Pith. sign in

REVIEW 4 major objections 6 minor 123 references

A single 48 kHz model can generate expressive multi-speaker speech and audio from either a natural-language caption or a reference voice clip, and it benchmarks ahead on several zero-shot and instruct tasks.

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 →

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

T0 review

2026-08-04 16:23 UTC pith:AHSW7G3E

load-bearing objection A well-engineered unified TTS/audio system whose headline expressiveness claims rest on uncalibrated LLM judges and in-house benchmarks; the engineering is worthy of refereeing, but the evidence is underdetermined. the 4 major comments →

arxiv 2608.02023 v1 pith:AHSW7G3E submitted 2026-08-03 eess.AS cs.SD

SwanTale: Unified Multi-Speaker Speech and Audio Generation for Instruct and Zero-Shot Tasks

Yu Zhang , Ruiqi Li , Changhao Pan , Ke Lei , Xiang Yin , Cheng Yang This is my paper
classification eess.AS cs.SD
keywords speech generationaudio generationzero-shot TTSinstruct TTSmulti-speaker speechflow matchingmixture-of-expertscaption annotation
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.

This paper claims that a single generative model can handle two previously separate jobs: producing expressive speech and audio from a natural-language caption (instruct) and cloning or extending a voice from a reference clip (zero-shot). To make this work, the authors build a data pipeline that converts roughly 70 million mixed-media clips into structured captions with environment, speaker-style, and content fields, and a 48 kHz autoencoder-latent architecture that outputs one waveform containing speech, background sound, and effects. They report that the resulting system beats existing systems on several zero-shot and instruct benchmarks, especially on expressiveness, and that it can follow complex instructions mixing dialogue and sound effects. A sympathetic reader would take the central claim as: unified multi-speaker speech-and-audio generation is feasible with one backbone if the data captions are rich enough and the model conditions on task, quality, and recurring caption patterns.

Core claim

On its own terms, the discovery is that a flow-matching diffusion transformer, operating on a 25 Hz continuous latent space and trained with a mixture of caption-only and reference-audio conditions, can generate expressive multi-speaker speech together with environmental and effect audio in a single pass at 48 kHz. The authors separate the two control paths—lexical text and speaker turns versus global caption and acoustic attributes—and add a quality flag that lets inference request the highest-quality output without reinforcement learning. They also show that routing computation per audio frame, rather than per task alone, improves instruction accuracy, acoustic quality, and expressiveness

What carries the argument

The load-bearing pieces are: (1) SwanVAE, a 48 kHz to 25 Hz continuous latent autoencoder with a decoder-side Transformer resampling block that keeps local acoustic detail; (2) a flow-matching DiT backbone with reward-conditioned quality control, which feeds quality scores as conditioning and fixes them to 'high' at inference; (3) Engram conditioning, a hash-based memory layer that recognizes recurring caption patterns; (4) Unified MoE, which combines a task router for sample-level priors and a time-aware dynamic Top-P audio router for frame-level expert selection; and (5) curriculum training plus GRPO post-training that progressively moves from zero-shot speech to full caption mixture and t

Load-bearing premise

The central comparisons—especially the 'best expressiveness' claims—depend on automatic LLM judges (Section 4.2) with no reported human correlation or agreement, and the caption training data is auto-annotated; if those judges are biased or the captions are noisy, the stated advantages may not reflect true perceptual quality or instruction fidelity.

What would settle it

Run a human best-worst or MOS listening test on the same SwanBench-Speech monologue and dialogue items used in the zero-shot table, comparing SwanTale against the strongest baseline on Expressive Richness and Hierarchy. If human raters do not prefer SwanTale, the headline expressiveness claim is not supported. Additionally, generate identical captions with 'low' vs 'high' quality flags and measure PESQ/STOI on the outputs; if the waveforms are statistically indistinguishable, reward-conditioned quality control is not doing the claimed work.

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

If this is right

  • A single model can act as both a voice cloner and a voice designer, so a character created from a caption can later be reused via a reference clip without retraining.
  • Complex instruct generation—multi-speaker dialogue plus environment beds and local effects—can be produced in one waveform rather than assembled by separate speech and audio pipelines.
  • Quality labels can be used as conditions rather than filters, letting moderate-quality training data contribute while pushing inference toward the highest quality.
  • Curriculum ordering—zero-shot first, then captions, then full mixtures—helps the model acquire caption grounding without losing reference-audio capability.
  • GRPO post-training with a KL reference term can improve pronunciation and stability without erasing multi-speaker and audio skills learned earlier.

Where Pith is reading between the lines

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

  • Beyond the paper's benchmarks, the two-stage decomposed classifier-free guidance suggests a general recipe: tie lexical content and speaker turns to one guidance scale and all other acoustic attributes to another, so users can trade off 'what is said' versus 'how it sounds'.
  • If the LLM-judge expressiveness results survive human listening checks, the largest gap may come from data diversity and expressive-data selection rather than from the MoE or attention details; that hypothesis is testable by ablating the data pipeline alone.
  • The auto-captioning pipeline could be adapted as a data engine for other audio-generation tasks, such as sound-effect or music captioning, since its environment/content/speaker schema is domain-neutral.
  • The time-aware expert budget in Unified MoE implies adaptive compute could be exploited at inference to skip experts in stable regions, potentially lowering cost for long-form audio, though the paper does not report speed or FLOP numbers.

