TaDiCodec: Text-aware Diffusion Speech Tokenizer for Speech Language Modeling

Dear reviewer, we want to thank for your carefully reading and providing the constructive comments! Below we address the concerns mentioned in the review.

Weakness 1: However, some claims from the paper seem wrong. Firstly, they refer the auto-regressive zero-shot TTS as "speech language models", which should include a broader range of models, such as Moshi-style text-speech LMs. So "improving the reconstruction-generation gap for SLMs" seems to be overclaiming.

Thank you for your suggestion. We acknowledge that the term speech language modeling can have broader interpretations depending on context. In our work, we use speech language modeling to specifically refer to generation speech tokens with language models, particularly in the context of text-conditional speech generation. This definition has also been adopted in prior works on speech tokenizers [1, 2, 3, 4], which similarly evaluate their models primarily on language-model-based TTS tasks.

We fully understand your concern about potential overclaiming, and we commit to clarifying our definition of speech language modeling in the revised version of the paper to avoid overclaiming.

Additionally, we are actively exploring the use of TaDiCodec (as the output) for building Moshi- or GLM4-Voice-style dialog speech LLMs, where both text and speech tokens are generated jointly. These models can decode speech outputs via the TaDiCodec decoder, which also motivates our tokenizer design.

Weakness 2: while they show a promising result on the zero-shot TTS task, there is no evidence on whether it benefits from the text-aware and the prompt, as other approaches don't have that. It will be beneficial to have ablation study on the text-aware component.

We compare all baseline TTS models in a zero-shot setting, where a prompt is provided for each sample to guide the generation. In fact, most of these baselines require prompts for their language models and diffusion decoders (if such components are present). In our case, we additionally provide text to the decoder, as this is a core part of our design.

For the ablation study on the text-aware component, we provide additional experiments demonstrating its impact in our response to Question 1.

Question 1: Can the authors provide an ablation study on the text-aware component?

We have conducted additional experiments to study the impact of the text-aware decoder. Since the core design of our tokenizer is to achieve high-quality reconstruction under extremely low token rates with the diffusion autoencoder and the guidance of text information, we started with the most basic setting: a GAN-based speech tokenizer. As shown below, the GAN-based model struggles to reconstruct intelligible speech at such low token rates and suffers from unstable training behavior compared to diffusion-based models.

We then replaced the tokenizer with an end-to-end trained diffusion autoencoder without any textual conditioning. We observe that both speech quality and similarity improve significantly, and the WER also decreases noticeably, though it still remains relatively high at this stage.

Finally, we introduced the text-aware decoder, which further improved performance across all metrics. Introducing the text-aware decoder significantly reduces the representational burden on the tokens, allowing them to focus on learning residual information beyond the text, while still potentially capturing some textual content, under very low token-rate settings. This design finally makes accurate reconstruction.

Question 2: From the reconstruction demo on the demo page, it seems the codec has some subtle changes to the prosody. As the codec is trained as a generative model, it makes sense to involve some randomness. Is it possible to introduce a metric to quantify that?

We provide additional evaluation metrics to measure prosodic deviation between the generated speech and the ground truth. Specifically, we compute the F0 Correlation (FPC) over all samples in the test set to assess overall prosody preservation.

To quantify the randomness introduced by the generative model, particularly due to the initial gaussian noise in flow matching models, we generate each sample five times and compute the standard deviation of FPC across these generations. We use the average standard deviation across the dataset as an indicator of prosodic variability introduced by sampling stochasticity.

We can observe that DualCodec achieves the highest mean FPC and zero mean FPC std, which can be attributed to its discriminative modeling approach and highest bitrate, allowing it to preserve prosody more precisely. In contrast, CosyVoice Tokenizer, Vevo Tokenizer, and TaDiCodec yield similar lower mean FPC scores,while TaDiCodec exhibits a higher mean FPC Std, likely due to its lower bitrate.

By the way, from a perceptual perspective, we observe that minor variations in prosody tend to occur more across different samples, rather than across multiple generations of the same sample. That is, the differences introduced by stochastic sampling (of the initial gaussian noise) are generally subtle.

Thanks again for your valuable comments, we would be grateful if we could hear your feedback regarding our answers. We would be happy to answer and discuss if you have further comments.

