in §3.1. Given a preference list Tc,l and its associated preference label vector ψc,l, we apply Listwise Preference Optimization to encourage the model to generate speech samples that better align with the desired emotion intensity ordering. The corresponding optimization objective is defined as:

where r(·) denotes a listwise learning-to-rank loss under the LiPO framework, and s represents the ranking scores induced by the current policy πθ relative to the reference policy πref , The score of each candidate Si ∈ Tc,l is computed as:

Following the original LiPO formulation [Liu et al., 2025a], r(·) is defined as:

where ψc,l = {(i, j) | ψc,l(i) > ψc,l(j)}, the relative orders of candidate pairs (Si, Sj). The weighting term λi,j serves as an intensity-aware coefficient that scales each pairwise loss term according to the implied emotion-intensity discrepancy between (Si, Sj), injecting fine-grained intensity information beyond the binary preference order and promoting ordinal consistency for more reliable intensity control. We then define the gain and discount functions, G(·) and D(·), based on the position index i:

and λi,j is derived as:

A larger λi,j implies a larger implied intensity gap between Si and Sj under the list ranking, so the model is penalized more strongly for violating their order.

4 ESD-plus Dataset

4.1 Dataset Overview

To evaluate the effectiveness of the proposed Emo-LiPO method, we construct ESD-plus, a speech dataset with explicit emotion intensity variations designed to support research on fine-grained emotion control in TTS systems. ESDplus is built upon the English portion of the ESD corpus [Zhou et al., 2022a], from which we select 350 parallel sentences as synthesis scripts. We retain the original ESD emotion categories, including four non-neutral emotions (happy, surprise, sad, and angry) and one neutral emotion. For each non-neutral emotion, we further introduce three distinct intensity levels, resulting in a total of 13 fine-grained emotion labels.

In total, ESD-plus contains 45,500 fine-grained emotional speech samples, with an average duration of 2.92 seconds per sample, amounting to approximately 36.89 hours of audio for training and evaluation. Following the official data partitioning strategy of ESD [Zhou et al., 2022a], we split the dataset into training, development, and test sets with 300, 20, and 30 utterances, respectively, while ensuring that all speaker-emotion variants of each utterance are assigned to the same split.

4.2 Dataset Construction and Quality Control

We construct ESD-plus using a prompt-based emotional speech synthesis pipeline built upon gpt-4o-mini-tts 4 . For each emotion-intensity category, we design a structured prompt that combines a natural-language emotion description with a scalar intensity indicator, providing unified and reusable conditional control signals. To ensure perceptually distinguishable emotional variations across different intensity levels, the scalar indicators are deliberately spaced (e.g., 1/5, 3/5, and 5/5) rather than uniformly incremental. 5 Moreover, we generate speech using 10 official English voices, comprising 4 male and 6 female speakers, enabling multi-speaker emotional synthesis.

4 https://openai.fm/

5 Details on dataset construction, quality control, and dataset statistics are provided in the supplementary material (Appendix A).

To ensure data reliability, we perform human verification on the development and test sets of ESD-plus, which are used exclusively for model evaluation. Annotators perform pairwise comparisons within each emotion category to verify that speech samples generated at different intensity levels exhibit consistent and monotonically increasing emotional intensity in accordance with their prompts. Samples that fail to demonstrate the expected intensity ordering are flagged and regenerated until the target ordering is satisfied, thereby ensuring the correctness of emotion-intensity labels in the development and test sets. Overall, 91.23% of the evaluated samples pass verification, demonstrating that ESD-plus provides highquality and reliable data for model training and evaluation. 6