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. The paper introduces SwanTale, a unified 48 kHz model for multi-speaker speech and audio generation supporting both zero-shot (reference-audio) and instruct (caption-only) modes within a single flow-matching DiT backbone. The authors contribute SwanData-Caption, a large-scale captioning pipeline with coverage design, preprocessing, annotation, and quality filtering; SwanVAE, a 25 Hz continuous-latent autoencoder; and several modeling components, including reward-conditioned quality control, Engram memory conditioning, and a Unified MoE with task- and audio-level routing. Training proceeds through a curriculum of zero-shot pretraining, dense caption adaptation, full-mixture MoE training, high-quality SFT, and GRPO post-training. Evaluations cover SwanVAE reconstruction on public test sets, zero-shot speech on SwanBench-Speech, instruction following on InstructTTSEval, perceptual quality on the new SwanBench-Scene, and an ablation on the new SwanBench-Caption. The abstract claims state-of-the-art zero-shot and instruct metrics, best expressiveness in both tasks, and support for complex multi-speaker instruct generation with audio.

Significance. If the reported results hold, SwanTale would be a notable step toward unified, controllable speech-and-audio synthesis in a single model: it combines zero-shot and instruction-following conditioning, multi-speaker dialogue, environmental audio, and local effects under one flow-matching backbone. The SwanVAE design, with its discriminators, alignment objectives, and 48 kHz continuous latents, is evaluated on public datasets (VCTK, GTSinger, FSD50K, MUSDB18-HQ) and appears competitive, which is a genuine strength. The human professional-listener MOS on SwanBench-Scene is also a positive feature, and the paper is candid about remaining limitations such as long-form generation and background-music transitions. However, the paper's most distinctive claims—best expressiveness and complex instruct generation—rest heavily on LLM-judge scores that are not calibrated against human judgments, on author-constructed benchmarks, and on ablations without external baselines. These validity gaps currently prevent the reader from accepting the headline conclusions at face value.

major comments (4)
  1. [§4.2, Tables 5 and 8] The central expressiveness claims rely on uncalibrated proprietary LLM judges. Section 4.2 states that Expressive Richness and Expressive Hierarchy on SwanBench-Speech are scored by Gemini 3 Pro and that SwanBench-Caption is scored by gemini-3.5-flash; no human correlation, inter-annotator agreement, or calibration against perceptual judgments is reported. The decisive margins are small: e.g., Table 5 monologue Richness 3.90 vs. 3.81 and Hierarchy 3.70 vs. 3.62; Table 8 Instruction Accuracy 3.39 vs. 3.02 in the same-model ablation. Without evidence that these judge scores track human perception, a systematic bias or unstable scale could erase or invert the claimed advantage. This is a load-bearing validity gap for the 'best expressiveness' claim. I would require either a human-correlation study on a held-out subset or a transparent calibration of the judge against human ratings, with ite
  2. [§4.6, Table 8] SwanBench-Caption, the only evaluation of complex multi-speaker audio-containing instruct generation, contains no external baselines. Table 8 compares only SwanTale without MoE, SwanTale, and SwanTale with a 32B caption encoder. Consequently, the abstract's claim that SwanTale 'supports complex instruct generation involving multi-speaker speech and audio' is not demonstrated relative to any existing system. The authors should add at least one or two strong external baselines (e.g., MiMo-Audio, Qwen3-TTS, or Seedance 2.0 under the same judge) or substantially temper the claim to 'the proposed components improve per internal ablation'.
  3. [§4.4–§4.6, Tables 3–8] No error bars, confidence intervals, or significance tests are reported anywhere. This matters because many headline margins are small: e.g., Table 7 overall Mean MOS 4.22 vs. 4.09 and Table 5 dialogue Expressive Hierarchy 3.85 vs. 3.71. For LLM-judge metrics and small author-constructed sets (SwanBench-Caption has 64 cases), per-item variability could be substantial. Please provide bootstrap confidence intervals or per-item score distributions, and state the number of independent evaluation runs or judge calls.
  4. [§2.3–§2.4 and §4.5–§4.6] Three of the four evaluation settings (SwanBench-Speech, SwanBench-Scene, and SwanBench-Caption) are derived from or filtered by the authors' own pipeline, and the caption training data itself is auto-annotated and filtered by internal components (SwanVerifier, human audit). The manuscript does not report an external validation of caption accuracy, such as a held-out human agreement study on the generated captions or an independent evaluation of SwanVerifier's filtering decisions. Given that caption quality is load-bearing for the instruct task and for the 'complex instruct' claim, I request a direct human validation of caption correctness on a random sample, or an external benchmark of caption-instruct TTS that is not authored by the same team, in addition to InstructorTTSEval.
minor comments (6)
  1. [§4.1] The data counts are not internally consistent at first reading: Section 2 says the current mixture contains approximately 70M caption records, while Section 4.1 says dense caption adaptation uses 70M clean samples but full-mixture training uses 10M SwanData-Caption samples. Please clarify the relationship among these numbers.
  2. [§4.2] The definition of 'Expressive Hierarchy' is intuitive but not operationalized beyond 'emotional variation, vocal dynamics, and scene appropriateness.' Please provide the exact rubric or prompt used for the Gemini judges, plus the judge's temperature/settings, so readers can assess determinism and reproducibility.
  3. [Table 6] The note says results for all models other than SwanTale are taken from the VoxCPM2 paper. Please confirm that the same judge version, prompt, and inference setting were used; otherwise the cross-paper comparisons may not be apples-to-apples.
  4. [Table 2] The 'Nominal Rate' column mixes a 16-bit continuous-scalar assumption with codebook-index counts for VQ systems. This is understandable but should be stated more prominently so the comparison is not overinterpreted.
  5. [§3.5] The GRPO objective uses K=8 trajectories and clipping epsilon is not specified. Please state epsilon and the GRPO learning rate; the current implementation details are incomplete for reproducibility.
  6. [§4.2] SwanBench-Scene contains 180 instructions and SwanBench-Caption 64 cases. Both are small; please include per-item score distributions or inter-annotator agreement for the human MOS, and similarly for the judge scores, to support the reported means.