Reference:

[1] Zhang X, Zhang D, Li S, et al. Speechtokenizer: Unified speech tokenizer for speech large language models[J]. arXiv preprint arXiv:2308.16692, 2023.

[2] Ji S, Jiang Z, Wang W, et al. Wavtokenizer: an efficient acoustic discrete codec tokenizer for audio language modeling[J]. arXiv preprint arXiv:2408.16532, 2024.

[3] Ye Z, Sun P, Lei J, et al. Codec does matter: Exploring the semantic shortcoming of codec for audio language model[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2025, 39(24): 25697-25705.

[4] Li J, Lin X, Li Z, et al. DualCodec: A Low-Frame-Rate, Semantically-Enhanced Neural Audio Codec for Speech Generation[J]. arXiv preprint arXiv:2505.13000, 2025.

Dear reviewer, we want to thank for your carefully reading and providing the constructive comments! Below we address the concerns mentioned in the review.

Weakness 1: Missing some recent works: WavTokenizer, StableCodec, Single-Codec, BiCodec, and DM-Codec.

In Table 1 of our paper, which presents the main results of our tokenizers, we compare against several representative baselines, including WavTokenizer, BiCodec, StableCodec (TAAE), as well as some other recent models such as DualCodec, x-codec 2, Mimi, BigCodec, Vevo Tokenizer, FireRedTTS Tokenizer, CosyVoice 1/2 Tokenizer, and Ints Tokenizer.

For Single-Codec, we did not include results as the model weight and inference code are not publicly available. We did not reproduce the model as our benchmark already includes a sufficiently diverse set of strong baselines.

Regarding DM-Codec, we were not previously aware of this work and will add the appropriate citation in our revision. While it adopts 8 layers of RVQ (which means it has a much higner token rate), its performance is comparable to that of SpeechTokenizer.

Weakness 2: The paper lacks evaluation on established comprehensive benchmarks: Codec-SUPERB, DASB, or ARCH benchmarks...

Our work primarily focuses on designing speech tokenizers tailored for speech generation with language models or as the output of speech language models. Under this setting, we explore how to compress speech representations as much as possible. So, our evaluation mainly targets speech reconstruction and speech generation with language modeling.

To address the reviewer's suggestion, we additionally evaluate our codec on the DASB benchmark [1]. Due to time constraints, we select two representative tasks: emotion recognition and speaker verification. Following the official DASB protocol, we train models using our codec embeddings (after projection) and compare the performance against SpeechTokenizer as reported in the original DASB paper:

Furthermore, we could not identify a benchmark named ARCH that is publicly available or specifically related to speech codec evaluation. Could you please provide more details about it? We would be happy to look into it!

Weakness 3: Codebook utilization rate is not reported.

Thank you for the suggestion. We adopt BSQ for quantization, which significantly alleviates the codebook collapse issue. In our experiments, with a codebook size of 16,384, our method achieves a utilization rate of over 99%.

Weakness 4: No analysis of model size or memory requirements compared to baselines.

We provide detailed model configurations in Appendix A.1, along with an ablation study on tokenizer model size in Table 4, and analysis of model size and real-time factor (RTF) for TTS in Table 6.

Since the core design of our work leverages a text-aware module and a diffusion autoencoder to enable high-quality reconstruction at an extremely low token rate (6.25 Hz), while maintaining efficient generation within the full speech generation with language modeling pipeline, we believe that the results and analysis presented in Table 6 sufficiently support our claims regarding efficiency and model compactness.

Weakness 5: TaDiCodec uses 100k hours while some baselines use different amounts. Is it a fair comparison?

Indeed, we are not able to ensure that all compared tokenizers are trained on exactly the same dataset. Our model, TaDiCodec, is trained on the Emilia dataset consisting of 100k hours of speech. A significant portion of the baselines are also trained on Emilia or its supersets (which implies they use more training data including Emilia), or they are trained on datasets larger than 100k hours collected with similar data pipelines. These include: Mimi , DualCodec (Emilia), X-codec 2 (Emilia + MLS), Vevo Tokenizer (Emilia), FireRedTTS Tokenizer (200k+ hours), CosyVoice 1/2 Tokenizer (170k+ hours), and Ints Tokenizer (Emilia).

Weakness 6: Only 5,000 samples for test-en and test-zh, which may be insufficient for robust statistical significance analysis.