5 Experiments

5.1 Experimental Setup

Baselines and Implementation. We adopt several LLMbased TTS models as baselines. Specifically, we include two supervised LLM-based TTS models: CosyVoice [Du et al., 2024a], which is built upon supervised semantic tokens, and EmoVoice [Yang et al., 2025a], which supports freestyle text prompting for emotional control. In addition, we consider preference-optimized LLM-based TTS methods following Emo-DPO [Gao et al., 2025], which primarily target coarse-grained emotion category control. To evaluate their effectiveness in modeling fine-grained emotion intensity, we design three Emo-DPO variants: (i) Emo-DPO (R), trained with randomly sampled pairwise preferences (Sc,l, ¯ Sc,l); (ii) Emo-DPO (E), trained using pairwise preferences across different emotion categories at the same intensity level (Sc,l, S¯ c,l); and (iii) Emo-DPO (I), trained using pairwise preferences across different intensity levels within the same emotion category (Sc,l, Sc,¯l). Our Emo-LiPO model is built upon CosyVoice as the backbone, specifically using the CosyVoice-300M-Instruct version.

EvaluationMetric. We evaluate the generated speech from two complementary perspectives: speech quality and emotion relevance. At the speech level, we assess intelligibility and perceptual quality. Speech intelligibility is measured using Word Error Rate (WER), computed with the WhisperLarge-v3 ASR model [Radford et al., 2023]. Speech naturalness and overall perceptual quality are evaluated using three objective metrics: NISQA [Mittag et al., 2021], DNSMOS [Reddy et al., 2021], and UTMOS [Saeki et al., 2022]. At the emotion level, we evaluate how well the generated speech conveys the intended emotional content. We employ the emotion2vec model [Ma et al., 2024] for speech emotion recognition (SER) and adopt two metrics: Emotion Similarity (EmoSIM) and Recall. EmoSIM is computed as the cosine similarity between emotion embeddings extracted from the generated and ground-truth speech. Recall measures the emotion classification accuracy of emotion2vec on the generated speech. Due to its limited out-of-domain performance, we further fine-tune emotion2vec on ESD-plus and report the resulting score as Recall-ft.

6 The training set is not modified during this process, as its overall label correctness is already high.

5.2 Emotion Category Control Performance

Effect of Intensity-Aware Preference Modeling. Table 1 summarizes the performance of different prompt-conditioned TTS models in terms of emotion category control, together with their speech quality metrics. Overall, Emo-LiPO demonstrates the strongest ability to generate speech that aligns with the intended emotion categories, consistently outperforming both supervised baselines and DPO-based methods, while maintaining comparable speech quality. Compared with supervised LLM-based TTS models, preferenceoptimized approaches show clear advantages in emotion relevance, indicating that incorporating preference signals is critical for learning effective emotional control without sacrificing speech naturalness. Among the DPO-based variants, models trained with structured emotion-related preferences achieve better emotion category recognition than those trained with randomly sampled pairs, highlighting the importance of informative preference construction. Notably, introducing intensity-aware preference modeling further improves emotion category control. Emo-DPO trained with intensitylevel preferences performs better than variants relying solely on cross-category comparisons, suggesting that modeling relative intensity ordering within emotions provides transferable supervision for category-level emotion recognition. Finally, Emo-LiPO consistently achieves the best performance across emotion relevance metrics, demonstrating that listwise preference optimization enables more accurate and robust emotion category control under fine-grained emotional prompting, without degrading speech quality.

The impact of weighting term λ. To analyze the role of the distance-based weighting term λ (Eq. 8) in learning stable intensity ordering, we conduct an ablation study in which λ is fixed to 1 for all preference pairs, resulting in a uniformly weighted LiPO loss, while keeping all other training settings unchanged. As shown in Table 1, adopting uniform weighting leads to a slight improvement in emotion similarity but consistently degrades emotion control metrics compared with the full model. These findings indicate that λ does not primarily improve category-level alignment; rather, it provides more informative supervision for intensity-aware ranking by emphasizing ordinal consistency across candidates, thereby enhancing controllability and generalization during evaluation.