Circularity Check

2 steps flagged

Expressiveness claims rest on self-authored LLM-judged benchmarks; GRPO attribute reward reuses the same verifier that filtered training captions

specific steps
  1. self citation load bearing [Section 4.2, Zero-shot task; Table 5; references [74] and [32]]
    "SwanBench-Speech [74] provides monologue and dialogue test cases paired with reference audio. ... Both expressiveness metrics are scored by Gemini 3 Pro [32, 74]. Expressive Richness assesses emotional resonance, character portrayal, and storytelling."

    The abstract's central claim—'achieves the best expressiveness scores in both tasks'—is supported by the Richness and Hierarchy columns in Table 5, which come from Gemini 3 Pro scoring SwanBench-Speech. Reference [74] is a benchmark authored by this same team (authors include Changhao Pan, Ruiqi Li, Yu Zhang, Ke Lei, and Xiang Yin). No human correlation, inter-annotator agreement, or calibration against perceptual judgments is reported for these LLM scores. The headline expressiveness result is therefore not an independent external measurement; it is an in-house benchmark scored by an unvalidated proprietary judge, making the central claim load-bearing on a self-citation rather than on an independent instrument.

  2. self definitional [Section 2.4 Data Refinement and Section 3.5 Reward-guided GRPO Post-training]
    "SwanVerifier, a lightweight waveform-grounded attribute model described in Appendix B, checks gender and age-range labels against the vocal stream and flags inconsistent speaker descriptions for removal. ... For the instruct task, SwanVerifier predicts age and gender from the generated speech, and the attribute reward measures agreement with the corresponding attributes in the full caption (Appendix B)."

    SwanVerifier is used at both ends of the training target. In Section 2.4 it filters which age/gender labels survive in SwanData-Caption; in Section 3.5 it is the reward model for instruct speaker-attribute control, measuring agreement between its own predictions on generated audio and caption attributes. GRPO therefore optimizes the model toward labels that were themselves selected by SwanVerifier. Any systematic bias in the verifier is baked into the training data and then reinforced as a reward, so the speaker-attribute control objective is partly self-defined rather than anchored to an independent ground truth. Human audit mitigates but does not break the loop, because the reward still uses the same verifier.

full rationale

The core model derivation is not circular: SwanVAE, the flow-matching DiT, Unified MoE, curriculum learning, and GRPO are described with explicit equations and standard objectives, and the reconstruction comparisons use external corpora (VCTK, GTSinger, FSD50K, MUSDB18-HQ). The instruction-following results also include the external InstructTTSEval benchmark, which provides some independent check. However, the paper's most distinctive expressiveness claims are supported mainly by self-authored benchmarks (SwanBench-Speech, SwanBench-Caption) scored by proprietary LLM judges without human correlation, and the GRPO speaker-attribute reward reuses the same SwanVerifier model that filtered the training captions. These are evaluation/training-target loops rather than algebraic reductions, so the overall circularity is moderate: score 4 rather than higher, because the central generation architecture and several objective results remain independently meaningful.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The paper's central claims depend on the quality of its internal caption data, the stability of LLM-judge scoring, and a set of hand-chosen training and inference thresholds. Most architectural pieces are standard flow matching, MoE, and RL techniques; the nonstandard components (Engram, quality conditioning, GRPO adaptations) carry the highest burden.

