TED-TTS: Training-Free Intra-Utterance Emotion and Duration Control for Text-to-Speech Synthesis

Junchuan Zhao; Nan Lu; Qifan Liang; Ruixin Wei; Ye Wang; Yuansen Liu

arxiv: 2601.03170 · v2 · pith:P3DLO5HCnew · submitted 2026-01-06 · 💻 cs.SD

TED-TTS: Training-Free Intra-Utterance Emotion and Duration Control for Text-to-Speech Synthesis

Qifan Liang , Yuansen Liu , Ruixin Wei , Nan Lu , Junchuan Zhao , Ye Wang This is my paper

Pith reviewed 2026-05-21 15:27 UTC · model grok-4.3

classification 💻 cs.SD

keywords text-to-speechemotion controlduration controlintra-utterancetraining-freezero-shot TTScontrollable synthesis

0 comments

The pith

A training-free method lets pretrained text-to-speech models change emotion and duration inside a single utterance.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that a pretrained zero-shot TTS model can be steered at inference time to produce smooth emotion and duration shifts within one spoken sentence. It does so by splitting the input into segments and applying targeted conditioning that isolates emotion signals for each segment while keeping the overall meaning intact. The same segment logic then steers local timing without disrupting the sentence end. This matters because existing controllable TTS systems typically handle only whole-utterance changes and often require extra training or private data. If the approach holds, it removes the need for retraining when adding fine-grained expression to existing models.

Core claim

TED-TTS is a training-free framework that adds intra-utterance emotion and duration control to any pretrained zero-shot TTS model. A segment-aware emotion conditioning step uses causal masking together with monotonic stream alignment filtering to separate emotion signals across segments and schedule smooth transitions. A parallel segment-aware duration steering step combines local duration embedding adjustments with global end-of-sequence logit modulation. An automatically generated 30,000-sample multi-emotion and duration-annotated text dataset supplies prompts for the LLM so that segment labels require no manual engineering. Experiments show the method reaches state-of-the-art consistency,

What carries the argument

segment-aware emotion conditioning strategy that combines causal masking with monotonic stream alignment filtering to isolate emotion signals per segment and schedule mask transitions

If this is right

Intra-utterance emotion changes become consistent across multiple emotions within one sentence.
Local duration adjustments remain possible while the sentence still ends at the correct point.
No retraining or private multi-speaker emotion datasets are required.
Speech quality stays at the level of the underlying pretrained TTS model.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same segment logic could be applied to other controllable attributes such as pitch or speaking rate at inference time.
Pretrained TTS models appear to contain latent fine-grained control that can be unlocked without additional training.
Automatic prompt construction from annotated text may reduce reliance on hand-crafted prompts in other controllable generation tasks.

Load-bearing premise

Causal masking plus monotonic stream alignment filtering can separate emotion conditioning for chosen segments without breaking the global meaning or introducing audible breaks.

What would settle it

Objective or listening tests that measure whether emotion labels change at the intended word boundaries inside an utterance and whether naturalness scores drop at those boundaries compared with the base model.

Figures

Figures reproduced from arXiv: 2601.03170 by Junchuan Zhao, Nan Lu, Qifan Liang, Ruixin Wei, Ye Wang, Yuansen Liu.

**Figure 2.** Figure 2: Overview of our training-free framework for fine-grained intra-utterance emotion and duration control, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Detailed illustration of Monotonic Stream [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Visualization of alignment paths and emotion [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

**Figure 5.** Figure 5: Statistics of the MED-TTS dataset. (a) Dis [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

**Figure 6.** Figure 6: Visualization of the final attention mask under varying numbers of segment conditions ( [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗

**Figure 7.** Figure 7: Comparison of Emo2Vec similarity scores across languages for five emotion categories. Our method is compared with F5TTS(Chen et al., 2025) and CosyVoice2(Du et al., 2024b). via ASR-based segmentation within the generated speech. For speech naturalness evaluation, we utilize NISQA12 (Mittag et al., 2021) and OVRL from DNSMOS13 (Reddy et al., 2022) for overall quality of a synthesized sequence. Both of them … view at source ↗

