WorksComputer Science
EmpDPO: Enabling Emphasis Control in TTS via Direct Preference Optimization
Abstract
Word-level emphasis conveys meaning that text alone cannot express and is therefore essential for a practical text-to-speech (TTS) system. In this work, we study how to add emphasis control post hoc to a pretrained TTS model that has never been trained with emphasis annotations. We present EmpDPO, a simple, efficient, and effective two-stage recipe that adds emphasis control to a multilingual zero-shot diffusion TTS model through supervised fine-tuning followed by direct preference optimization (DPO). The training signals for both stages come from an off-the-shelf multilingual emphasis detector, so that no manual emphasis annotation or task-specific reward design is needed. We find that DPO can substantially strengthen emphasis control with only a few thousand preference pairs and training steps. Objective and subjective evaluations against strong public systems show that our model produces perceptible emphasis while preserving naturalness and overall quality.
Introduction
Current text-to-speech (TTS) systems can already generate highly realistic and natural speech [1, 2]. However, human speech carries information that is difficult to fully convey through text alone. Emphasis is one such example: emphasizing different words in the same sentence leads to different contrastive focus [3]. For example, consider the sentence “She called today.”
She called today (not someone else). She called today (not messaged). She called today (not yesterday).
Therefore, when TTS systems are used in practice, users may need finer-grained control to accurately convey the intended meaning.
A straightforward approach to fine-grained control is to treat control signals as additional input and learn their realization through supervised training. For emphasis control, there are two groups of methods: (1) directly learning emphasis from emphasis tags or features fed to the model [4, 5, 6, 7, 8]; and (2) controlling emphasis indirectly through more concrete prosodic attributes, such as pitch, energy, and duration [9, 10, 11, 12]. The former leaves the acoustic realization of emphasis entirely to the model, but requires training data with word-level emphasis annotations, which are costly to obtain; the latter restricts emphasis to a few prosodic parameters and tends to be language-dependent and less natural than end-to-end generation.
To further improve the adherence to control signals, recent work has explored feedback-based post-training for fine-grained speech control, ranging from online reinforcement learning such as Group
Relative Policy Optimization (GRPO) [13] to offline methods such as Direct Preference Optimization (DPO) [14, 15]. For example, EMORL-TTS [16] defines emphasis as a pitch/energy contrast reward computed on generated speech and reinforces it with a GRPO-style objective. Fish Audio S2 [17] first uses supervised fine-tuning (SFT) to teach localized vocal instructions such as [emphasis], and then applies GRPO to generated speech, using a rich transcriber to penalize omitted instructions.
Inspired by these approaches, we adopt a two-stage recipe—SFT followed by preference optimization—to add multilingual emphasis control to a pretrained TTS system that has never seen emphasis labels. Our method involves three key design choices. First, for preference optimization, we use DPO, which learns directly from fixed chosen/rejected pairs without requiring a separate reward model. Second, we use the same off-the-shelf multilingual emphasis detector [18] to generate pseudo-labels for both SFT and preference optimization, eliminating the need for manually annotated emphasis data. Third, we build on a latent diffusion TTS backbone trained with flow matching [19], similar to the teacher model of DMO-Speech [20], which provides strong zero-shot voice cloning and efficient inference. Our main contributions are as follows:
A simple recipe. We present a detector-in-the-loop SFT and Diffusion-DPO recipe for emphasis control that needs no manual emphasis annotation or task-specific reward design.
Efficient training. We show that DPO substantially strengthens emphasis control with only a few thousand preference pairs and training steps, and that the gain holds across six languages.
Effective control. Objective and subjective evaluations against strong public systems show that our model produces perceptible emphasis while preserving naturalness and overall quality.
Method
Base model
Our base model is a multilingual zero-shot latent diffusion TTS model, similar to the teacher model of DMO-Speech [20]. Its input consists of three parts: the raw text , an IPA phoneme sequence , and a reference audio . The text is encoded by a frozen pretrained T5 encoder into [21], and the phonemes by a learned embedding table into . Both the target audio and the reference audio are encoded by the same frozen Descript Audio Codec (DAC) [22] into latent frames and . The generation process can be written as a single conditional distribution:
and DAC then decodes the result back to a waveform .
