Pith. sign in

REVIEW 5 major objections 6 minor 30 references

XEmoRAG: Cross-Lingual Emotion Transfer with Controllable Intensity Using Retrieval-Augmented Generation

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read XEmoRAG claims that a Chinese emotional recording can set the emotion, intensity, and speaker timbre of synthesized Thai speech without parallel emotional data or explicit emotion labels.

desk verdict Solid systems paper; Chinese-to-Thai zero-shot emotion transfer is new, but intensity control and cross-lingual alignment evidence are both under-developed. read the letter →

arxiv 2508.07302 v2 pith:YO3IO7F2 submitted 2025-08-10 eess.AS

Tianlun Zuo , Jingbin Hu , Yuke Li , Xinfa Zhu , Hai Li , Ying Yan , Junhui Liu , Danming Xie
show 1 more author
Lei Xie
This is my paper
classification eess.AS
keywords cross-lingualemotiontransfertext-to-speechretrieval-augmentedgenerationintensitycontrolzero-shotspeechsynthesisflowmatchingChinese-to-ThaiLLM-basedTTS
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that a Chinese emotional recording—with no emotion label and no paired Thai recording—can be used to synthesize Thai speech that carries the same emotion. It argues this is possible because emotional content can be embedded in a language-agnostic space, so a Thai utterance with matching emotion can be retrieved and used as a prosody prompt. The authors build this into an LLM-based text-to-speech system with a flow-matching aligner, and they report that emotional similarity, intelligibility, and speaker similarity all improve over strong baselines. If correct, this removes the main data bottleneck for expressive cross-lingual synthesis and adds weak-to-strong intensity control for free.

What carries the argument

Two mechanisms carry the argument. First, the RAG module: a pre-trained Emo2Vec model maps Chinese and Thai speech into a supposedly language-agnostic emotional embedding space; K-means clustering over the Thai prompt database and an intensity filter (weak/normal/strong) select the prosody prompt at inference. Second, flow-matching alignment: a transformer-based 1D U-Net vector-field predictor, conditioned on speaker reference embeddings, maps the discrete X-Codec2 tokens to mel-spectrograms, correcting pitch and duration mismatches so the transfer sounds natural rather than foreign-accented.

What would settle it

Take a held-out set of Chinese emotional references with known Thai emotion labels. Measure whether Emo2Vec nearest-neighbor retrieval from Chinese to Thai matches the labeled emotion significantly better than chance, and repeat for a third language not seen in tuning. If cross-lingual retrieval accuracy is at chance—or if replacing the Thai prompt pool with another language's pool does not degrade the emotion-similarity rating in a listener test—the central transfer mechanism fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that zero-shot cross-lingual emotion transfer can be decomposed into retrieval plus alignment instead of requiring parallel emotional data. Emo2Vec embeds both Chinese and Thai speech in a shared space where the nearest Thai neighbor of a Chinese emotional reference is, the paper claims, the right emotional prompt. That retrieved Thai utterance—audio and text—is prepended to the target text and fed to Llasa, an autoregressive LLM that predicts X-Codec2 tokens; a flow-matching U-Net then aligns those tokens to mel-spectrograms while conditioning on a speaker reference, and DSPGAN converts them to audio. The paper reports that removing either the retr

Load-bearing premise

The load-bearing assumption is that Emo2Vec's embedding space puts the same emotion close together across Chinese and Thai, so the nearest Thai utterance to a Chinese reference really carries that emotion; this is supported in the paper mainly by a qualitative t-SNE plot and retrieval accuracy on one manually labeled set, not by a direct cross-lingual alignment metric.