We follow standard benchmarks commonly used in prior work on speech generation, and provide a detailed discussion of our evaluation datasets in Appendix F.

Specifically, for tokenizer evaluation, we use SeedTTS test-en (1,000 samples) and test-zh (2,000 samples), along with multilingual test sets in French, German, Japanese, and Korean, each consisting of 300 utterances randomly sampled from Common Voice.

For TTS evaluation, in addition to SeedTTS test-en and test-zh, we include 400 samples from Articulatory-en, 400 from Articulatory-zh, 500 from Code-switching-en, 500 from Code-switching-zh, 500 from Cross-lingual-zh2en, and 500 from Cross-lingual-en2zh. While many previous works only evaluate on SeedTTS test-en/test-zh or LibriSpeech. In fact, the total number of evaluation samples used in our work exceeds that of most related papers.

Weakness 7: Insufficient Ablation Studies.

Thank you for the valuable suggestion! We are more than happy to provide additional analysis as follows:

The effect of text quality and performance degradation when text-speech misalignment:

We use ground-truth text during inference. To assess the robustness of our system to imperfect text input, we additionally provide results using transcripts generated by Whisper-large-v3. On the SeedTTS test-en set, Whisper-large-v3 achieves a low WER of 2.15. We observe a slight degradation in WER when using ASR transcripts (from 2.73 to 3.24), while SIM and UTMOS remain nearly unchanged. This aligns with our motivation: modern ASR systems are already powerful enough for high-quality text transcription. Moreover, in the speech generation setting, the target text is typically predefined or user-specified, and thus always accessible. Therefore, using ground-truth text for evaluation does not affect the fairness or validity of our TTS results.

Ablation on different text encoder architectures:

In fact, our work does not involve any dedicated text encoder for either the tokenizer or the TTS models. We follow a minimal design principle: for both the tokenizer and the TTS system, the text is simply treated as a prefix of the speech token sequence and fed jointly into a standard Transformer architecture using self-attention. No additional modules are used to separately process the text input (e.g., via a dedicated text encoder). While incorporating such components may offer potential improvements, this direction is orthogonal to our core design goals.

Weakness 8: Limited Analysis of Failure Cases.

In our demo page, we showcase several challenging cases such as singing and utterances with significant pitch variation, demonstrating that our method performs better than other baselines (with low token rates) in these scenarios. As for very long speech samples, we observe that performance degrades when the model processes utterances longer than 40 seconds in a single pass. The reason is the training data primarily consists of utterances shorter than 30 seconds. However, this issue can be mitigated by chunking long utterances into segments (e.g., 30 seconds each), which is straightforward to implement.

Weakness 9: No evaluation on speech understanding tasks. Why evaluated only evaluated on TTS tasks?

We provide a detailed explanation in our response to Weakness 2.

Question 1: How does the model perform when text input is noisy or contains errors (e.g., from imperfect ASR)? Have you evaluated robustness to text quality degradation?

We provide a detailed explanation in our response to Weakness 7.

Question 2: Can you provide more insight into why text conditioning enables such extreme compression? Is there a theoretical formulation that explains the information-theoretic relationship between text and the required speech token rate?

We're happy to provide more high-level insights. This is closely related to the question raised in question 1 of reviewer rwbf.

TaDiCodec is inspired by recent NAR TTS systems such as E2TTS, which have demonstrated that it is possible to generate speech features directly from text using diffusion models. However, these models operate on high-dimensional continuous acoustic representations, which are often difficult to model accurately and comprehensively.

In contrast, TaDiCodec introduces discrete tokens that encode essential information beyond semantics, including prosody, speaker characteristics, and text-speech alignment, which are critical for accurate reconstruction. Additionally, the powerful generative capacity of the diffusion decoder allows it to synthesize fine-grained acoustic details that do not need to be explicitly encoded in the tokens. This design enables TaDiCodec to achieve high-quality reconstruction and more accurately capture the target speech distribution.

Due to space limitations, we provide a more detailed discussion in our response to question 1 of reviewer rwbf. Please refer to it for further insights.

Question 3: Have you explored distillation methods (e.g., consistency models) to reduce the number of diffusion steps to 1-2 while maintaining quality? This could significantly improve inference efficiency.

It's a very promising direction, and we are actively exploring this problem by finetuning the decoder through distillation. We plan to share results in future work.