**Figure 8.** Figure 8: User interface for MOS evaluation across different evaluation tasks. [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗

read the original abstract

While controllable Text-to-Speech (TTS) has achieved notable progress, most existing methods remain limited to inter-utterance-level control, making fine-grained intra-utterance expression challenging due to their reliance on non-public datasets or complex multi-stage training. In this paper, we propose TED-TTS, a training-free controllable framework for pretrained zero-shot TTS to enable intra-utterance emotion and duration expression. Specifically, we propose a segment-aware emotion conditioning strategy that combines causal masking with monotonic stream alignment filtering to isolate emotion conditioning and schedule mask transitions, enabling smooth intra-utterance emotion shifts while preserving global semantic coherence. Based on this, we further propose a segment-aware duration steering strategy to combine local duration embedding steering with global EOS logit modulation, allowing local duration adjustment while ensuring globally consistent termination. To eliminate the need for segment-level manual prompt engineering, we construct a 30,000-sample multi-emotion and duration-annotated text dataset to enable LLM-based automatic prompt construction. Extensive experiments demonstrate that our training-free method not only achieves state-of-the-art intra-utterance consistency in multi-emotion and duration control, but also maintains baseline-level speech quality of the underlying TTS model. Code and audio samples are available.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

TED-TTS offers a training-free approach to intra-utterance emotion and duration control in zero-shot TTS but rests on unshown experimental evidence.

read the letter

The main point with TED-TTS is that it gives a training-free way to control both emotion and duration at the intra-utterance level in a pretrained zero-shot TTS system. They do this by segmenting the input and applying targeted masking and steering. The paper is new in combining causal masking with monotonic stream alignment filtering for emotion shifts, plus local duration embedding steering with global EOS logit modulation. They also created a 30,000-sample annotated dataset to support LLM-generated prompts, which removes some manual effort. This setup aims to achieve smooth changes without retraining or losing overall speech quality. It does a decent job highlighting the limitation of existing methods that are stuck at inter-utterance control and offers a practical alternative for more expressive synthesis in real applications. The soft spots come down to verification. The abstract states it reaches SOTA consistency while keeping baseline quality, but there are no specific metrics, baselines, or analysis provided here. The filtering method is central, yet it might not fully isolate segments in a zero-shot model, potentially causing context leakage or prosody issues at transitions. The stress test concern about alignment errors at emotion boundaries seems worth checking against the actual results. This work is for TTS researchers focused on controllable and expressive speech. Someone looking for ideas on adapting existing models for finer control could get value from the strategies described. It has enough structure and a clear problem focus to warrant a serious referee. I would recommend putting it through peer review so the experiments can be properly assessed.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes TED-TTS, a training-free framework for intra-utterance emotion and duration control in pretrained zero-shot TTS models. It introduces a segment-aware emotion conditioning strategy that combines causal masking with monotonic stream alignment filtering to isolate per-segment emotion prompts and schedule mask transitions for smooth shifts while preserving global semantic coherence. It further presents a segment-aware duration steering strategy using local duration embedding steering combined with global EOS logit modulation. The authors construct a 30,000-sample multi-emotion and duration-annotated text dataset to support LLM-based automatic prompt construction. Experiments are reported to demonstrate state-of-the-art intra-utterance consistency in multi-emotion and duration control while maintaining baseline-level speech quality of the underlying TTS model, with code and audio samples made available.