Editorial extensions

If this is right

  • Chinese-to-Thai emotional dubbing becomes a zero-shot task: any Chinese emotional clip can drive Thai delivery without a Thai emotional recording of the same line or speaker.
  • Low-resource languages need only a small pool of emotional Thai utterances for prompting, not a parallel bilingual emotional corpus.
  • Intensity control is obtained without retraining: choosing weak, normal, or strong restricts retrieval to that subset of prompts.
  • Speaker identity and naturalness survive transfer because flow matching conditions on a speaker reference and aligns pitch and duration.
  • The same recipe is claimed to port to other target languages because the emotional embedding is language-agnostic.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A decisive next test the paper does not run is a direct labeled retrieval metric across several language pairs; that would convert the t-SNE argument into a number.
  • Intensity control could be made continuous by interpolating between retrieved embeddings instead of restricting to binned subsets, a natural extension the paper leaves open.
  • Because the evaluation uses a single Chinese source speaker, the emotion-transfer result may partly reflect speaker-prosody transfer; testing multiple source speakers would separate those factors.
  • If the embedding-space alignment holds, the pipeline should work for other typologically distant pairs such as English-to-Thai or Chinese-to-Arabic with no architectural change.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper presents XEmoRAG, a retrieval-augmented LLM-based TTS system for zero-shot emotion transfer from Chinese reference speech to Thai synthesis. The system uses Emo2Vec embeddings to retrieve emotionally matched Thai prompt utterances from a curated database, employs K-means clustering with intensity-conditioned subsets for retrieval, and uses flow-matching alignment to map X-Codec2 tokens to mel-spectrograms. It builds on Llasa-1B-Multilingual, with two-stage fine-tuning on Thai data. Experiments report EMOS 4.65 vs. 3.89 for DelightfulTTS, NMOS 4.38, CER 5.95%, and SIM 0.89, along with ablations and a retrieval scaling study. The central claim is that Chinese-to-Thai emotion transfer can be achieved without parallel emotional data or explicit emotion category labels, with controllable intensity.

Significance. If the central claim holds, the paper demonstrates a practical route to cross-lingual emotion transfer in a low-resource language pair by combining a pretrained LLM TTS backbone, emotion embedding retrieval, and flow-matching alignment. The system is evaluated with human MOS, objective CER/SIM, ablations, and retrieval efficiency comparisons, and a demo page is provided. The paper also clearly identifies the resource constraints it addresses. However, the load-bearing assumption that Emo2Vec embeddings align emotions across Chinese and Thai is only weakly evidenced, and the claimed intensity control is not experimentally validated. These gaps currently prevent full confidence in the results.

major comments (5)
  1. [Section III-C, Fig. 2, Table II] The central mechanism reduces to: Chinese reference -> Emo2Vec embedding -> nearest Thai utterance -> prosody prompt. Thus the output emotion is inherited almost entirely from the retrieved Thai prompt. The paper asserts language-agnostic Emo2Vec embeddings, but this is supported only by a qualitative t-SNE plot and retrieval accuracy on a manually labeled set. Retrieval accuracy is not a direct cross-lingual alignment metric: it does not report per-emotion confusion, chance-level performance, or a within-language baseline, and the best accuracy of 86.3% implies roughly one in seven retrieved prompts is emotionally mismatched. The authors should provide a quantitative alignment evaluation, such as a cross-lingual emotion classification probe, per-emotion retrieval recall, a within-language vs. cross-lingual retrieval comparison, and a sensitivity analysis of EMOS to retrieval errors.
  2. [Section III-C and Section IV] Intensity control (weak/normal/strong) is claimed in the title and abstract and implemented by restricting retrieval to intensity subsets, but no experiment evaluates it. There is no MOS, no acoustic analysis, and no demonstration that the synthesized output actually varies in perceived emotional intensity. This is a load-bearing claim for the paper. Please add perceptual and/or acoustic evaluation across intensity levels and clarify whether intensity annotations are available at inference. This also bears on the claim of operating 'without explicit emotion labels.'
  3. [Section IV-C, Table I] Typhoon2-Audio 8B Instruct is listed as a baseline but receives no subjective evaluation; only CER is reported. Consequently, the claim that XEmoRAG outperforms baselines in expressiveness and naturalness is supported only against DelightfulTTS. Either include EMOS/NMOS for this baseline or qualify the claim accordingly.
  4. [Section IV-C, Table I] Objective metrics CER and SIM are reported as single numbers without error bars, confidence intervals, or significance tests. The CER difference of 5.95% vs. 7.33% could be meaningful, but the paper does not establish this statistically. Please provide multiple runs or bootstrap intervals, and clarify the ASR language used for CER and the reference speaker used for SIM.
  5. [Section IV-D, Table II] The K-means retrieval is not fully specified: the number of clusters K is not reported. Retrieval accuracy is computed on a manually labeled test set, but the emotion label taxonomy, label counts, and annotator reliability are not described, making the 86.3% figure difficult to interpret. Additionally, the text states that emotion labels in the Thai dataset are unseen during inference, yet inference uses intensity labels; please clarify exactly which labels are unseen.
minor comments (6)
  1. [Fig. 1 vs. Section III-C] The figure uses 'Emotion2Vec' while the text consistently uses 'Emo2Vec'; unify the terminology.
  2. [Section IV-C] Naturalness of Thai speech is rated by 15 listeners, of whom only 5 are native Thai. If the Chinese listeners are not fluent in Thai, the NMOS results should be interpreted with caution; please state their language proficiency.
  3. [Table I] The 'w/o RAG Module' row has a blank SIM value; report it or explain why it is omitted.
  4. [References] References [3] and [13] are the same paper (DelightfulTTS); consolidate to avoid duplication.
  5. [Section IV-C] The phrase 'significantly outperforming' is used without a statistical test. Please either add significance testing or soften the wording.
  6. [Section III-B] The description of the flow-matching U-Net is terse; specifying the upsampling ratio '1.6 : 1' and the role of the 1D U-Net more precisely would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the system is evaluated against external baselines and human perception; internal components are ablated rather than fitted to the target metric.