free parameters (4)
  • Data quality filtering thresholds = PESQ >= 2.0, STOI >= 0.85, SI-SDR >= 0, MOS >= 2.5
    Hand-chosen thresholds in Section 2.4 define which samples survive to training; changing them changes the data distribution and all downstream metrics.
  • CFG guidance weights = [1.5, 3.0] plus annealing (a, b, p) = (0.6, 0.6, 1.0)
    Selected for inference in Section 4.1; no sensitivity analysis is reported, so there is no evidence that other weights would not change the reported scores.
  • GRPO reward weights = not reported
    Equation 35 aggregates multiple speech-side rewards with hand-set weights; values are not stated, so the reported GRPO gains cannot be independently reproduced or checked.
  • MoE budget hyperparameters = not reported (pmax, pmin, bnull, cmax, cmin)
    Section 3.3 defines the time-dependent budget q(t) and Top-P routing but does not list chosen values; routing behavior and capacity depend on them.
axioms (5)
  • standard math Flow matching and Euler-Maruyama SDE discretization are valid for the noise-to-data convention used.
    Used in Section 3.5 for GRPO likelihoods and in Equation 7 for training; standard techniques, cited to [65] and [91].
  • standard math Gumbel-Softmax reparameterization and Top-P routing are valid for differentiable expert selection.
    Used in Section 3.3 (Equations 19-24); standard method, cited to [46].
  • domain assumption Auto-annotated captions from Seed2.0 Lite plus style matrices are accurate enough to supervise instruct generation.
    Section 2.3; if caption noise is systematic, errors propagate into the trained model. Human audit and SwanVerifier mitigate but do not eliminate this premise.
  • domain assumption LLM judges (Gemini 3 Pro, gemini-3.5-flash) reflect human perceptual quality and instruction-following accuracy.
    Section 4.2 defines metrics scored by LLMs; no human correlation, inter-annotator agreement, or calibration is reported, so this is a load-bearing evaluation assumption.
  • ad hoc to paper Engram hash-memory layers improve caption conditioning enough to justify the added complexity.
    Section 3.2 introduces Engram as an enhancement, but no ablation or sensitivity analysis is provided; the claim that it strengthens fixed-pattern recognition is untested in this system.
invented entities (2)
  • SwanBench-Caption no independent evidence
    purpose: Evaluate complex instruction-based speech and audio generation (64 cases)
    New benchmark authored by the same team; no external validation or release is provided, so it offers no falsifiable handle outside this paper.
  • SwanVerifier no independent evidence
    purpose: Waveform-grounded check of perceived age and gender labels in captions
    Internal filtering component; accuracy reported only on a held-out split in Appendix B, no release or external evaluation.

pith-pipeline@v1.3.0-daily-deepseek · 32472 in / 12066 out tokens · 112576 ms · 2026-08-04T16:23:51.162213+00:00 · methodology

0 comments
read the original abstract

Speech and audio generation is often needed in animation dubbing, audio drama, movies, advertising, games, podcasts, and short-video production. In these scenarios, creators may need to design voices without reference recordings, control speaker styles with natural language, support acoustic scenes with environments and audio effects, and later reuse the designed voices. Therefore, it is important to support multi-speaker speech and audio generation for both instruct and zero-shot tasks. The instruct task requires a caption of the environment, speaker styles, and fine-grained content, while the zero-shot task uses reference audio together with the same fine-grained content. We address these tasks from both the data and model sides. First, we propose SwanData-Caption, which cleans raw speech and audio data, adds targeted synthetic coverage, and annotates diverse and accurate multi-level captions. Then, we propose SwanTale, a multi-speaker expressive speech and audio generation model that supports both zero-shot and instruct tasks. We introduce SwanVAE to support high-quality multi-audio-modality generation. Then, we adopt reward-conditioned quality control and Engram conditioning, along with Unified MoE for multi-task and multi-audio-modality modeling. In addition, we use curriculum learning and GRPO post-training to let the model progressively learn and strengthen its capabilities. Experimental results show that SwanTale leads on multiple key zero-shot and instruct metrics, achieves the best expressiveness scores in both tasks, and supports complex instruct generation involving multi-speaker speech and audio. Demos can be found at https://swanaigc.github.io/\#swantale.

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

123 extracted references · 1 canonical work pages

  1. [1]

    Funaudiollm: Voice understanding and generation foundation models for natural interaction between humans and llms.arXiv preprint arXiv:2407.04051, 2024

    Keyu An, Qian Chen, Chong Deng, Zhihao Du, Changfeng Gao, Zhifu Gao, Yue Gu, Ting He, Hangrui Hu, Kai Hu, et al. Funaudiollm: Voice understanding and generation foundation models for natural interaction between humans and llms.arXiv preprint arXiv:2407.04051, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    Ultimate vocal remover.GitHub repository, 2020

    Anjok07 and aufr33. Ultimate vocal remover.GitHub repository, 2020. URL https://github.com/Anjok07/ ultimatevocalremovergui

    2020

  3. [3]

    Qwen technical report.arXiv preprint arXiv:2309.16609, 2023

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report.arXiv preprint arXiv:2309.16609, 2023

    Pith/arXiv arXiv 2023

  4. [4]

    Whisperx: Time-accurate speech transcription of long-form audio

    Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. Whisperx: Time-accurate speech transcription of long-form audio. arXiv preprint arXiv:2303.00747, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    Curriculum learning

    Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. InProceedings of the 26th Annual International Conference on Machine Learning, pages 41–48. ACM, 2009

    2009

  6. [6]

    Seed2.0, 2026

    ByteDance Seed Team. Seed2.0, 2026. URLhttps://seed.bytedance.com/en/seed2. Model website

    2026

  7. [7]

    Seed Speech ASR 2.0 Documentation.BytePlus documentation, 2026

    BytePlus. Seed Speech ASR 2.0 Documentation.BytePlus documentation, 2026. URLhttps://docs.byteplus. com/en/docs/byteplusvoice/speechtotextv2. Accessed: 2026-06-01

    2026

  8. [8]

    FlexiVoice: Enabling flexible style control in zero-shot TTS with natural language instructions

    Dekun Chen, Xueyao Zhang, Yuancheng Wang, Kenan Dai, Li Ma, and Zhizheng Wu. FlexiVoice: Enabling flexible style control in zero-shot TTS with natural language instructions. InInternational Conference on Learning Representations, 2026

    2026

  9. [9]

    Gigaspeech: An evolving, multi-domain asr corpus with 10,000 hours of transcribed audio