5.3 Emotion Intensity Control Performance

Automatic Evaluation Comparison. Intuitively, higher emotional intensity leads to more perceptually salient speech and should therefore result in higher emotion recognition accuracy. Accordingly, we evaluate emotion recognition accuracy (Recall-ft) across different emotion intensity levels, as shown in Figure 3. Emo-LiPO achieves the strongest overall performance and is the only model that exhibits a clear and stable monotonic increase in Recall-ft from low to high intensity levels. In contrast, supervised baselines and DPO-based variants display weaker or non-monotonic trends, indicating inconsistent alignment between textual intensity descriptions and the generated speech. These results demonstrate that Emo-LiPO more faithfully translates relative emotion intensity specifications into perceptually distinguishable speech, validating its effectiveness in fine-grained emotion intensity control.

Human Evaluation Comparison. To further validate the effectiveness of Emo-LiPO in emotion intensity control, we conduct a human Arena-style evaluation. We perform pairwise comparisons between Emo-LiPO and three representative baselines (CosyVoice, Emo-DPO (E), and Emo-DPO (I)) along three perceptual dimensions: Speech Quality, Emotion Expression, and Intensity Control. For each comparison, annotators are presented with the same text prompt and two candidate models, each providing a pair of speech samples generated under the same script and emotion category at adjacent intensity levels. The evaluated intensity transitions include (Sneu, Sc,llow ), (Sc,llow , Sc,lmid ), and (Sc,lmid , Sc,lhigh ). Annotators select the model that performs better with respect to a single evaluation dimension, with a tie option allowed when performance is comparable. The order of models is randomized, and five annotators participate in the evaluation. 7

Table 2 reports the Arena win rates of Emo-LiPO compared with three baseline models across three perceptual dimensions. Overall, Emo-LiPO consistently outperforms all baselines in all evaluation dimensions, demonstrating strong perceptual advantages under human judgment. When compared with Emo-DPO (I), the advantage of Emo-LiPO in intensity control is less pronounced, suggesting that incorporating intensity-aware pairwise preferences already captures part of the fine-grained intensity information. However, EmoLiPO consistently outperforms Emo-DPO (I) in Speech Quality, indicating that listwise preference optimization better preserves acoustic naturalness while maintaining competitive intensity controllability. This comparison highlights a favorable trade-off achieved by Emo-LiPO, where global intensity ordering is modeled without sacrificing speech quality.

Score Distance Visualization. To diagnose whether the Emo-LiPO policy πθ induces a separable preference structure consistent with the rule-based list supervision in Eq. 1, we visualize the distribution of score margins between the target sample Sc,l and the candidates in Tc,l. Based on Eq. 5, we compute three absolute margins aligned with the list components: (i) the margin to closest same emotion sample, ∆sclosest = |s(Sc,l) - s(Sc,lclosest )|; (ii) the margin to neutral sample, ∆sneu = |s(Sc,l) - s(Sneu)|; and (iii) the margin to other emotion sample, ∆s¯ c = |s(Sc,l) - s(S¯ c)|. For comparability across scales, we plot the logarithm of these margins. As shown in Fig. 4, the three margins exhibit a stable hierarchy in both location and central tendency: ∆s¯ c is the largest and most concentrated, followed by ∆sneu, while ∆sclosest is the smallest and closer to zero with a longer tail, reflecting the finer and more challenging discrimination between nearest intensities within the same emotion. Overall, this distributional hierarchy is consistent with the rule-based order in Eq. 1, indicating that Emo-LiPO shapes a scoring landscape that prioritizes cross-emotion separation while preserving fine-grained intensity hierarchy modeling within the same emotion.

7 Detailed human evaluation settings, including the evaluation scale, platform, and annotation guidelines, are provided in the supplementary material (Appendix B).

5.4 Ablation Studies and Analysis

Effect of the Number of Negative Samples S¯ c. We investigate how the number of negative samples affects intensity learning in listwise optimization by varying the number of negatives while keeping the target intensity chain fixed. Each negative sample is drawn from a non-target emotion category under the same text condition. As shown in Figure 5 (a), introducing a single negative sample yields the best overall performance and produces a clearer monotonic trend across intensity levels. In contrast, increasing the number of negative samples to two or three does not lead to further gains and instead degrades performance. These results suggest that a limited amount of cross-emotion contrast is sufficient for effective intensity learning, whereas excessive negative samples weaken the supervision signal for modeling intensity ordering.

