Pith. sign in

REVIEW 4 major objections 6 minor 31 references

A speech synthesizer that tracks word-level dialogue interactions produces more natural conversational prosody.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-05 04:28 UTC pith:PN2SRYNU

load-bearing objection Real architectural step for CSS, but the headline empirical claim is only as strong as the unverified baseline protocol. the 4 major comments →

arxiv 2509.06074 v1 pith:PN2SRYNU submitted 2025-09-07 cs.CL

Multimodal Fine-grained Context Interaction Graph Modeling for Conversational Speech Synthesis

classification cs.CL
keywords conversational speech synthesismultimodal dialogue historysemantic interaction graphprosody interaction graphword-level context modelinggraph neural networkDailyTalkFastSpeech 2
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Conversational speech synthesis needs to know not just what was said earlier, but which words mattered and how they were spoken. This paper claims that modeling those word-level effects explicitly—through two interaction graphs, one for meaning and one for prosody—makes synthesized responses sound more natural and more expressively appropriate. The system builds word-level text and speech nodes for the dialogue history, propagates their influence forward through the dialogue, and feeds the pooled interaction features into a FastSpeech-2-style synthesizer. On the DailyTalk dataset, raters preferred it over seven prior systems on naturalness and prosody, and removing either graph lowers the scores. The implied lesson is that fine-grained cross-modal context, not just utterance summaries, carries conversational prosody.

Core claim

MFCIG-CSS constructs two graphs over the dialogue history. A semantic interaction graph links word-level text and speech nodes to an utterance-level text backbone, so word meanings and delivery update each later turn's meaning. A prosody interaction graph uses the same word nodes against a speech-utterance backbone to update later prosody. Each graph ends in a pooling node that compresses the history into one semantic feature and one prosody feature, both fed to the synthesizer. On DailyTalk, the paper reports listener-rated naturalness 3.980 vs 3.858 for the best baseline, listener-rated prosody 3.899 vs 3.795, and lower pitch and mel-cepstral error. Ablations removing either graph lower th

What carries the argument

The two multimodal fine-grained interaction graphs are the central object. In each graph, word-level text embeddings and word-level speech/prosody embeddings act as nodes, and an utterance-level backbone carries information forward through the dialogue; a GraphSAGE convolution aggregates word nodes into the next utterance's node. The semantic graph uses a text backbone, while the prosody graph uses a speech backbone. At the end of each backbone sits a special interaction node that pools all utterance-level nodes, yielding two context vectors—one semantic, one prosodic—that condition the synthesizer. The design's core move is to make word-level modalities the carriers of influence across turn

Load-bearing premise

The central claim assumes the baseline scores were produced with the same DailyTalk split, the same evaluation protocol, and the same listener pool as the proposed system; the paper does not state that these conditions were matched.

What would settle it

Run a controlled replication: retrain the proposed system and all seven baselines on the identical DailyTalk train/validation/test split and have the same listener group rate all systems in one session. If the naturalness and prosody gaps shrink to within the reported confidence intervals, the claimed advantage is not established. A complementary behavioral test: construct dialogue pairs that share the same words but differ in the emotional prosody of a keyword, and check whether the synthesized next turn's pitch and energy shift measurably in the predicted direction.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the graph interaction is doing what the paper says, future systems can condition each new utterance on which words in prior turns carried meaning or prosodic weight, not just on whole-utterance embeddings.
  • The two-graph split means semantic context and prosodic context can be tuned independently, for example strengthening the prosody graph when expressive delivery matters more than lexical content.
  • Since the graphs produce fixed interaction feature vectors, they can be plugged into other synthesizer backbones and losses, not only the FastSpeech-2-style architecture used here.
  • The ablation pattern implies that word-level interaction modeling is not redundant with the utterance-level context encoders already present in the baselines; it adds information those encoders miss.