Thanks again for your valuable comments, we would be grateful if we could hear your feedback regarding our answers. We would be happy to answer and discuss if you have further comments.

Reference

[1] Mousavi P, Della Libera L, Duret J, et al. Dasb-discrete audio and speech benchmark[J]. arXiv preprint arXiv:2406.14294, 2024.

Thank you very much for your response and your consideration of raising the score. We understand that you still have some remaining concerns, and we sincerely appreciate your thoughtful feedback. We will strive to address these issues in our future work. Meanwhile, we would also like to further clarify some of the questions.

For model comparison, DualCodec, X-codec 2, Vevo Tokenizer, and Ints Tokenizer were all trained on Emilia. We will explicitly indicate the training datasets used for each baseline in the table of our paper to ensure a fairer and more transparent comparison.

For "Failure case analysis", we provide some quantitative analysis on the slight prosodic variations introduced by randomness in the response to Reviewer WHX2's Question 2.

For Ablations, we provide the ablation of the effect of text quality and performance degradation when text-speech misalignment in the response to Weakness 7: Insufficient Ablation Studies, we also provide ablation of the text-aware component in the response to Reviewer WHX2's Question 1. We provide it here:

Thank you again for your valuable feedback!

Dear reviewer, we want to thank for your carefully reading and providing the constructive comments! Below we address the concerns mentioned in the review.

Weakness 1: Limited Application Scenarios... In contrast, the speech tokenizers used in models like Moshi and GLM-4-Voice are capable of handling both speech understanding and generation tasks.

In this work, we primarily focus on exploring the use of diffusion-based speech tokenizers for TTS and as output representations for speech language models. The motivation stems from the observation that, in TTS scenarios and most speech language models, the corresponding text of the target speech is typically available. For example, in TTS, the target text is always known, and in most end-to-end spoken dialogue systems, text and speech tokens are generated jointly. Under this setting, we aim to push the limits of compression in speech tokenization.

We are actively investigating how diffusion speech tokenizers can be used for both speech generation and understanding. One promising direction is to inject textual information into the speech tokens during training, either through explicit ASR supervision or by enabling the encoder to generate an additional set of tokens aligned with the text—thereby eventually removing the decoder's dependence on textual input. This remains an exciting area for future work. In this paper, however, we focus on the text-aware decoder setting, where the decoder is conditioned on textual information during generation.

Weakness 2: The comparison in Table 1 is not fair... I believe that at least one more column should be added to indicate whether each codec requires extra information for reconstruction, such as: none, prompt speech, global features, or text condition, etc.

Thank you for your suggestion. We will revise the table in the paper accordingly and add an additional column to clarify the extra inputs required by each codec. We will also compute an adjusted token rate upper bound that accounts for these additional conditions, to provide a clearer and more accurate comparison.

Question 1: Regarding the diffusion decoder... Are these tokens encoding the residual information, maybe speech-text-global? Providing more analysis on this point would add valuable insights to the paper.

Thank you for the insightful question. Prior works such as [1, 2] predict mel-spectrograms or continuous latent features directly from text and prompts, but still require an additional duration predictor to explicitly align text and speech. Meanwhile, studies like [3, 4] have shown that it is possible to predict speech features without explicit alignment modeling. However, all of these approaches directly model complex continuous acoustic representations, which are often difficult to capture comprehensively the target distribution.

Qualitatively, these models can be viewed as a special case of TaDiCodec with a token rate of zero. That is, they first predict text, encode the audio into "discrete tokens with sequence length of zero", then decode from text and "zero length tokens". It is evident that such approaches struggle to accurately reconstruct the original speech, especially in more expressive scenarios. Our demo page demonstrates that our model is capable of reconstructing challenging speech types such as singing, whispering, and speech with significant prosodic variation, with minimal timbre or pitch distortion.

Returning to the TTS and language modeling setting, we argue that predicting TaDiCodec tokens and decoding them is more tractable than directly predicting high-dimensional continuous features. This is because the distribution over compressed discrete tokens is easier to model, and the tokens can be decoded into high-quality audio. Our TTS results, which outperform strong baselines such as F5-TTS, further support this claim.