    Guoguo Chen, Shuzhou Chai, Guanbo Wang, Jiayu Du, Wei-Qiang Zhang, Chao Weng, Dan Su, Daniel Povey, Jan Trmal, Junbo Zhang, et al. Gigaspeech: An evolving, multi-domain asr corpus with 10,000 hours of transcribed audio. arXiv preprint arXiv:2106.06909, 2021

    Pith/arXiv arXiv 2021

  10. [10]

    Wavlm: Large-scale self-supervised pre-training for full stack speech processing

    Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, et al. Wavlm: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing, 16(6):1505–1518, 2022

    2022

  11. [11]

    3d-speaker-toolkit: An open-source toolkit for multimodal speaker verifi- cation and diarization

    Yafeng Chen, Siqi Zheng, Hui Wang, Luyao Cheng, Tinglong Zhu, Rongjie Huang, Chong Deng, Qian Chen, Shiliang Zhang, Wen Wang, et al. 3d-speaker-toolkit: An open-source toolkit for multimodal speaker verifi- cation and diarization. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025

    2025

  12. [12]

    SeniorTalk: A chinese conversation dataset with rich annotations for super-aged seniors

    Yang Chen, Hui Wang, Shiyao Wang, Junyang Chen, Jiabei He, Jiaming Zhou, Xi Yang, Yequan Wang, Yonghua Lin, and Yong Qin. SeniorTalk: A chinese conversation dataset with rich annotations for super-aged seniors. arXiv preprint arXiv:2503.16578, 2025. URLhttps://arxiv.org/abs/2503.16578

    arXiv 2025

  13. [13]

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching.arXiv preprint arXiv:2410.06885, 2024

    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

    Pith/arXiv arXiv 2024

  14. [14]

    Conditional memory via scalable lookup: A new axis of sparsity for large language models.arXiv preprint arXiv:2601.07372, 2026

    Xin Cheng, Wangding Zeng, Damai Dai, Qinyu Chen, Bingxuan Wang, Zhenda Xie, Kezhao Huang, Xingkai Yu, Zhewen Hao, Yukun Li, Han Zhang, Huishuai Zhang, Dongyan Zhao, and Wenfeng Liang. Conditional memory via scalable lookup: A new axis of sparsity for large language models.arXiv preprint arXiv:2601.07372, 2026

    Pith/arXiv arXiv 2026

  15. [15]

    Michael Chinen, Felicia S. C. Lim, Jan Skoglund, Nikita Gureev, Feargus O’Gorman, and Andrew Hines. ViSQOL v3: An open source production ready objective speech and audio metric. In2020 TwelfthInternationalConference on Quality of Multimedia Experience (QoMEX), pages 1–6, 2020. doi: 10.1109/QoMEX48832.2020.9123150

    arXiv 2020

  16. [16]

    An unsupervised autoregressive model for speech representation learning

    Yu-An Chung, Wei-Ning Hsu, Hao Tang, and James Glass. An unsupervised autoregressive model for speech representation learning. arXiv preprint arXiv:1904.03240, 2019

    Pith/arXiv arXiv 1904

  17. [17]

    GLM-4-Voice: Towards intelligent and human-like end-to-end spoken chatbot.arXiv preprint arXiv:2507.01006, 2025

    Jiayan Cui, Zhihan Yang, Naihan Li, Jiankun Tian, Xingyu Ma, Yi Zhang, Guangyu Chen, Runxuan Yang, Yuqing Cheng, Yizhi Zhou, Guochen Yu, Xiaotao Gu, and Jie Tang. GLM-4-Voice: Towards intelligent and human-like end-to-end spoken chatbot.arXiv preprint arXiv:2507.01006, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Damai Dai, Chengqi Deng, Chenggang Zhao, R. X. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, Zhenda Xie, Y. K. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, and Wenfeng Liang. DeepSeekMoE: Towards ultimate expert specialization in mixture-of-experts language models.arXiv preprint arXiv:2401.06066, 2024. 24

    Pith/arXiv arXiv 2024

  19. [19]

    High fidelity neural audio compression.arXiv preprint arXiv:2210.13438, 2022

    Alexandre Defossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High fidelity neural audio compression.arXiv preprint arXiv:2210.13438, 2022

    Pith/arXiv arXiv 2022

  20. [20]

    ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification

    Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck. ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification. InInterspeech, pages 3830–3834, 2020

    2020

  21. [21]

    Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens

    Zhihao Du, Qian Chen, Shiliang Zhang, Kai Hu, Heng Lu, Yexin Yang, Hangrui Hu, Siqi Zheng, Yue Gu, Ziyang Ma, et al. Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens. arXiv preprint arXiv:2407.05407, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Cosyvoice 3: Towards in-the-wild speech generation via scaling-up and post-training.arXiv preprint arXiv:2505.17589, 2025

    Zhihao Du, Changfeng Gao, Yuxuan Wang, Fan Yu, Tianyu Zhao, Hao Wang, Xiang Lv, Hui Wang, Chongjia Ni, Xian Shi, et al. Cosyvoice 3: Towards in-the-wild speech generation via scaling-up and post-training.arXiv preprint arXiv:2505.17589, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Stable audio open

    Zach Evans, Julian D Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio open. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025

    2025

  24. [24]