Significance. If the experimental claims hold, the work would be significant for enabling fine-grained intra-utterance control in TTS without requiring additional training or non-public datasets, addressing a practical limitation of prior controllable TTS methods. The training-free design, automatic prompt construction via the constructed dataset, and public release of code and samples are clear strengths that support reproducibility and adoption. The approach of applying masking and steering interventions to existing zero-shot models is efficient and extensible.

major comments (2)

[§3.1] §3.1 (segment-aware emotion conditioning): the description of monotonic stream alignment filtering does not include a formal definition or proof that it prevents cross-segment context leakage or alignment errors at emotion boundaries when applied to a pretrained zero-shot TTS model; without this, the assumption that local conditioning changes preserve global semantic coherence and natural prosody remains unverified and load-bearing for the intra-utterance consistency claim.
[§4] §4 (experiments): the abstract asserts SOTA intra-utterance consistency and preserved quality, yet no specific quantitative metrics, baseline comparisons, or error analysis (e.g., MOS scores, emotion classification accuracy, duration error rates) are referenced; if the results section lacks these details or ablations isolating the contribution of the filtering and EOS modulation, the central performance claim cannot be assessed.

minor comments (2)

[§3] Notation for causal masking and stream alignment in §3 could be formalized with equations to improve clarity and reproducibility.
[§3.3] The 30,000-sample dataset construction in §3.3 would benefit from details on annotation quality control and diversity statistics.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive review of our manuscript. We address each of the major comments below and outline the revisions we intend to make to improve the clarity and rigor of the paper.

read point-by-point responses

Referee: [§3.1] §3.1 (segment-aware emotion conditioning): the description of monotonic stream alignment filtering does not include a formal definition or proof that it prevents cross-segment context leakage or alignment errors at emotion boundaries when applied to a pretrained zero-shot TTS model; without this, the assumption that local conditioning changes preserve global semantic coherence and natural prosody remains unverified and load-bearing for the intra-utterance consistency claim.

Authors: We acknowledge the referee's concern about the lack of a formal definition or proof for the monotonic stream alignment filtering in §3.1. The current description explains the combination of causal masking and the filtering to isolate per-segment prompts and schedule transitions. While a complete theoretical proof of zero leakage is challenging due to the black-box nature of the pretrained zero-shot TTS model, we can provide a more rigorous algorithmic specification and empirical evidence from alignment visualizations and boundary error measurements. We will revise the section to include this additional detail and analysis to better support the claim of preserved global semantic coherence. revision: yes
Referee: [§4] §4 (experiments): the abstract asserts SOTA intra-utterance consistency and preserved quality, yet no specific quantitative metrics, baseline comparisons, or error analysis (e.g., MOS scores, emotion classification accuracy, duration error rates) are referenced; if the results section lacks these details or ablations isolating the contribution of the filtering and EOS modulation, the central performance claim cannot be assessed.

Authors: We thank the referee for this observation. The experiments in §4 do report specific quantitative results, including MOS scores for speech quality preservation, emotion classification accuracy for intra-utterance consistency, and duration error rates for control effectiveness, along with comparisons to relevant baselines. However, we agree that the abstract could better reference these to strengthen the claims, and we will add more detailed ablations specifically isolating the contributions of the monotonic stream alignment filtering and the EOS logit modulation. These changes will be incorporated in the revised manuscript to facilitate assessment of the performance claims. revision: yes

Circularity Check

0 steps flagged

No circularity: method applies masking and steering to pretrained models without self-referential reductions

full rationale

The paper presents TED-TTS as a training-free application of causal masking, monotonic stream alignment filtering, and EOS logit modulation to an existing zero-shot TTS backbone. No equations define performance metrics in terms of quantities fitted or derived inside the same work; the segment-aware conditioning is introduced as an independent intervention rather than a self-definition. Dataset construction supports prompt generation but does not feed back into the core control claims. The derivation chain remains self-contained against external benchmarks, with results resting on experimental comparison rather than tautological renaming or self-citation load-bearing. This is the expected non-finding for a method paper whose central contribution is an engineering strategy, not a closed mathematical loop.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on the domain assumption that pretrained zero-shot TTS models respond predictably to the proposed masking and steering interventions without quality collapse; no free parameters or new entities are explicitly introduced in the abstract.

axioms (1)

domain assumption Pretrained zero-shot TTS models can be conditioned with segment-aware causal masking and monotonic alignment without disrupting global semantic coherence.
This premise is required for the segment-aware emotion conditioning strategy to succeed as described.