The diffusion model is trained to denoise Gaussian noise into the latent . Specifically, is interpolated with noise to obtain a noisy latent , where is the noise level. The model predicts a velocity , which is trained to match the true velocity [23, 19]:

Figure 1. Overview of EMPDPO. SFT stage: a multilingual emphasis detector labels real speech with emphasis markers. DPO stage: the same detector scores candidates sampled from the SFT model to build chosen/rejected pairs for Diffusion-DPO.
where . At inference, we start from pure noise and numerically integrate the predicted velocity (e.g., with Euler steps) to obtain the clean latent .
Supervised fine-tuning with emphasis
To add emphasis control on top of the base model, we wrap each emphasized word with markers on both the text and phoneme paths. For example,
Text input: She <emp>called</emp> today. Phoneme input: ʃiː <emp>kɔːld</emp> təˈdeɪ
We use an off-the-shelf multilingual emphasis detector [18]. The detector is trained on seven languages (en, de, es, fr, it, pt, zh) and is applied zero-shot to other languages. Given a waveform and its transcript, it assigns an emphasis score to each word. We further pick a threshold (0.35) and mark every word whose score exceeds it, on both the text and phoneme paths. The model is then fine-tuned on a mixture of marked and unmarked data, giving it an initial ability to follow emphasis markers.
Direct preference optimization (DPO)
After SFT, we further strengthen the emphasis-following ability with DPO. DPO training requires a pair of chosen and rejected targets for the same input, where the chosen one is preferred. For emphasis control, the chosen audio should sound more emphasized than the rejected one. Both are encoded by the same DAC encoder into a chosen latent and a rejected latent (winner/loser). We denote the conditioning input by . The original DPO objective raises the likelihood of the chosen latent relative to a reference model and lowers that of the rejected one:
where is the log-likelihood ratio against the reference model , the final model of the SFT stage, which serves as the anchor. controls the allowed deviation, and is the sigmoid function. Unlike online reinforcement learning, DPO requires only offline preference pairs, with no separate reward model.
For a diffusion model, the log-likelihood cannot be evaluated directly, so Diffusion-DPO [15] replaces it with the denoising objective and applies the same relative comparison at each noise level:
is the velocity-prediction error on the corresponding clean latent. Intuitively, this asks the policy to drive the denoising error on chosen below that of the frozen reference model, and to push the error on rejected above it.
In practice, optimizing the preference loss alone is unstable: the model drifts away from the reference and the denoising quality of both chosen and rejected samples degrades. We therefore add a replay term, i.e., the original denoising loss of Eq. (2) on additional data, and optimize the total loss with . The replay data can come from two sources: the real speech used for SFT (corpus replay), or the chosen samples of the preference pairs themselves (chosen replay). Corpus replay anchors the model to the real data distribution, whereas chosen replay requires no additional data pipeline; we compare both in Sec. 4.2.
DPO pair construction
Given input text and reference audio (markers are added to and identically; we write only below), we use the SFT model to generate candidates whose text carries emphasis markers, and baselines without markers. We score all candidates and baselines with the emphasis detector.
Denote the set of marked target words in by ; the improvement of a target word in candidate and the selectivity of candidate are defined as
where is the detector score of target word in candidate , is its average score over the baseline samples, and the selectivity is the fraction of non-target words whose scores fall below the lowest target-word score . The composite score of candidate is:
where is the smallest improvement across target words, is the average improvement, and are the weights of the three terms. We use . Intuitively, the first two terms reward candidates that emphasize every target word (the minimum term penalizes leaving any of them out), while the selectivity term rewards candidates in which non-target words stay unemphasized, so the emphasis is not simply achieved by making the whole utterance louder. When , i.e., with a single emphasized target word, , and reduces to the fraction of non-target words scoring below the target word. We take the candidates with the highest and lowest as chosen and rejected , respectively.
Experimental setup
Training configuration
The model is trained in three stages: pre-training without emphasis markers, SFT on the marked/unmarked mixture, and DPO with preference pairs and replay. During pre-training, the model is trained for 700k steps on 300k hours of internal speech data across 21 languages, including the six evaluation languages. SFT uses about 258k high-quality training samples (711 hours), 45% of which carry <emp> markers.
We then use the SFT model to generate the preference pairs. For each input, we generate candidates and baselines. In addition to the composite ranking score defined in § 2.4, we apply the following filters:
The candidate must satisfy ( for ja/zh), ensuring that it realizes the intended emphasis to some degree.
Both chosen and rejected outputs must satisfy CER , NISQA , SIM , and a clipping fraction to ensure speech quality.