    Stable audio 3

    Zach Evans, Julian D Parker, Matthew Rice, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio 3. arXiv preprint arXiv:2605.17991, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    Falk, Chenxi Zheng, and Wai-Yip Chan

    Tiago H. Falk, Chenxi Zheng, and Wai-Yip Chan. A non-intrusive quality and intelligibility measure of reverberant and dereverberated speech.IEEE Transactions on Audio, Speech, and Language Processing, 18(7):1766–1774,

  26. [26]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.Journal of Machine Learning Research, 23(120):1–39, 2022

    William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.Journal of Machine Learning Research, 23(120):1–39, 2022

    2022

  27. [27]

    Rectifying the emotional flow: Aligning priors and dynamic guidance for high-arousal text-to-speech

    Fangming Feng, Dongjie Fu, Zequn Xie, Yu Zhang, Yangyang Wu, Zhou Zhao, and Tao Jin. Rectifying the emotional flow: Aligning priors and dynamic guidance for high-arousal text-to-speech. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 21874–21888, San Diego, California, United States,...

    work page doi:10.18653/v1/2026.acl-long.998 2026

  28. [28]

    Fish-Speech: Leveraging large language models for advanced multilingual text-to-speech synthesis

    Fish Audio Team. Fish-Speech: Leveraging large language models for advanced multilingual text-to-speech synthesis. arXiv preprint arXiv:2411.01156, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Fsd50k: an open dataset of human-labeled sound events.IEEE/ACM Transactionson Audio, Speech, and Language Processing, 30:829–852, 2021

    Eduardo Fonseca, Xavier Favory, Jordi Pons, Frederic Font, and Xavier Serra. Fsd50k: an open dataset of human-labeled sound events.IEEE/ACM Transactionson Audio, Speech, and Language Processing, 30:829–852, 2021

    2021

  30. [30]

    Ace-step: A step towards music generation foundation model

    Junmin Gong, Sean Zhao, Sen Wang, Shengyuan Xu, and Joe Guo. Ace-step: A step towards music generation foundation model. arXiv preprint arXiv:2506.00045, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Gemini 2.5 Pro model card, 2025

    Google DeepMind. Gemini 2.5 Pro model card, 2025. URLhttps://storage.googleapis.com/deepmind-media/ Model-Cards/Gemini-2-5-Pro-Model-Card.pdf. Model card

    2025

  32. [32]

    Gemini 3 Pro model card, 2025

    Google DeepMind. Gemini 3 Pro model card, 2025. URLhttps://deepmind.google/models/model-cards/ gemini-3-pro/. Model card

    2025

  33. [33]

    Gemini 3.5 Flash model card, 2026

    Google DeepMind. Gemini 3.5 Flash model card, 2026. URLhttps://deepmind.google/models/model-cards/ gemini-3-5-flash/. Model card

    2026

  34. [34]

    Gray and John D

    Augustine H. Gray and John D. Markel. Distance measures for speech processing.IEEE Transactions on Acoustics, Speech, and Signal Processing, 24(5):380–391, 1976. doi: 10.1109/TASSP.1976.1162849

    arXiv 1976

  35. [35]

    MRSAudio: A large-scale multimodal recorded spatial audio dataset with refined annotations

    Wenxiang Guo, Changhao Pan, Zhiyuan Zhu, Xintong Hu, Yu Zhang, Li Tang, Rui Yang, Han Wang, Zongbao Zhang, Yuhan Wang, Yixuan Chen, Hankun Xu, Ke Xu, Pengfei Fan, Zhetao Chen, Yanhao Yu, Qiange Huang, Fei Wu, and Zhou Zhao. MRSAudio: A large-scale multimodal recorded spatial audio dataset with refined annotations. In Advances in Neural Information Process...

    2025

  36. [36]

    Techsinger: Technique controllable multilingual singing voice synthesis via flow matching.arXiv preprint arXiv:2502.12572, 2025

    Wenxiang Guo, Yu Zhang, Changhao Pan, Rongjie Huang, Li Tang, Ruiqi Li, Zhiqing Hong, Yongqi Wang, and Zhou Zhao. Techsinger: Technique controllable multilingual singing voice synthesis via flow matching.arXiv preprint arXiv:2502.12572, 2025. 25

    Pith/arXiv arXiv 2025

  37. [37]

    STARS: A unified framework for singing transcription, alignment, and refined style annotation

    Wenxiang Guo, Yu Zhang, Changhao Pan, Zhiyuan Zhu, Ruiqi Li, ZheTao Chen, Wenhao Xu, Fei Wu, and Zhou Zhao. STARS: A unified framework for singing transcription, alignment, and refined style annotation. In Findings of the Association for Computational Linguistics: ACL 2025, pages 15081–15093, Vienna, Austria,

    2025

  38. [38]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Pith/arXiv arXiv 2022

  39. [39]

    VoiceSculptor: Your voice, designed by you.arXiv preprint arXiv:2601.10629, 2026

    Jingbin Hu, Huakang Chen, Linhan Ma, Dake Guo, Qirui Zhan, Wenhao Li, Haoyu Zhang, Kangxiang Xia, Ziyu Zhang, Wenjie Tian, Chengyou Wang, Jinrui Liang, Shuhan Guo, Zihang Yang, Bengu Wu, Binbin Zhang, Pengcheng Zhu, Pengyuan Xie, Chuan Xie, Qiang Zhang, Jie Liu, and Lei Xie. VoiceSculptor: Your voice, designed by you.arXiv preprint arXiv:2601.10629, 2026

    arXiv 2026

  40. [40]