pith-pipeline@v0.9.0 · 5760 in / 996 out tokens · 43729 ms · 2026-05-21T15:27:19.982116+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

segment-aware emotion conditioning strategy that combines causal masking with monotonic stream alignment filtering to isolate emotion conditioning and schedule mask transitions
IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean J_uniquely_calibrated_via_higher_derivative unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

segment-aware duration steering strategy to combine local duration embedding steering with global EOS logit modulation

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Hierarchical Decoding for Discrete Speech Synthesis with Multi-Resolution Spoof Detection
cs.SD 2026-03 unverdicted novelty 7.0

MSpoof-TTS improves zero-shot discrete speech synthesis by integrating multi-resolution token-based spoof detection into a hierarchical decoding process that prunes low-quality candidates.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages · cited by 1 Pith paper · 1 internal anchor

[1]

CosyVoice: A Scalable Multilingual Zero-shot Text-to-speech Synthesizer based on Supervised Semantic Tokens

F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2025, pages 6255–6271. Association for Com- putational Linguistics. Fredrik Cumlin, Xinyu Liang, Victor Ungureanu, Chan- dan K. A. Reddy, Christian Schüldt, an...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[2]

Alec Radford, Jong Wook Kim, Tao Xu, Greg Brock- man, Christine McLeavey, and Ilya Sutskever

ISCA. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brock- man, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak su- pervision. InInternational Conference on Machine Learning, ICML 2023, volume 202, pages 28492– 28518. PMLR. Chandan K. A. Reddy, Vishak Gopal, and Ross Cut- ler. 2022. Dnsmos P.835: A non-intrusive pe...

work page arXiv 2023
[3]

Tianxin Xie, Shan Yang, Chenxing Li, Dong Yu, and Li Liu

IEEE. Tianxin Xie, Shan Yang, Chenxing Li, Dong Yu, and Li Liu. 2025. Emosteer-tts: Fine- grained and training-free emotion-controllable text- to-speech via activation steering.arXiv preprint arXiv:2508.03543. Dongchao Yang, Songxiang Liu, Rongjie Huang, Chao Weng, and Helen Meng. 2024. Instructtts: Modelling expressive TTS in discrete latent space with n...

work page arXiv 2025
[4]

Indextts2: A breakthrough in emotionally expressive and duration-controlled auto-regressive zero-shot text-to-speech,

Emovoice: Llm-based emotional text-to- speech model with freestyle text prompting. InPro- ceedings of the 33rd ACM International Conference on Multimedia, page 10748–10757. ACM. Kun Zhou, Berrak Sisman, Rui Liu, and Haizhou Li. 2022. Emotional voice conversion: Theory, databases and ESD.Speech Commun., 137:1–18. Siyi Zhou, Yiquan Zhou, Yi He, Xun Zhou, Ji...

work page arXiv 2022
[5]

To evaluate the smoothness of transitions in both emotion and speaking rate, we adopt the DNSMOS Pro10 (Cumlin et al., 2024), referred as DNSM

ASR model to calculate Word Error Rate (WER)8, while for Chinese audio, we utilize a Paraformer (Gao et al., 2022) ASR model to cal- culate Character Error Rate (CER) for Chinese to quantify transcription accuracy9. To evaluate the smoothness of transitions in both emotion and speaking rate, we adopt the DNSMOS Pro10 (Cumlin et al., 2024), referred as DNS...

work page arXiv 2022
[7]

then", "now

Text Utterance: - Length: 15-25 words (corresponding to 5-10 seconds of speech). - The text MUST contain all emotions in the given sequence, each clearly identifiable. - Emotional transitions MUST be conveyed through changes in language tone, imagery, internal reactions, or perspective. - CRITICAL: Do NOT use explicit temporal markers such as "then", "now...

work page
[8]