Here, CER is the character error rate between the Qwen3-ASR [24] transcript and the reference text; NISQA is the MOS predicted by a no-reference speech-quality model [25]; and SIM is the cosine similarity between the speaker embeddings of the generated audio and the reference audio, computed by an ECAPA-TDNN speaker-verification model on WavLM-Large features [26, 27]. Only pairs satisfying all of the above hard filters are included in the DPO dataset. The resulting dataset contains 5,171 preference pairs (en 957, de 878, es 1,314, fr 796, ja 662, zh 564). We run DPO for 10k steps on 8 GPUs with a batch of 32 pairs plus 64 replay clips per step, using AdamW with a learning rate of and .
Baselines
We compare our approach against three public systems.
Fish-S2 [17] (Fish Audio S2 Pro) supports zero-shot voice cloning from a reference audio and provides an inline [emphasis] tag, which makes it the closest match to our setting.
WordVoice [12] introduces discrete word-level (character-level for Chinese) attribute tokens for duration, energy, pitch, boundary, and tone. Attributes not specified are predicted by the model itself. It has no explicit notion of emphasis and supports only English and Chinese. To use it, we first perform a generation pass and do force-alignment on the output to obtain the timing of each word. We then generate a second time with the target word’s pitch set to the mean pitch of the input reference audio +4.5 semitones, its energy set to the mean +4 dB, and its duration set to 1.45× (1.35× for Chinese) the duration of that word in the first pass, to simulate emphasis. These values were chosen in a small-scale listening test to balance emphasis strength and naturalness.
Qwen3-TTS [28] (Custom Voice variant) controls emphasis through a natural-language instruction field but does not support voice cloning. We therefore instruct the model to emphasize the target word and use its built-in voices.
Evaluation protocol and metrics
We conduct objective and subjective evaluations of word-level emphasis control in six languages, measuring how strongly each model emphasizes the target word and how natural the resulting speech is.
The evaluation texts cover six languages (de, en, es, fr, ja, zh), with 35 sentences per language and one annotated target word per sentence, for 210 sentences in total. None of them overlap with the training texts. We use 8 reference audio clips, 4 male and 4 female.
This yields distinct inputs per model. Since WordVoice supports only en and zh, it covers inputs. Qwen3-TTS uses 8 of its preset voices (also 4 male and 4 female) and is included only in the objective evaluation, since its voices differ from the reference audio. To reduce the effect of sampling randomness, we generate 3 outputs with emphasis markers and 3 without for every input. For the subjective evaluation, one of the 3 marked outputs is selected at random. The objective metrics are:
is the target-word gain defined in § 2.4 averaged over all 3 outputs and all inputs.
CER, SIM, and NISQA are defined as in § 3.1; SIM is not reported for Qwen3-TTS, which takes no reference audio.
UTMOS is the naturalness score of a neural MOS predictor [29]; together with NISQA it serves as an audio quality metric.
For all metrics other than , we report the mean over all marked inputs and over all unmarked inputs separately, denoted emphasis and plain, respectively.
Subjective listening tests remain the more reliable way to evaluate emphasis generation [9, 10, 11, 12, 16]. We use the same input set as in the objective evaluation and evaluate Ours-Base, Ours-SFT, Ours-DPO, Fish-S2, and WordVoice (en and zh only). Ours-Base is the pre-trained model without any emphasis marker, serving as a no-emphasis control. Native speakers of each of the six languages are recruited through Prolific. Each participant completes 20 trials; the interface and the three questions are shown in Figure 2: emphasis strength of the highlighted target word (Q1) and overall naturalness (Q2) on a 1–5 scale, and whether any other word sounds emphasized (Q3). Each participant’s 20 trials cover all systems, and between-system comparisons use within-subject paired -tests.

Figure 2. Listening test interface (English version): emphasis strength of the highlighted word (Q1) and naturalness (Q2) on a 1–5 scale, and whether another word sounds emphasized (Q3).