    Word level timestamp generation for automatic speech recognition and translation

    Ke Hu, Krishna Puvvada, Elena Rastorgueva, Zhehuai Chen, He Huang, Shuoyang Ding, Kunal Dhawan, Hainan Xu, Jagadeesh Balam, and Boris Ginsburg. Word level timestamp generation for automatic speech recognition and translation. arXiv preprint arXiv:2505.15646, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    python-pinyin: pypinyin, 2023

    Huang Huang, Xiaoquan Kong, et al. python-pinyin: pypinyin, 2023. URLhttps://github.com/mozillazg/ python-pinyin. Software, version 0.48.0

    2023

  42. [42]

    InstructTTSEval: Benchmarking complex natural-language instruction following in text-to-speech systems

    Kexin Huang, Qian Tu, Liwei Fan, Chenchen Yang, Dong Zhang, Shimin Li, Zhaoye Fei, Qinyuan Cheng, and Xipeng Qiu. InstructTTSEval: Benchmarking complex natural-language instruction following in text-to-speech systems. arXiv preprint arXiv:2506.16381, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    MOSS-VoiceGenerator: Create realistic voices with natural language descriptions.arXiv preprint arXiv:2603.28086, 2026

    Kexin Huang, Liwei Fan, Botian Jiang, Yaozhou Jiang, Qian Tu, Jie Zhu, Yuqian Zhang, Yiwei Zhao, Chenchen Yang, Zhaoye Fei, Shimin Li, Xiaogui Yang, Qinyuan Cheng, and Xipeng Qiu. MOSS-VoiceGenerator: Create realistic voices with natural language descriptions.arXiv preprint arXiv:2603.28086, 2026

    arXiv 2026

  44. [44]

    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 (NeurIPS), 2022

    2022

  45. [45]

    Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models

    Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In International Conference on Machine Learning, pages 13916–13932. PMLR, 2023

    2023

  46. [46]

    Categorical reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144, 2016

    Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144, 2016

    Pith/arXiv arXiv 2016

  47. [47]

    Univnet: A neural vocoder with multi- resolution spectrogram discriminators for high-fidelity waveform generation.arXiv preprint arXiv:2106.07889, 2021

    Won Jang, Dan Lim, Jaesam Yoon, Bongwan Kim, and Juntae Kim. Univnet: A neural vocoder with multi- resolution spectrogram discriminators for high-fidelity waveform generation.arXiv preprint arXiv:2106.07889, 2021

    Pith/arXiv arXiv 2021

  48. [48]

    Wavtokenizer: an efficient acoustic discrete codec tokenizer for audio language modeling

    Shengpeng Ji, Ziyue Jiang, Wen Wang, Yifu Chen, Minghui Fang, Jialong Zuo, Qian Yang, Xize Cheng, Zehan Wang, Ruiqi Li, et al. Wavtokenizer: an efficient acoustic discrete codec tokenizer for audio language modeling. arXiv preprint arXiv:2408.16532, 2024

    Pith/arXiv arXiv 2024

  49. [49]

    Latent-domain predictive neural speech coding

    Xue Jiang, Xiulian Peng, Huaying Xue, Yuan Zhang, and Yan Lu. Latent-domain predictive neural speech coding. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2111–2123, 2023

    2023

  50. [50]

    Mega-tts 2: Boosting prompting mechanisms for zero-shot speech synthesis

    Ziyue Jiang, Jinglin Liu, Yi Ren, Jinzheng He, Zhenhui Ye, Shengpeng Ji, Qian Yang, Chen Zhang, Pengfei Wei, Chunfeng Wang, et al. Mega-tts 2: Boosting prompting mechanisms for zero-shot speech synthesis. InThe TwelfthInternational Conference on Learning Representations, 2024

    2024

  51. [51]

    Megatts 3: Sparse alignment enhanced latent diffusion transformer for zero-shot speech synthesis

    Ziyue Jiang, Yi Ren, Ruiqi Li, Shengpeng Ji, Boyang Zhang, Zhenhui Ye, Chen Zhang, Bai Jionghao, Xiaoda Yang, Jialong Zuo, et al. Megatts 3: Sparse alignment enhanced latent diffusion transformer for zero-shot speech synthesis. arXiv preprint arXiv:2502.18924, 2025

    Pith/arXiv arXiv 2025

  52. [52]

    Mooncast: High-quality zero-shot podcast generation.arXiv preprint arXiv:2503.14345, 2025

    Zeqian Ju, Dongchao Yang, Jianwei Yu, Kai Shen, Yichong Leng, Zhengtao Wang, Xu Tan, Xinyu Zhou, Tao Qin, and Xiangyang Li. Mooncast: High-quality zero-shot podcast generation.arXiv preprint arXiv:2503.14345, 2025

    Pith/arXiv arXiv 2025

  53. [53]

    Libriheavy: A 50,000 hours asr corpus with punctuation casing and context

    Wei Kang, Xiaoyu Yang, Zengwei Yao, Fangjun Kuang, Yifan Yang, Liyong Guo, Long Lin, and Daniel Povey. Libriheavy: A 50,000 hours asr corpus with punctuation casing and context. InICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10991–10995. IEEE, 2024. 26