Where Pith is reading between the lines

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

  • The paper leaves baseline reproduction details unstated, so the exact sizes of the reported gains (0.122 naturalness, 0.104 prosody) should be treated as provisional until all systems are retrained on the same split with the same listening protocol.
  • A natural stress test the paper does not run: take the same keyword and vary the speaker's emotional delivery, then measure whether the synthesized next turn changes pitch and energy in the predicted direction; the paper's motivating wallet-versus-pen example is exactly this experiment in qualitative form.
  • Because only English two-speaker dialogues are tested, the graph's behavior on multi-party turns, overlapping speakers, or code-switching remains open; applying the same construction there would show whether word-level forward propagation generalizes beyond the current dataset.
  • A simpler alternative—replacing the hand-built graph edges with attention between all word and utterance nodes—would test whether the explicit graph structure is the active ingredient or whether any fine-grained fusion of word-level features would suffice.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. Jia et al. propose MFCIG-CSS, a conversational speech synthesis system that augments a FastSpeech 2 backbone with two graph modules: a semantic interaction graph (SIG) and a prosody interaction graph (PIG). The graphs use word-level text (TOD-BERT) and speech (Wav2Vec2.0) features and utterance-level nodes to model how words in the dialogue history influence subsequent utterances. The encoded interaction features are injected into the feature aggregator. Experiments on DailyTalk compare against seven CSS baselines and three ablations, reporting DMOS, MAE-P, MAE-E, and MCD. The authors claim the method outperforms all baselines in prosodic expressiveness; code and samples are released.

Significance. If the controlled-comparison issues are resolved, the paper makes a useful contribution: it is among the first to shift CSS context modeling from utterance-level to word-level interaction graphs, and the 0.12-0.15 DMOS gains over a strong I3-CSS baseline in the authors' setup are promising. The release of code/samples and the inclusion of ablations are strengths. However, the headline claim is currently overstated and the evaluation protocol for baselines is not documented. With additional statistical validation and a precise graph definition, the result could support the paper's conclusions.

major comments (4)
  1. [§3.3, Table 1, §2.5] The central claim requires that all baseline rows in Table 1 are produced under the same conditions as MFCIG-CSS, but Section 3.3 and Appendix A.1 only name the baseline models; the paper never states whether these numbers were rerun by the authors, taken from original papers, or produced with a shared protocol. A specific internal warning sign is the gap between Abl.Exp.3 (no graphs; N-DMOS 3.592, P-DMOS 3.512) and I3-CSS (N-DMOS 3.858, P-DMOS 3.795): since §2.5 says MFCIG-CSS uses the same synthesizer as I3-CSS, removing the two graph modules should approximate I3-CSS unless I3-CSS contains additional context modules that the ablation also removes—a difference the paper does not document. Without controlled baselines, the reported gains may reflect evaluation heterogeneity rather than the proposed graphs.
  2. [Abstract & §3.5, Table 1] The abstract's 'outperforms all baseline models' is not supported by MAE-E: MFCIG-CSS reports 0.314 and I3-CSS reports 0.310, so MFCIG-CSS is second-best on this metric. Section 3.5 correctly says 'average performance', but the abstract and §1 'outperforms all baseline models' remain overstated. Also, the text says MAE-E is '0.004 lower' than the best baseline, but it is higher (worse). Please qualify the claim (e.g., 'best on average; second-best on MAE-E') and correct the direction.
  3. [§3.4, Table 1] Subjective DMOS comparisons lack a significance test. The paper reports means with ± values (undefined whether standard error or confidence interval) for 20 raters, but no paired test between MFCIG-CSS and each baseline, no multiple-comparison correction, and no inter-rater reliability. The reported gains of 0.122 N-DMOS and 0.104 P-DMOS could be within rating noise. Please add an appropriate significance test and report inter-rater agreement (e.g., ICC or Krippendorff's alpha) and the number of rated utterances.
  4. [§2.3–2.4, Eqs. (1)–(2)] The graph construction is under-specified. The paper lists three interaction branches but does not formally define the node/edge sets or the adjacency structure of G_s and G_p; Eq. (1) writes SAGE(F_i^t, W_i^t, W_i^s), which is not the standard GraphSAGE aggregation over a neighbor set. It is unclear how word-level nodes are connected to utterance-level nodes and what 'Average Pooling(F_{1→J}^t, I_s)' pools over. Please provide a complete graph definition (nodes, edges, aggregation equations) to make the method reproducible.
minor comments (6)
  1. [Table 1] The I3-CSS row is formatted incorrectly: '0.4500.31011.47' should be separated columns. Please check all table alignment.
  2. [§2.3] MFA is used without expansion or citation at first use. Specify the forced aligner and its configuration.
  3. [§2.5] The feature aggregator is said to add I'_s and I'_p into 'P_t^C', but P_t^C is never defined. Please define it in the text or Figure 1.
  4. [§2.3 vs §2.4] The paper uses 'Wav2Vec2.0' for word-level speech features in SIG and 'Wav2Vec2.0-IEMOCAP' for utterance-level speech nodes in PIG. Clarify whether the same word-level extractor is used for both graphs and why different models are chosen.
  5. [§3.1] The data split is described only as '8:1:1'. State whether the split is by dialogue or by utterance, and confirm that no speaker or dialogue overlaps between partitions.
  6. [References] The I3-CSS baseline is cited only as an arXiv preprint (Jia and Liu, 2024). If a peer-reviewed version exists, cite it; otherwise state clearly that the numbers are from the preprint.