Effect of Ranking Strategy for Preference List Construction. To study how the ranking strategy for constructing preference lists Tc,l affects intensity learning, we compare three ordering schemes on the same candidate set. Specifically, we evaluate the proposed rule-based strategy against two acoustic-driven alternatives, where candidates are ranked by their distance to the target speech in terms of average pitch or energy. As shown in Figure 5 (b), the rule-based ranking achieves the best overall performance among the three strategies. In comparison, pitch-based ranking performs the worst, while energy-based ranking yields intermediate results. These findings suggest that rankings derived from a single acoustic attribute provide weaker supervision signals, whereas the rule-based strategy offers a more effective and stable ordering for listwise preference optimization.

Effect of Weighting Mechanism in Listwise Optimization. We study the effect of the weighting mechanism λ in listwise optimization under identical training settings. By default, Emo-LiPO adopts an index-based preference labeling scheme for ψc,l, from which the distance-aware weighting term λi,j is derived. We compare this default setting with two variants: (i) fixed-λ, which disables weighting by setting λi,j = 1 for all pairs; and (ii) score-λ, where λi,j is computed from pre-assigned preference score gaps. As shown in Figure 5 (c), introducing distance-aware weighting consistently improves performance over the fixed-λ variant. Index-based and score-based weighting achieve similar results, indicating that the performance gains mainly stem from the weighting mechanism itself rather than the specific definition of ψc,l.

Impact of Negative Sampling Strategies. We study how the sampling strategy for the single negative sample S¯ c affects intensity learning in listwise optimization. By default, our method samples the negative uniformly at random across all intensity levels (Rand). Each preference list includes one negative sample drawn from a non-target emotion category. We further compare this default strategy with three alternatives that restrict the negative sample to a specific intensity level (Low, Mid, or High). As shown in Figure 5 (d), the default random strategy achieves the best overall performance, while constraining negatives to a single intensity level consistently degrades performance. These results indicate that sampling negatives across diverse intensity levels provides more balanced supervision, whereas single-level sampling introduces bias and weakens stable intensity ordering.

6 Conclusion

This work studies fine-grained emotion intensity control in prompt-conditioned LLM-based text-to-speech systems and identifies the limitations of existing supervised and preference-aligned methods in modeling perceptually meaningful intensity variations. To address this challenge, we propose Emo-LiPO, a listwise preference optimization framework that formulates emotion intensity control as a learningto-rank problem and explicitly models global intensity ordering under fixed transcripts. We further introduce ESD-plus, a multi-speaker emotional speech dataset with explicit intensity variations to support systematic training and evaluation. Extensive automatic and human evaluations demonstrate that Emo-LiPO consistently outperforms strong supervised and DPO-based LLM TTS baselines in emotion accuracy and intensity controllability, while maintaining high speech quality. Overall, this work highlights the importance of listwise preference modeling for fine-grained emotional control and provides a principled direction for expressive and controllable speech generation.

Ethical Statement

There are no ethical issues.

Acknowledgments

This research is supported by National Natural Science Foundation of China (Grant No. 62271432), the Program for Guangdong Introducing Innovative and Entrepreneurial Teams(Grant No. 2023ZT10X044), Shenzhen Science and Technology Program(Shenzhen Key Laboratory, Grant No. ZDSYS20230626091302006), and the Shenzhen Stability Science Program 2023. This research was also supported by the Internal Project of the Shenzhen Research Institute of Big Data (SRIBD), the internal project of the Guangdong Provincial Key Laboratory of Big Data Computing under Grant B10120210117-KP02 (The Chinese University of Hong Kong, Shenzhen), and Shenzhen Key Lab of MultiModal Cognitive Computing.

Contribution Statement

Yihang Lin and Li Zhou are co-first authors and contributed equally.