We also provide analyses to better understand what TaDiCodec learns. Since our codec achieves high-fidelity reconstruction, the learned tokens encode residual factors such as prosody, speaker information, and text-speech alignment. This is evident from our demo page, which showcases reconstructions of singing and other expressive speech styles. Furthermore, because our system is trained end-to-end without explicit disentanglement constraints, the tokens naturally capture some semantic information as well. This behavior reflects the tradeoff governed by the information bottleneck, which is influenced by factors such as codebook size and token rate. Our preliminary experiments show that with a codebook size of 16,384, the TTS model achieves excellent performance, suggesting that the alignment between text and speech tokens can be easily learned under this capacity. However, when the codebook size is reduced to 4,096, we observe that while the reconstruction quality remains high, TTS performance drops significantly. We attribute this to a loss of semantic content in the tokens, which makes it harder for the language model to learn the alignment between text and speech tokens.

We also provide additional evaluations of our tokenizers on the DASB benchmark to show that TaDiCodec learns acoustic information. Due to time constraints, we select two representative tasks: emotion recognition and speaker verification. Following the official DASB protocol, we train models using our codec embeddings (after projection) and compare the performance against SpeechTokenizer as reported in the original DASB paper:

Question 2: In Figure 3a, the WER of reconstruction is generally better than that of generation... I believe that discussing this unconventional phenomenon would be very interesting and insightful.

Interesting question! This phenomenon has also been observed in some advanced TTS systems. We believe the underlying reason lies in certain biases introduced by the way WER is computed. For most test sets, paired speech and text are provided, and in theory, the ground-truth speech should yield a WER of zero. However, this is often not the case in practice, as ASR models learn to predict from an averaged distribution, and may not perfectly align with the real ground-truth audio.

In contrast, the generated speech tokens are guided directly by the text, which in some cases may be more easily captured by the ASR model, resulting in lower WER.

A similar phenomenon is also observed between the reconstructed audio from TaDiCodec and the ground-truth speech. For example, on the Chinese test set SeedTTS-zh, we find that TaDiCodec-reconstructed speech yields a better WER than the original ground-truth speech (0.94 vs. 1.254). One possible explanation is that TaDiCodec uses a text-aware decoder, which may lead to better alignment between the predicted speech and the corresponding text. From the perspective of an averaged ASR distribution, this alignment could make the generated speech more recognizable to the ASR model.

Thanks again for your valuable comments, we would be grateful if we could hear your feedback regarding our answers. We would be happy to answer and discuss if you have further comments.

Reference

[1] Shen K, Ju Z, Tan X, et al. NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers[C]//ICLR. 2024.

[2] Le, Matthew, et al. "Voicebox: Text-guided multilingual universal speech generation at scale." Advances in neural information processing systems 36 (2023): 14005-14034.

[3] Eskimez S E, Wang X, Thakker M, et al. E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts[C]//2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024: 682-689.

[4] Chen Y, Niu Z, Ma Z, et al. F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching[J]. arXiv preprint arXiv:2410.06885, 2024.

Dear reviewer, we want to thank for your carefully reading and providing the constructive comments! Below we address the concerns mentioned in the review.

Weakness 1: The paper lacks clarity on some key implementation details. For instance, the specific downsampling and upsampling methods used before and after quantization are not described. The paper states that it uses mel-spectrograms as the reconstruction target, but it does not specify which vocoder model was used to generate the final audio waveforms. This is a critical detail needed when evaluating it compared to other methods.

Thank you for your suggestion! We are happy to provide more implementation details. Since our tokenizer operates at a frame rate of 6.25 Hz, while both the input and output mel-spectrograms are at 50 Hz, we perform downsampling before quantization and upsampling after quantization to match the temporal resolutions.

To keep it simple, we use basic interpolation for both operations. Let the embedding before quantization be denoted as z, and the embedding after quantization as z_q. The following pseudocode illustrates this process using torch.nn.functional.interpolate: import torch.nn.functional as F

# Downsample from 50 Hz to 6.25 Hz before VQ
z_downsampled = F.interpolate(z, scale_factor=1/8, mode="linear", align_corners=True)
# Quantize
z_q = vq(z_downsampled)
# Upsample back to 50 Hz after VQ
z_q_upsampled = F.interpolate(z_q, scale_factor=8, mode="linear", align_corners=True)

We find this simple approach to be effective and sufficient for aligning the representations temporally across stages.