Circularity Check

0 steps flagged

No significant circularity; the empirical comparison is independent of the graph modules' definitions.

full rationale

The paper's central claim is empirical: MFCIG-CSS outperforms baselines on DailyTalk after training and evaluation on held-out data. The proposed SIG and PIG modules are defined by Equations (1) and (2), which specify graph encoders over word- and utterance-level features, but these equations do not define the evaluation metrics (N-DMOS, P-DMOS, MAE-P, MAE-E, MCD) in terms of the model inputs. The loss follows FastSpeech 2, so there is no fitted parameter being renamed as a prediction. The only self-citation is the I3-CSS synthesizer backbone (Jia and Liu, 2024), which is used as an architectural component; the comparison to I3-CSS is an experiment, not a derivation, and no load-bearing theoretical premise is imported from the cited paper. The concern that Table 1 baseline scores may come from different splits or rater pools is a verification/comparability risk, not circularity, because the reported numbers are not derived by construction from those baselines. Equations are model definitions, not circular reductions, and no uniqueness theorem or ansatz is smuggled in via citation. The Limitations section acknowledges scope constraints (FastSpeech 2 only, no emotion/pause features) but these are not circularity. Therefore no circular step is identifiable by the paper's own equations or citations.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No new physical or formal entities are introduced. The interaction graphs are computational constructs. The free parameters are standard hyperparameters, not fitted constants in a derivation. The load-bearing assumptions are empirical: dataset representativeness, baseline comparability, feature adequacy, encoder efficacy, and rater reliability.

free parameters (3)
  • Graph node feature dimension (256) = 256
    Hand-set to match the FastSpeech2-based synthesizer; standard dimension, not fitted to maximize the reported metrics.
  • Speaker embedding dimension (256) = 256
    Hand-set hyperparameter, identical across the system; no tuning reported against the evaluation metrics.
  • Training steps (400k) and batch size (16) = 400k steps, batch 16
    Stopping criterion and batch size chosen following prior work (FastSpeech2 / I3-CSS); not fitted to the test set.
axioms (5)
  • domain assumption The DailyTalk dataset is representative enough to evaluate conversational prosody expressiveness.
    The paper validates only on DailyTalk (~2541 dialogue pairs); conclusions about general CSS prosody assume this dataset captures the relevant phenomena (Section 3.1).
  • domain assumption Baseline results in Table 1 are directly comparable to MFCIG-CSS results.
    The paper does not state whether baselines were re-trained on the same split with the same evaluation protocol; comparability is assumed (Sections 3.3 and 3.5).
  • domain assumption Word-level prosody can be extracted from Wav2Vec2.0 features averaged over MFA-aligned word segments.
    The pipeline assumes that average-pooled Wav2Vec2.0 frames preserve word-level prosody relevant to conversational interaction (Section 2.3, SIG Construction).
  • domain assumption SAGE graph convolution encodes the intended semantic and prosody interactions from the constructed nodes.
    The paper relies on SAGE (Hamilton et al. 2017) as the message-passing layer without evidence that it captures the specific word-to-utterance interactions better than alternative encoders (Equations 1 and 2).
  • domain assumption DMOS ratings from 20 graduate students are a reliable measure of prosody without inter-rater agreement or significance testing.
    Subjective scores are reported as means with standard errors, but no statistical test or inter-rater reliability is provided (Section 3.4).

pith-pipeline@v1.3.0-alltime-deepseek · 8334 in / 9021 out tokens · 93567 ms · 2026-08-05T04:28:42.728041+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Multimodal Fine-grained Context Interaction Graph Modeling for Conversational Speech Synthesis." pith.science (2026). https://pith.science/paper/PN2SRYNU