full rationale

XEmoRAG is an empirical systems paper rather than a formal derivation. Its central claim (Chinese-to-Thai zero-shot emotion transfer with controllable intensity) rests on a pipeline of pretrained components (Emo2Vec retrieval, Llasa generation, flow-matching alignment) and is validated by external baselines (DelightfulTTS, Typhoon2-Audio), human MOS judgments, CER, SIM, and ablations. No equation is defined in terms of the quantity it is said to predict, and no fitted parameter is renamed as a prediction. The reliance on the authors' prior models (Llasa, X-Codec2, DSPGAN, and related-work papers [15], [19]) is normal use of published systems; those citations are not invoked as a uniqueness theorem or to forbid alternatives, and the main contribution is evaluated against independent perceptual and objective metrics. The language-agnostic property of Emo2Vec is an empirical premise supported by t-SNE visualization and retrieval accuracy on a manually labeled set; even if that evidence is weak (retrieval accuracy is not a direct alignment metric), it is an assumption subject to external falsification, not a conclusion that reduces to its own input. Retrieval accuracy is reported as an evaluation, not used to fit the synthesis model. Hence there is no identifiable circular step within the paper's derivation chain.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The system is empirical; it introduces no new theoretical entity. The main unproven assumptions are the language-agnostic property of the emotion embeddings, the correctness of the Thai database annotations, and the reliability of the pre-trained components.

free parameters (2)
  • Number of clusters K in K-means retrieval
    The clustering-based retrieval, which is the default configuration, requires selecting K. The paper does not state K or show sensitivity to it, yet retrieval accuracy (Table II) depends on this choice.
  • Intensity level annotations (weak/normal/strong) = weak, normal, strong
    The intensity control is implemented by gating retrieval to predefined subsets. The placement of these levels is not empirically calibrated or validated; no experiment demonstrates that different intensity selections produce perceptibly different output.
assumptions (3)
  • domain assumption Emo2Vec embeddings are language-agnostic for Chinese and Thai emotions.
    The RAG retrieval relies on the assumption that emotion expressed in Chinese and Thai maps to similar regions in Emo2Vec embedding space. The paper supports this only with a t-SNE visualization (Fig. 2) and retrieval accuracy (Table II).
  • domain assumption The Thai emotional database's annotations (emotion categories and intensity levels) are accurate and representative.
    The retrieval pool and the intensity control depend on correct annotation of the Thai utterances; no annotation protocol or inter-annotator agreement is reported.
  • domain assumption Pre-trained models (Llasa, X-Codec2, DSPGAN, WeSpeaker, Zipformer) perform as documented and their internal biases do not invalidate the evaluation.
    The system's behavior depends on the reliability of these external models, which are taken as given without validation in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of XEmoRAG: Cross-Lingual Emotion Transfer with Controllable Intensity Using Retrieval-Augmented Generation." pith.science (2026). https://pith.science/paper/YO3IO7F2

@misc{pith2026250807302,
  author       = {Pith},
  title        = {Pith review of: XEmoRAG: Cross-Lingual Emotion Transfer with Controllable Intensity Using Retrieval-Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YO3IO7F2}},
  note         = {Machine review of arXiv:2508.07302}
}
read the original abstract