Results
Comparison with existing systems
Table 1 reports the objective results; denotes the 95% confidence interval (CI). Main observations are:
| System | CER plain / emph. | SIM plain / emph. | UTMOS plain / emph. | NISQA plain / emph. | |
| Ours-Base | — | 0.008 / — | 0.49 / — | 3.61 / — | 4.34 / — |
| Ours-SFT | 0.069 ± 0.003 | 0.008 / 0.007 | 0.47 / 0.47 | 3.60 / 3.58 | 4.42 / 4.42 |
| Ours-DPO | 0.136 ± 0.005 | 0.008 / 0.008 | 0.49 / 0.49 | 3.60 / 3.55 | 4.36 / 4.34 |
| Fish-S2 | 0.129 ± 0.005 | 0.039 / 0.034 | 0.51 / 0.51 | 4.20 / 4.20 | 4.42 / 4.39 |
| Qwen3-TTS | 0.017 ± 0.003 | 0.034 / 0.036 | N/A | 3.56 / 3.48 | 4.18 / 4.06 |
| WordVoice (en/zh) | 0.074 ± 0.008 | 0.047 / 0.040 | 0.46 / 0.46 | 3.92 / 3.86 | 4.77 / 4.74 |
Table 1. Objective evaluation results (210 sentences × 8 voices = 1,680 inputs per system; 560 for WordVoice).
SFT responds to the marker only weakly (); DPO reaches 0.136, slightly above the strongest public system, Fish-S2 (0.129). The explicit acoustic control of WordVoice works to some extent (0.074), while the natural-language instruction of Qwen3-TTS has almost no effect (0.017).
The higher CER of the public systems is equally high on plain inputs, i.e., intrinsic to these models rather than caused by the emphasis marker.
All systems have similar SIM, UTMOS, and NISQA, suggesting no obvious quality issues.
| System | Q1 | Q2 | Q3 | |
| Ours-Base | 364 | 2.73 ± 0.08 | 3.45 ± 0.08 | 42% |
| Ours-SFT | 364 | 3.15 ± 0.08 | 3.52 ± 0.08 | 34% |
| Ours-DPO | 364 | 3.56 ± 0.07 | 3.47 ± 0.08 | 32% |
| Fish-S2 | 364 | 3.25 ± 0.08 | 2.90 ± 0.11 | 29% |
| WordVoice (en/zh) | 140 | 3.37 ± 0.12 | 3.13 ± 0.15 | 36% |
Table 2. Subjective evaluation results: emphasis strength (Q1), naturalness (Q2), and mis-emphasis rate (Q3). : number of native listeners; mean ± 95% CI over participants.
Table 2 reports the subjective results and Figure 3 breaks them down by language. Main observations are:
Figure 3. Per-language subjective results: Q1 emphasis strength (top) and Q2 naturalness (bottom). WordVoice covers only en/zh. Bars: mean over participants; error bars: 95% CI.
Emphasis strength (Q1): DPO is rated highest, significantly above Base, SFT, and Fish-S2, and on par with WordVoice on en/zh.
Naturalness (Q2): DPO stays at the level of Base and SFT, whereas Fish-S2 and WordVoice are significantly lower. Together with Q1, this shows that DPO strengthens emphasis control with only a few thousand training steps and pairs.
Mis-emphasis rate (Q3): DPO is comparable to SFT and slightly worse than Fish-S2.
Per language, Fish-S2 beats DPO on en but loses on the others. By checking the Chinese (zh) samples (see the audio samples page), we found that Fish-S2 output carries a noticeable foreign accent. We suspect that Fish-S2 carries over the accent of the English reference audio together with the voice, which would explain why it does well only on en.
WordVoice matches DPO in strength but not in naturalness. By checking samples, we found that with fixed pitch/energy/duration settings it merely raises the loudness of the target word, which sounds natural only in some contexts.
Comparison of replay strategies
Table 3 compares the two replay sources of Sec. 2.3—the real SFT corpus (corpus) and the model’s own chosen samples (chosen)—with identical initialization, preference pairs, and hyperparameters; an additional row continues SFT for another 10k steps as a control. Besides the metrics of Table 1, we report the integrated loudness (LUFS); all metrics are computed on the emphasis outputs only.
| Variant | CER | LUFS | UTMOS | NISQA | |
| SFT (init) | 0.069 ± 0.003 | 0.007 | −19.7 | 3.58 | 4.42 |
| SFT +10k | 0.069 ± 0.003 | 0.005 | −19.7 | 3.56 | 4.41 |
| DPO-chosen | 0.136 ± 0.005 | 0.006 | −14.0 | 3.59 | 3.74 |
| DPO-corpus | 0.136 ± 0.005 | 0.008 | −19.6 | 3.55 | 4.34 |
Table 3. Replay strategies, evaluated on the emphasis outputs. SFT +10k: SFT trained for 10k more steps; corpus / chosen: replay on real speech / on self-generated chosen samples; LUFS: integrated loudness.