@misc{pith2026250906074,
  author       = {Pith},
  title        = {Pith review of: Multimodal Fine-grained Context Interaction Graph Modeling for Conversational Speech Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PN2SRYNU}},
  note         = {Machine review of arXiv:2509.06074}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Conversational Speech Synthesis (CSS) aims to generate speech with natural prosody by understanding the multimodal dialogue history (MDH). The latest work predicts the accurate prosody expression of the target utterance by modeling the utterance-level interaction characteristics of MDH and the target utterance. However, MDH contains fine-grained semantic and prosody knowledge at the word level. Existing methods overlook the fine-grained semantic and prosodic interaction modeling. To address this gap, we propose MFCIG-CSS, a novel Multimodal Fine-grained Context Interaction Graph-based CSS system. Our approach constructs two specialized multimodal fine-grained dialogue interaction graphs: a semantic interaction graph and a prosody interaction graph. These two interaction graphs effectively encode interactions between word-level semantics, prosody, and their influence on subsequent utterances in MDH. The encoded interaction features are then leveraged to enhance synthesized speech with natural conversational prosody. Experiments on the DailyTalk dataset demonstrate that MFCIG-CSS outperforms all baseline models in terms of prosodic expressiveness. Code and speech samples are available at https://github.com/AI-S2-Lab/MFCIG-CSS.

Figures

Figures reproduced from arXiv: 2509.06074 by Berrak Sisman, Haizhou Li, Rui Liu, Zhenqi Jia.

Figure 1
Figure 1. Figure 1: The overview of MFCIG-CSS consists of Multimodal Fine-grained Dialogue Semantic Interaction Graph, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 22 canonical work pages · 2 internal anchors

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449--12460

  4. [4]

    Santiago Castro, Devamanyu Hazarika, Ver \'o nica P \'e rez-Rosas, Roger Zimmermann, Rada Mihalcea, and Soujanya Poria. 2019. https://doi.org/10.18653/v1/P19-1455 Towards multimodal sarcasm detection ( A n \_ O bviously \_ P erfect P aper) . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4619--4629, Flore...

  5. [5]

    Qi Chen, Mingkui Tan, Yuankai Qi, Jiaqiu Zhou, Yuanqing Li, and Qi Wu. 2022. V2c: Visual voice cloning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21242--21251

  6. [6]

    Yayue Deng, Jinlong Xue, Yukang Jia, Qifei Li, Yichen Han, Fengping Wang, Yingming Gao, Dengfeng Ke, and Ya Li. 2024. Concss: Contrastive-based context comprehension for dialogue-appropriate prosody in conversational speech synthesis. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10706--10710. IEEE

  7. [7]

    Yasser El Miedany and Yasser El Miedany. 2019. Virtual reality and augmented reality. Rheumatology teaching: the art and science of medical education, pages 403--427

  8. [8]

    Haohan Guo, Shaofei Zhang, Frank K Soong, Lei He, and Lei Xie. 2021. Conversational end-to-end tts for voice agents. In 2021 IEEE Spoken Language Technology Workshop (SLT), pages 403--409. IEEE

  9. [9]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems, 30

  10. [10]

    Yifan Hu, Rui Liu, Guanglai Gao, and Haizhou Li. 2024. Fctalker: Fine and coarse grained context modeling for expressive conversational speech synthesis. In 2024 IEEE 14th International Symposium on Chinese Spoken Language Processing (ISCSLP), pages 299--303. IEEE

  11. [11]

    Mahyuzie Jenal, Athira Nabilla Omar, Muhammad Azizi Aswad Hisham, Wan Najmi Wan Mohd Noh, and Zul Adib Izzuddin Razali. 2022. Smart home controlling system. Journal of Electronic Voltage and Application, 3(1):92--104

  12. [12]

    Zhenqi Jia and Rui Liu. 2024. Intra-and inter-modal context interaction modeling for conversational speech synthesis. arXiv preprint arXiv:2412.18733

  13. [13]

    Robert Kubichek. 1993. Mel-cepstral distance measure for objective speech quality assessment. In Proceedings of IEEE pacific rim conference on communications computers and signal processing, volume 1, pages 125--128. IEEE

  14. [14]

    Keon Lee, Kyumin Park, and Daeyoung Kim. 2023. Dailytalk: Spoken dialogue dataset for conversational text-to-speech. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1--5. IEEE

  15. [15]

    Deyi Li, Jialun Yin, Tianlei Zhang, Wei Han, and Hong Bao. 2024. The four most basic elements in machine cognition. Data Intelligence, 6(2):297--319