Zero-shot emotion transfer in cross-lingual speech synthesis refers to generating speech in a target language, where the emotion is expressed based on reference speech from a different source language. However, this task remains challenging due to the scarcity of parallel multilingual emotional corpora, the presence of foreign accent artifacts, and the difficulty of separating emotion from language-specific prosodic features. In this paper, we propose XEmoRAG, a novel framework to enable zero-shot emotion transfer from Chinese to Thai using a large language model (LLM)-based model, without relying on parallel emotional data. XEmoRAG extracts language-agnostic emotional embeddings from Chinese speech and retrieves emotionally matched Thai utterances from a curated emotional database, enabling controllable emotion transfer without explicit emotion labels. Additionally, a flow-matching alignment module minimizes pitch and duration mismatches, ensuring natural prosody. It also blends Chinese timbre into the Thai synthesis, enhancing rhythmic accuracy and emotional expression, while preserving speaker characteristics and emotional consistency. Experimental results show that XEmoRAG synthesizes expressive and natural Thai speech using only Chinese reference audio, without requiring explicit emotion labels. These results highlight XEmoRAG's capability to achieve flexible and low-resource emotional transfer across languages. Our demo is available at https://tlzuo-lesley.github.io/Demo-page/ .

Figures

Figures reproduced from arXiv: 2508.07302 by the authors.

Figure 1
Figure 1. Cross-lingual emotional speech synthesis system based on retrieval-augmented generation and flow matching. The user [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualization of emotional representations ex [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 22 canonical work pages

  1. [1]

    Fastspeech 2: Fast and high-quality end-to-end text to speech,

    Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu, “Fastspeech 2: Fast and high-quality end-to-end text to speech,” arXiv preprint arXiv:2006.04558 , 2020

    arXiv 2006

  2. [2]

    Adaspeech: Adaptive text to speech for custom voice,

    Mingjian Chen, Xu Tan, Bohan Li, Yanqing Liu, Tao Qin, Sheng Zhao, and Tie-Yan Liu, “Adaspeech: Adaptive text to speech for custom voice,” arXiv preprint arXiv:2103.00993 , 2021

    arXiv 2021

  3. [4]

    Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,

    Jaehyeon Kim, Jungil Kong, and Juhee Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” in International Conference on Machine Learning. PMLR, 2021, pp. 5530– 5540

    work page 2021

  4. [5]

    Generspeech: Towards style transfer for generalizable out-of-domain text-to-speech,

    Rongjie Huang, Yi Ren, Jinglin Liu, Chenye Cui, and Zhou Zhao, “Generspeech: Towards style transfer for generalizable out-of-domain text-to-speech,” Advances in Neural Information Processing Systems , vol. 35, pp. 10970–10983, 2022

    work page 2022

  5. [6]

    Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,

    Ziqiang Zhang, Long Zhou, Chengyi Wang, Sanyuan Chen, Yu Wu, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al., “Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,” arXiv preprint arXiv:2303.03926 , 2023

    arXiv 2023

  6. [7]

    Xtts: a massively multilingual zero-shot text-to-speech model,

    Edresson Casanova, Kelly Davis, Eren G ¨olge, G ¨orkem G ¨oknar, Iulian Gulea, Logan Hart, Aya Aljafari, Joshua Meyer, Reuben Morais, Samuel Olayemi, et al., “Xtts: a massively multilingual zero-shot text-to-speech model,” arXiv preprint arXiv:2406.04904 , 2024

    arXiv 2024

  7. [8]

    Zero-shot Cross-lingual Voice Transfer for TTS

    Fadi Biadsy, Youzheng Chen, Isaac Elias, Kyle Kastner, Gary Wang, Andrew Rosenberg, and Bhuvana Ramabhadran, “Zero-shot cross- lingual voice transfer for tts,” arXiv preprint arXiv:2409.13910 , 2024

    work page Pith review arXiv 2024

  8. [9]

    Multilingual video dubbing—a tech- nology review and current challenges,

    Dan Bigioi and Peter Corcoran, “Multilingual video dubbing—a tech- nology review and current challenges,” Frontiers in signal processing , vol. 3, pp. 1230755, 2023

    work page 2023

Show all 30 references
  1. [10]

    Dse-tts: Dual speaker embedding for cross-lingual text-to-speech,

    Sen Liu, Yiwei Guo, Chenpeng Du, Xie Chen, and Kai Yu, “Dse-tts: Dual speaker embedding for cross-lingual text-to-speech,” arXiv preprint arXiv:2306.14145, 2023

    2023 arXiv

  2. [11]