For the vocoder, we use the Vocos [1] architecture with increased model capacity (around 120M parameters) and adopt the loss functions and GAN design from DACodec [2]. Unlike traditional vocoders that require upsampling, the Vocos decoder directly predicts the STFT spectrogram, which leads to higher inference efficiency compared to BigVGAN. Additionally, we find that it achieves strong reconstruction quality for both speech and singing.

We will include this information in the appendix of the paper, and we commit to releasing our models and training code to further support transparency and reproducibility.

Weakness 2: While the current architecture, which relies on bidirectional attention and an iterative diffusion decoder, is primarily suited for audio compression and offline zero-shot TTS, this limits its applicability for streaming use cases where audio chunks must be generated sequentially before the full text token stream is available.

Thank you for the insightful comment! This is indeed one of the directions we plan to explore in future work. From an engineering perspective, even without modifying the model architecture, we can perform chunk-wise inference by generating a fixed number of text tokens, predicting the corresponding speech tokens, and decoding them with TaDiCodec, without waiting for all text tokens to be generated.

A more promising direction is to design the decoder itself to be chunk-wise during training, supporting autoregressive diffusion over both text and speech tokens. Concretely, the current decoder input follows the pattern:

[text tokens, speech tokens]

We propose to restructure it as:

[text_chunk_1, speech_chunk_1, ..., text_chunk_n, speech_chunk_n]

This interleaved organization pattern of text and speech tokens has also been adopted in many speech LLMs, such as GLM4-Voice [3] and the streaming version of CosyVoice 2 [4].

where the length of speech tokens in each chunk is set to a fixed ratio relative to the text tokens. Within each chunk, diffusion is used for generation, while causal attention is applied across chunks. This approach has been validated in prior works such as [5, 6].

Our preliminary experiments also show that introducing a slight shift (or look-ahead) of approximately 1 second between speech tokens and the target mel-spectrogram improves chunk-wise streaming inference. Under a 1-second chunk size, the model can achieve real-time performance with minimal degradation in quality.

Question 1: Since the vocoder is not specified, it could be assumed that the RTF calculation excludes vocoder inference time. Please clarify this explicitly.

For the vocoder, we use the Vocos [1] architecture with increased model capacity (around 120M parameters) and adopt the loss functions and GAN design from DACodec [2]. Unlike traditional vocoders that require upsampling, the Vocos decoder directly predicts the STFT spectrogram, which leads to higher inference efficiency compared to BigVGAN. Additionally, we find that it achieves strong reconstruction quality for both speech and singing. It is worth noting that in the RTF calculation for TTS, we observe that the vocoder contributes minimally to the overall inference time across all baselines. The majority of the computational cost lies in the language model and the flow-matching decoder.

Question 2: While the contribution of this work and its significance is not diminished without this discussion, to broaden the work's impact on the wider community, it would be beneficial for the authors to discuss future directions for adapting TaDiCodec for streaming scenarios and its integration with real-time language models.

Thanks for your suggestion! We discuss this question in Weakness 2.

Thanks again for your valuable comments, we would be grateful if we could hear your feedback regarding our answers. We would be happy to answer and discuss if you have further comments.

Reference:

[1] Siuzdak H. Vocos: Closing the gap between time-domain and fourier-based neural vocoders for high-quality audio synthesis[J]. arXiv preprint arXiv:2306.00814, 2023.

[2] Kumar R, Seetharaman P, Luebs A, et al. High-fidelity audio compression with improved rvqgan[J]. Advances in Neural Information Processing Systems, 2023, 36: 27980-27993

[3] Zeng A, Du Z, Liu M, et al. Glm-4-voice: Towards intelligent and human-like end-to-end spoken chatbot[J]. arXiv preprint arXiv:2412.02612, 2024.

[4] Du Z, Wang Y, Chen Q, et al. Cosyvoice 2: Scalable streaming speech synthesis with large language models[J]. arXiv preprint arXiv:2412.10117, 2024.

[5] Liu Z, Wang S, Inoue S, et al. Autoregressive diffusion transformer for text-to-speech synthesis[J]. arXiv preprint arXiv:2406.05551, 2024.

[6] Ju Z, Yang D, Yu J, et al. MoonCast: High-quality zero-shot podcast generation[J]. arXiv preprint arXiv:2503.14345, 2025.