  16. [16]

    Jingbei Li, Yi Meng, Chenyi Li, Zhiyong Wu, Helen Meng, Chao Weng, and Dan Su. 2022 a . Enhancing speaking styles in conversational text-to-speech synthesis with graph-based multi-modal context modeling. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7917--7921. IEEE

  17. [17]

    Jingbei Li, Yi Meng, Xixin Wu, Zhiyong Wu, Jia Jia, Helen Meng, Qiao Tian, Yuping Wang, and Yuxuan Wang. 2022 b . Inferring speaking styles from multi-modal conversational context by multi-scale relational graph convolutional networks. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5811--5820

  18. [18]

    Xinhang Li, Xiangyu Zhao, Jiaxing Xu, Yong Zhang, and Chunxiao Xing. 2023. Imf: interactive multimodal fusion model for link prediction. In Proceedings of the ACM Web Conference 2023, pages 2572--2580

  19. [19]

    Guan-Ting Lin, Cheng-Han Chiang, and Hung-yi Lee. 2024. https://doi.org/10.18653/v1/2024.acl-long.358 Advancing large language models to capture varied speaking styles and respond properly in spoken conversations . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6626--6642, Bangkok,...

  20. [20]

    Rui Liu, Yifan Hu, Yi Ren, Xiang Yin, and Haizhou Li. 2024 a . Emotion rendering for conversational speech synthesis with heterogeneous graph-based context modeling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18698--18706

  21. [21]

    Rui Liu, Yifan Hu, Yi Ren, Xiang Yin, and Haizhou Li. 2024 b . Generative expressive conversational speech synthesis. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4187--4196

  22. [22]

    Rui Liu, Zhenqi Jia, Feilong Bao, and Haizhou Li. 2025 a . Retrieval-augmented dialogue knowledge aggregation for expressive conversational speech synthesis. Information Fusion, 118:102948

  23. [23]

    Rui Liu, Zhenqi Jia, Feilong Bao, and Haizhou Li. 2025 b . https://doi.org/10.1016/j.inffus.2025.102948 Retrieval-augmented dialogue knowledge aggregation for expressive conversational speech synthesis . Information Fusion, page 102948

  24. [24]

    Rui Liu, Zhenqi Jia, Jie Yang, Yifan Hu, and Haizhou Li. 2024 c . Emphasis rendering for conversational text-to-speech with multi-modal multi-scale context modeling. arXiv preprint arXiv:2410.09524

  25. [25]

    Wei Peng, Yue Hu, Luxi Xing, Yuqiang Xie, Xingsheng Zhang, and Yajing Sun. 2022. https://doi.org/10.1109/ICASSP43922.2022.9747565 Modeling intention, emotion and external world in dialogue systems . In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7042--7046

  26. [26]

    Nils Reimers and Iryna Gurevych. 2019. https://doi.org/10.18653/v1/D19-1410 Sentence- BERT : Sentence embeddings using S iamese BERT -networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982--3992, Hong Kong, Chi...

  27. [27]

    Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. 2021. https://openreview.net/forum?id=piLPYqxtWuA Fastspeech 2: Fast and high-quality end-to-end text to speech . In International Conference on Learning Representations

  28. [28]

    Robert C Streijl, Stefan Winkler, and David S Hands. 2016. Mean opinion score (mos) revisited: methods and applications, limitations and alternatives. Multimedia Systems, 22(2):213--227

  29. [29]

    Minh Duc Vu, Han Wang, Zhuang Li, Jieshan Chen, Shengdong Zhao, Zhenchang Xing, and Chunyang Chen. 2024. Gptvoicetasker: Llm-powered virtual assistant for smartphone. arXiv preprint arXiv:2401.14268

  30. [30]

    Hoi, Richard Socher, and Caiming Xiong

    Chien-Sheng Wu, Steven C.H. Hoi, Richard Socher, and Caiming Xiong. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.66 TOD - BERT : Pre-trained natural language understanding for task-oriented dialogue . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 917--929, Online. Association for Computationa...

  31. [31]

    Jinlong Xue, Yayue Deng, Fengping Wang, Ya Li, Yingming Gao, Jianhua Tao, Jianqing Sun, and Jiaen Liang. 2023. M 2-ctts: End-to-end multi-scale multi-modal conversational text-to-speech synthesis. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1--5. IEEE