    Diclet-tts: Diffusion model based cross- lingual emotion transfer for text-to-speech – a study between english and mandarin,

    Tao Li, Chenxu Hu, Jian Cong, Xinfa Zhu, Jingbei Li, Qiao Tian, Yuping Wang, and Lei Xie, “Diclet-tts: Diffusion model based cross- lingual emotion transfer for text-to-speech – a study between english and mandarin,” 2023

    2023

  3. [12]

    Llasa: Scaling train- time and inference-time compute for llama-based speech synthesis,

    Zhen Ye, Xinfa Zhu, Chi-Min Chan, Xinsheng Wang, Xu Tan, Jiahe Lei, Yi Peng, Haohe Liu, Yizhu Jin, Zheqi DAI, et al., “Llasa: Scaling train- time and inference-time compute for llama-based speech synthesis,” arXiv preprint arXiv:2502.04128 , 2025

    2025 arXiv

  4. [13]

    Delightfultts: The microsoft speech synthesis system for blizzard challenge 2021,

    Yanqing Liu, Zhihang Xu, Gang Wang, Kuan Chen, Bohan Li, Xu Tan, Jinzhu Li, Lei He, and Sheng Zhao, “Delightfultts: The microsoft speech synthesis system for blizzard challenge 2021,” arXiv preprint arXiv:2110.12612, 2021

    2021 arXiv

  5. [14]

    iemotts: Toward robust cross- speaker emotion transfer and control for speech synthesis based on disentanglement between prosody and timbre,

    Guangyan Zhang, Ying Qin, Wenjie Zhang, Jialun Wu, Mei Li, Yutao Gai, Feijun Jiang, and Tan Lee, “iemotts: Toward robust cross- speaker emotion transfer and control for speech synthesis based on disentanglement between prosody and timbre,” 2023

    2023

  6. [15]

    Zero-shot emotion transfer for cross-lingual speech synthesis,

    Yuke Li, Xinfa Zhu, Yi Lei, Hai Li, Junhui Liu, Danming Xie, and Lei Xie, “Zero-shot emotion transfer for cross-lingual speech synthesis,” in 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8

    2023

  7. [16]

    Zet- speech: Zero-shot adaptive emotion-controllable text-to-speech synthesis with diffusion and style-based models,

    Minki Kang, Wooseok Han, Sung Ju Hwang, and Eunho Yang, “Zet- speech: Zero-shot adaptive emotion-controllable text-to-speech synthesis with diffusion and style-based models,” 2023

    2023

  8. [17]

    Learning to speak fluently in a foreign language: Multilingual speech synthesis and cross-language voice cloning,

    Yu Zhang, Ron J Weiss, Heiga Zen, Yonghui Wu, Zhifeng Chen, RJ Skerry-Ryan, Ye Jia, Andrew Rosenberg, and Bhuvana Ramabhad- ran, “Learning to speak fluently in a foreign language: Multilingual speech synthesis and cross-language voice cloning,” arXiv preprint arXiv:1907.04448, 2019

    1907 arXiv

  9. [19]

    Metts: Multilingual emotional text-to-speech by cross- speaker and cross-lingual emotion transfer,

    Xinfa Zhu, Yi Lei, Tao Li, Yongmao Zhang, Hongbin Zhou, Heng Lu, and Lei Xie, “Metts: Multilingual emotional text-to-speech by cross- speaker and cross-lingual emotion transfer,” 2023

    2023

  10. [20]

    Cosyvoice 3: Towards in-the-wild speech generation via scaling-up and post-training,

    Zhihao Du, Changfeng Gao, Yuxuan Wang, Fan Yu, Tianyu Zhao, Hao Wang, Xiang Lv, Hui Wang, Chongjia Ni, Xian Shi, Keyu An, Guanrou Yang, Yabin Li, Yanni Chen, Zhifu Gao, Qian Chen, Yue Gu, Mengzhe Chen, Yafeng Chen, Shiliang Zhang, Wen Wang, and Jieping Ye, “Cosyvoice 3: Toward...