Wind whispered through the parched cornstalks, its voice fraying like worn silk

Text Category Constraint: ${ - vivid_descriptive: Vivid descriptive sentences (novel prose style). Example: "Wind whispered through the parched cornstalks, its voice fraying like worn silk." | - emotional_dialogue: Emotionally charged dialogue excerpts (natural spoken lines). Example: "I’ve asked you three times! Why is the door still locked?" | - observa...

work page
[9]

text": "<generated single-sentence utterance>

Output Format: Provide your response in the following JSON structure ONLY: { "text": "<generated single-sentence utterance>", "text_category": "${text_category: vivid_descriptive | emotional_dialogue | observational_phrase}$" } Examples: Example 1 ${Example: Vivid Descriptive Input Emotion Sequence:

work page
[10]

text": "Warm light drifts around me, a sudden sharp gust jolts the calm, and a muted heaviness settles quietly over my thoughts

Sad Output: { "text": "Warm light drifts around me, a sudden sharp gust jolts the calm, and a muted heaviness settles quietly over my thoughts.", "text_category": "vivid_descriptive" } }$ ... Now generate a text utterance for the given emotion sequence. Listing 1: Example prompt for generating content text with emotion shifts using GPT-4o. Role: You are a...

work page
[11]

${Emotion_3}$ Requirements:

work page
[12]

- CRITICAL: Segments MUST correspond to the emotion sequence IN ORDER

Segmentation Rules: - Produce EXACTLY the same number of segments as emotions in the sequence. - CRITICAL: Segments MUST correspond to the emotion sequence IN ORDER. The first segment maps to the first emotion, the second to the second emotion, etc. - Each segment MUST be a continuous span from the original text. Do NOT rewrite, reorder, omit, or add any ...

work page
[13]

- The description should focus on auditory characteristics (e.g., pitch, intensity, pacing), not on events or semantics

Emotion Description (for TTS prosody reference): - Provide a short vocal-affect description (5-15 words) focusing on auditory qualities. - The description should focus on auditory characteristics (e.g., pitch, intensity, pacing), not on events or semantics. - The description MUST align with the assigned emotion

work page
[14]

2.4"). Output Format (JSON ONLY): {

Speaking Time Estimation: - Estimate speaking duration in seconds using the guideline: 0.18-0.30 seconds per word as a baseline. - The estimated duration should also reflect the emotional tone of the segment, as different emotions naturally influence speaking pace (e.g., excited or tense delivery tends to be quicker, while somber or reflective delivery te...

work page

[1] [1]

CosyVoice: A Scalable Multilingual Zero-shot Text-to-speech Synthesizer based on Supervised Semantic Tokens

F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2025, pages 6255–6271. Association for Com- putational Linguistics. Fredrik Cumlin, Xinyu Liang, Victor Ungureanu, Chan- dan K. A. Reddy, Christian Schüldt, an...

work page internal anchor Pith review Pith/arXiv arXiv 2025

[2] [2]

Alec Radford, Jong Wook Kim, Tao Xu, Greg Brock- man, Christine McLeavey, and Ilya Sutskever

ISCA. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brock- man, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak su- pervision. InInternational Conference on Machine Learning, ICML 2023, volume 202, pages 28492– 28518. PMLR. Chandan K. A. Reddy, Vishak Gopal, and Ross Cut- ler. 2022. Dnsmos P.835: A non-intrusive pe...

work page arXiv 2023

[3] [3]

Tianxin Xie, Shan Yang, Chenxing Li, Dong Yu, and Li Liu

IEEE. Tianxin Xie, Shan Yang, Chenxing Li, Dong Yu, and Li Liu. 2025. Emosteer-tts: Fine- grained and training-free emotion-controllable text- to-speech via activation steering.arXiv preprint arXiv:2508.03543. Dongchao Yang, Songxiang Liu, Rongjie Huang, Chao Weng, and Helen Meng. 2024. Instructtts: Modelling expressive TTS in discrete latent space with n...

work page arXiv 2025

[4] [4]

Indextts2: A breakthrough in emotionally expressive and duration-controlled auto-regressive zero-shot text-to-speech,