Training SFT for 10k more steps barely changes any metric, so the DPO gain does not come from the additional updates. We also tried DPO without any replay, but the denoising errors on both chosen and rejected samples explode after about 2k steps, and the final model can no longer generate intelligible speech. We suspect that the DPO loss constrains only the difference between the two denoising errors and does not prevent the policy from drifting away from both samples simultaneously. Replay provides the missing anchor.
The two replay sources reach the same but differ in a side effect: with chosen replay, the output becomes 5.7 dB louder than SFT and NISQA drops from 4.4 to 3.7, whereas corpus replay keeps loudness at the SFT level. Given this result, we use corpus replay in the objective and subjective evaluations.
Conclusion
We presented EmpDPO, a simple, efficient, and effective recipe for word-level emphasis control in a multilingual zero-shot diffusion TTS model. A single off-the-shelf emphasis detector supplies the training signal for both SFT and DPO, so no manual emphasis annotation or task-specific reward design is needed. DPO substantially strengthens emphasis control with only a few thousand preference pairs and training steps, provided that a replay term anchors the model to real speech. Objective and subjective evaluations against strong public systems show that our model produces perceptible emphasis while preserving naturalness and overall quality.
Acknowledgments (AI-use disclosure). LLM-based coding agents (Anthropic Claude, OpenAI GPT, and Moonshot Kimi models) assisted with the training/evaluation code, produced Fig. 1 and Fig. 3 from the authors’ design and data, and rephrased and proofread the text of all sections; the authors verified all content.
References
References
- [1]Yushen Chen, Zhikang Niu, Ziyang Ma, Keqi Deng, Chunhui Wang, Jian Zhao, Kai Yu, and Xie Chen, “F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching,” arXiv preprint arXiv:2410.06885, 2024.arxiv.org/abs/2410.06885
- [2]Tianxin Xie, Yan Rong, Pengfei Zhang, Wenwu Wang, and Li Liu, “Towards controllable speech synthesis in the era of large language models: A systematic survey,” in Proc. EMNLP, 2025.arxiv.org/abs/2412.06602
- [3]Maureen de Seyssel, Antony D’Aviro, Adina Williams, and Emmanuel Dupoux, “EmphAssess: A prosodic benchmark on assessing emphasis transfer in speech-to-speech models,” in Proc. EMNLP, 2024.
- [4]Antti Suni, Sofoklis Kakouros, Martti Vainio, and Juraj Šimko, “Prosodic prominence and boundaries in sequence-to-sequence speech synthesis,” in Proc. Speech Prosody, 2020.DOI
- [5]Liangqi Liu, Jiankun Hu, Zhiyong Wu, Song Yang, Songfan Yang, Jia Jia, and Helen Meng, “Controllable emphatic speech synthesis based on forward attention for expressive speech synthesis,” in Proc. IEEE Spoken Language Technology Workshop (SLT), 2021.
- [6]Slava Shechtman, Raul Fernandez, and David Haws, “Supervised and unsupervised approaches for controlling narrow lexical focus in sequence-to-sequence speech synthesis,” in Proc. IEEE Spoken Language Technology Workshop (SLT), 2021.
- [7]Yiwei Guo, Chenpeng Du, and Kai Yu, “Unsupervised word-level prosody tagging for controllable speech synthesis,” in Proc. ICASSP, 2022.