    2025

  11. [21]

    Maskgct: Zero-shot text-to-speech with masked generative codec transformer,

    Yuancheng Wang, Haoyue Zhan, Liwei Liu, Ruihong Zeng, Haotian Guo, Jiachen Zheng, Qiang Zhang, Xueyao Zhang, Shunsi Zhang, and Zhizheng Wu, “Maskgct: Zero-shot text-to-speech with masked generative codec transformer,” 2024

    2024

  12. [22]

    Glm-4-voice: Towards intelligent and human-like end-to-end spoken chatbot,

    Aohan Zeng, Zhengxiao Du, Mingdao Liu, Kedong Wang, Shengmin Jiang, Lei Zhao, Yuxiao Dong, and Jie Tang, “Glm-4-voice: Towards intelligent and human-like end-to-end spoken chatbot,” 2024

    2024

  13. [23]

    Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,

    Ziqiang Zhang, Long Zhou, Chengyi Wang, Sanyuan Chen, Yu Wu, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, Lei He, Sheng Zhao, and Furu Wei, “Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,” 2023

    2023

  14. [24]

    Retrieval- augmented generation for knowledge-intensive nlp tasks,

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K ¨uttler, Mike Lewis, Wen tau Yih, Tim Rockt ¨aschel, Sebastian Riedel, and Douwe Kiela, “Retrieval- augmented generation for knowledge-intensive nlp tasks,” 2021

    2021

  15. [25]

    Autostyle-tts: Retrieval-augmented generation based automatic style matching text-to-speech synthesis,

    Dan Luo, Chengyuan Ma, Weiqin Li, Jun Wang, Wei Chen, and Zhiyong Wu, “Autostyle-tts: Retrieval-augmented generation based automatic style matching text-to-speech synthesis,” arXiv preprint arXiv:2504.10309, 2025

    2025 arXiv

  16. [26]

    Wavrag: Audio-integrated retrieval augmented generation for spoken dialogue models,

    Yifu Chen, Shengpeng Ji, Haoxiao Wang, Ziqing Wang, Siyu Chen, Jinzheng He, Jin Xu, and Zhou Zhao, “Wavrag: Audio-integrated retrieval augmented generation for spoken dialogue models,” 2025

    2025

  17. [27]

    Codec does matter: Exploring the semantic shortcoming of codec for audio language model,

    Zhen Ye, Peiwen Sun, Jiahe Lei, Hongzhan Lin, Xu Tan, Zheqi Dai, Qiuqiang Kong, Jianyi Chen, Jiahao Pan, Qifeng Liu, et al., “Codec does matter: Exploring the semantic shortcoming of codec for audio language model,” in Proceedings of the AAAI Conference on Artificial Intellige...

    2025

  18. [28]

    Emo2vec: Learning emotional embeddings via multi-emotion cate- gory,

    Shuo Wang, Aishan Maoliniyazi, Xinle Wu, and Xiaofeng Meng, “Emo2vec: Learning emotional embeddings via multi-emotion cate- gory,” ACM Transactions on Internet Technology (TOIT) , vol. 20, no. 2, pp. 1–17, 2020

    2020

  19. [29]

    Dspgan: A gan-based universal vocoder for high-fidelity tts by time-frequency domain supervision from dsp,

    Kun Song, Yongmao Zhang, Yi Lei, Jian Cong, Hanzhao Li, Lei Xie, Gang He, and Jinfeng Bai, “Dspgan: A gan-based universal vocoder for high-fidelity tts by time-frequency domain supervision from dsp,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Si...

    2023

  20. [30]

    Typhoon 2: A family of open text and multi- modal thai large language models,

    Kunat Pipatanakul, Potsawee Manakul, Natapong Nitarach, Warit Siri- chotedumrong, Surapon Nonesung, Teetouch Jaknamon, Parinthapat Pengpun, Pittawat Taveekitworachai, Adisai Na-Thalang, Sittipong Sri- paisarnmongkol, et al., “Typhoon 2: A family of open text and multi- modal t...

    2024 arXiv

  21. [31]

    Wespeaker: A research and production oriented speaker embedding learning toolkit,

    Hongji Wang, Chengdong Liang, Shuai Wang, Zhengyang Chen, Binbin Zhang, Xu Xiang, Yanlei Deng, and Yanmin Qian, “Wespeaker: A research and production oriented speaker embedding learning toolkit,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal...

    2023

  22. [32]

    Zipformer: A faster and better encoder for automatic speech recognition,

    Zengwei Yao, Liyong Guo, Xiaoyu Yang, Wei Kang, Fangjun Kuang, Yifan Yang, Zengrui Jin, Long Lin, and Daniel Povey, “Zipformer: A faster and better encoder for automatic speech recognition,” arXiv preprint arXiv:2310.11230, 2023

    2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.