Emovoice: Llm-based emotional text-to- speech model with freestyle text prompting. InPro- ceedings of the 33rd ACM International Conference on Multimedia, page 10748–10757. ACM. Kun Zhou, Berrak Sisman, Rui Liu, and Haizhou Li. 2022. Emotional voice conversion: Theory, databases and ESD.Speech Commun., 137:1–18. Siyi Zhou, Yiquan Zhou, Yi He, Xun Zhou, Ji...

work page arXiv 2022

[5] [5]

To evaluate the smoothness of transitions in both emotion and speaking rate, we adopt the DNSMOS Pro10 (Cumlin et al., 2024), referred as DNSM

ASR model to calculate Word Error Rate (WER)8, while for Chinese audio, we utilize a Paraformer (Gao et al., 2022) ASR model to cal- culate Character Error Rate (CER) for Chinese to quantify transcription accuracy9. To evaluate the smoothness of transitions in both emotion and speaking rate, we adopt the DNSMOS Pro10 (Cumlin et al., 2024), referred as DNS...

work page arXiv 2022

[6] [7]

then", "now

Text Utterance: - Length: 15-25 words (corresponding to 5-10 seconds of speech). - The text MUST contain all emotions in the given sequence, each clearly identifiable. - Emotional transitions MUST be conveyed through changes in language tone, imagery, internal reactions, or perspective. - CRITICAL: Do NOT use explicit temporal markers such as "then", "now...

work page

[7] [8]

Wind whispered through the parched cornstalks, its voice fraying like worn silk

Text Category Constraint: ${ - vivid_descriptive: Vivid descriptive sentences (novel prose style). Example: "Wind whispered through the parched cornstalks, its voice fraying like worn silk." | - emotional_dialogue: Emotionally charged dialogue excerpts (natural spoken lines). Example: "I’ve asked you three times! Why is the door still locked?" | - observa...

work page

[8] [9]

text": "<generated single-sentence utterance>

Output Format: Provide your response in the following JSON structure ONLY: { "text": "<generated single-sentence utterance>", "text_category": "${text_category: vivid_descriptive | emotional_dialogue | observational_phrase}$" } Examples: Example 1 ${Example: Vivid Descriptive Input Emotion Sequence:

work page

[9] [10]

text": "Warm light drifts around me, a sudden sharp gust jolts the calm, and a muted heaviness settles quietly over my thoughts

Sad Output: { "text": "Warm light drifts around me, a sudden sharp gust jolts the calm, and a muted heaviness settles quietly over my thoughts.", "text_category": "vivid_descriptive" } }$ ... Now generate a text utterance for the given emotion sequence. Listing 1: Example prompt for generating content text with emotion shifts using GPT-4o. Role: You are a...

work page

[10] [11]

${Emotion_3}$ Requirements:

work page

[11] [12]

- CRITICAL: Segments MUST correspond to the emotion sequence IN ORDER

Segmentation Rules: - Produce EXACTLY the same number of segments as emotions in the sequence. - CRITICAL: Segments MUST correspond to the emotion sequence IN ORDER. The first segment maps to the first emotion, the second to the second emotion, etc. - Each segment MUST be a continuous span from the original text. Do NOT rewrite, reorder, omit, or add any ...

work page

[12] [13]

- The description should focus on auditory characteristics (e.g., pitch, intensity, pacing), not on events or semantics

Emotion Description (for TTS prosody reference): - Provide a short vocal-affect description (5-15 words) focusing on auditory qualities. - The description should focus on auditory characteristics (e.g., pitch, intensity, pacing), not on events or semantics. - The description MUST align with the assigned emotion

work page

[13] [14]

2.4"). Output Format (JSON ONLY): {

Speaking Time Estimation: - Estimate speaking duration in seconds using the guideline: 0.18-0.30 seconds per word as a baseline. - The estimated duration should also reflect the emotional tone of the segment, as different emotions naturally influence speaking pace (e.g., excited or tense delivery tends to be quicker, while somber or reflective delivery te...

work page