- [8]Yi Zhong, Chen Zhang, Xule Liu, Chenxi Sun, Weishan Deng, and Haifeng Hu, “EE-TTS: Emphatic expressive TTS with linguistic information,” in Proc. Interspeech, 2023.DOI
- [9]Shreyas Seshadri, Tuomo Raitio, Dan Castellani, and Jiangchuan Li, “Emphasis control for parallel neural TTS,” in Proc. Interspeech, 2022.arxiv.org/abs/2110.03012
- [10]Judith Bauer, Frank Zalkow, Meinard Müller, and Christian Dittmar, “Explicit emphasis control in text-to-speech synthesis,” in Proc. 13th ISCA Speech Synthesis Workshop (SSW), 2025.DOI
- [11]Arnaud Joly, Marco Nicolis, Ekaterina Peterova, Alessandro Lombardi, Ammar Abbas, and Arent van Korlaar, “Controllable emphasis with zero data for text-to-speech,” in Proc. 12th ISCA Speech Synthesis Workshop (SSW), 2023.arxiv.org/abs/2307.07062
- [12]Sihang Nie, Jinxin Ji, Xiaofen Xing, Deyi Tuo, et al., “WordVoice: Explicit and decoupled multi-dimensional word-level control for LLM-based TTS,” arXiv preprint arXiv:2607.06461, 2026.arxiv.org/abs/2607.06461
- [13]Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, et al., “DeepSeekMath: Pushing the limits of mathematical reasoning in open language models,” arXiv preprint arXiv:2402.03300, 2024.arxiv.org/abs/2402.03300
- [14]Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn, “Direct preference optimization: Your language model is secretly a reward model,” in Proc. NeurIPS, 2023.arxiv.org/abs/2305.18290
- [15]Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik, “Diffusion model alignment using direct preference optimization,” in Proc. CVPR, 2024.arxiv.org/abs/2311.12908
- [16]Haoxun Li, Yu Liu, Yuqing Sun, Hanlei Shi, Leyuan Qu, and Taihao Li, “EMORL-TTS: Reinforcement learning for fine-grained emotion control in LLM-based TTS,” in Proc. ICASSP, 2026.arxiv.org/abs/2510.05758
- [17]Shijia Liao, Yuxuan Wang, Songting Liu, Yifan Cheng, et al., “Fish audio S2 technical report,” arXiv preprint arXiv:2603.08823, 2026.arxiv.org/abs/2603.08823
- [18]Megan Wei, Deepali Aneja, Jiaqi Su, Yun Yun Wang, Haonan Chen, and Zeyu Jin, “Do speech emphasis models generalize across languages and emotions?,” arXiv preprint arXiv:2606.27717, 2026.arxiv.org/abs/2606.27717
- [19]Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le, “Flow matching for generative modeling,” in Proc. ICLR, 2023.arxiv.org/abs/2210.02747
- [20]Yinghao Aaron Li, Ritesh Kumar, and Zeyu Jin, “DMO-Speech: Direct metric optimization via distilled diffusion model in zero-shot speech synthesis,” arXiv preprint arXiv:2410.11097, 2024.arxiv.org/abs/2410.11097
- [21]Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,” Journal of Machine Learning Research, vol. 21, no. 140, pp. 1–67, 2020.arxiv.org/abs/1910.10683
- [22]Ritesh Kumar, Prem Seetharaman, Alejandro Luebs, Ishaan Kumar, and Kundan Kumar, “High-fidelity audio compression with improved RVQGAN,” in Proc. NeurIPS, 2023.arxiv.org/abs/2306.06546
- [23]Tim Salimans and Jonathan Ho, “Progressive distillation for fast sampling of diffusion models,” in Proc. ICLR, 2022.arxiv.org/abs/2202.00512
- [24]Xian Shi, Xiong Wang, Zhifang Guo, Yongqi Wang, Pei Zhang, Xinyu Zhang, Zishan Guo, Hongkun Hao, Yu Xi, et al., “Qwen3-ASR technical report,” arXiv preprint arXiv:2601.21337, 2026.arxiv.org/abs/2601.21337
- [25]Gabriel Mittag, Babak Naderi, Assmaa Chehadi, and Sebastian Möller, “NISQA: A deep CNN-self-attention model for multidimensional speech quality prediction with crowdsourced datasets,” in Proc. Interspeech, 2021.DOI
- [26]Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yosh ioka, Xiong Xiao, et al., “WavLM: Large-scale self-supervised pre-training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022.DOI
- [27]Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck, “ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification,” in Proc. Interspeech, 2020.DOI
- [28]Hangrui Hu, Xinfa Zhu, Ting He, Dake Guo, Bin Zhang, Xiong Wang, Zhifang Guo, Ziyue Jiang, Hongkun Hao, et al., “Qwen3-TTS technical report,” arXiv preprint arXiv:2601.15621, 2026.arxiv.org/abs/2601.15621
- [29]Takaaki Saeki, Detai Xin, Wataru Nakata, Tomoki Koriyama, Shinnosuke Takamichi, and Hiroshi Saruwatari, “UTMOS: UTokyo-SaruLab system for VoiceMOS challenge 2022,” in Proc. Interspeech, 2022.arxiv.org/abs/2204.02152