    2024

  54. [54]

    AudioCaps: Generating captions for audios in the wild

    Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. AudioCaps: Generating captions for audios in the wild. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 119–132. Association for Computational Linguistics, 2019. doi: 10.18653/v1/N19-1011

    doi:10.18653/v1/n19-1011 2019

  55. [55]

    Best-worst scaling more reliable than rating scales: A case study on sentiment intensity annotation

    Svetlana Kiritchenko and Saif Mohammad. Best-worst scaling more reliable than rating scales: A case study on sentiment intensity annotation. InProceedings of the 55th AnnualMeeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 465–470, Vancouver, Canada, 2017. Association for Computational Linguistics

    2017

  56. [56]

    Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis.Advances in neural information processing systems, 33:17022–17033, 2020

    Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis.Advances in neural information processing systems, 33:17022–17033, 2020

    2020

  57. [57]

    Kubichek

    Robert F. Kubichek. Mel-cepstral distance measure for objective speech quality assessment. InProceedings of the IEEE Pacific Rim Conference on Communications, Computers and Signal Processing, volume 1, pages 125–128, 1993. doi: 10.1109/PACRIM.1993.407206

    arXiv 1993

  58. [58]

    Torchaudio-squim: Reference-less speech quality and intelligibility measures in torchaudio

    Anurag Kumar, Ke Tan, Zhaoheng Ni, Pranay Manocha, Xiaohui Zhang, Ethan Henderson, and Buye Xu. Torchaudio-squim: Reference-less speech quality and intelligibility measures in torchaudio. InICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, 2023

    2023

  59. [60]

    High-fidelity audio compression with improved rvqgan.arXiv preprint arXiv:2306.06546, 2023

    Rithesh Kumar, Prem Seetharaman, Alejandro Luebs, Ishaan Kumar, and Kundan Kumar. High-fidelity audio compression with improved rvqgan.arXiv preprint arXiv:2306.06546, 2023

    Pith/arXiv arXiv 2023

  60. [61]

    Parler-TTS.GitHub repository, 2024

    Yoach Lacombe, Vaibhav Srivastav, and Sanchit Gandhi. Parler-TTS.GitHub repository, 2024. URLhttps: //github.com/huggingface/parler-tts

    2024

  61. [62]

    Towards streaming synchronized spatial audio generation via autoregressive diffusion transformer

    Ke Lei, Yu Zhang, Changhao Pan, Xueyi Pu, Wenxiang Guo, Ruiqi Li, and Zhou Zhao. Towards streaming synchronized spatial audio generation via autoregressive diffusion transformer. InProceedings of the 43rd International Conference on Machine Learning, 2026

    2026

  62. [63]

    Robust singing voice transcription serves synthesis.arXiv preprint arXiv:2405.09940, 2024

    Ruiqi Li, Yu Zhang, Yongqi Wang, Zhiqing Hong, Rongjie Huang, and Zhou Zhao. Robust singing voice transcription serves synthesis.arXiv preprint arXiv:2405.09940, 2024

    Pith/arXiv arXiv 2024

  63. [64]

    SwanVoice: Expressive long-form zero-shot speech synthesis for both monologue and dialogue.arXiv preprint arXiv:2605.30993, 2026

    Ruiqi Li, Yu Zhang, Changhao Pan, Ke Lei, Xiang Yin, and Cheng Yang. SwanVoice: Expressive long-form zero-shot speech synthesis for both monologue and dialogue.arXiv preprint arXiv:2605.30993, 2026

    Pith/arXiv arXiv 2026

  64. [65]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations, 2022

    2022

  65. [66]

    Flow-grpo: Training flow matching models via online rl.arXiv preprint arXiv:2505.05470, 2025

    Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. Flow-grpo: Training flow matching models via online rl.arXiv preprint arXiv:2505.05470, 2025

    Pith/arXiv arXiv 2025

  66. [67]

    UniMoE- Audio: Unified speech and music generation with dynamic-capacity MoE.arXiv preprint arXiv:2510.13344, 2025

    Zhenyu Liu, Yunxin Li, Xuanyu Zhang, Qixun Teng, Shenyuan Jiang, Xinyu Chen, Haoyuan Shi, Jinchao Li, Qi Wang, Haolan Chen, Fanbo Meng, Mingjun Zhao, Yu Xu, Yancheng He, Baotian Hu, and Min Zhang. UniMoE- Audio: Unified speech and music generation with dynamic-capacity MoE.arXiv preprint arXiv:2510.13344, 2025

    arXiv 2025

  67. [68]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2019

    2019

  68. [69]

    Natural language guidance of high-fidelity text-to-speech with synthetic annotations

    Dan Lyth and Simon King. Natural language guidance of high-fidelity text-to-speech with synthetic annotations. arXiv preprint arXiv:2402.01912, 2024

    Pith/arXiv arXiv 2024

  69. [70]

    MOSS-TTSD

    MOSI. MOSS-TTSD. Project website, February 2026. URLhttps://mosi.cn/models/moss-ttsd

    2026

  70. [71]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  71. [72]

    A multimodal evaluation framework for spatial audio playback systems: From localization to listener preference

    Changhao Pan, Wenxiang Guo, Yu Zhang, Zhiyuan Zhu, Zhetao Chen, Han Wang, and Zhou Zhao. A multimodal evaluation framework for spatial audio playback systems: From localization to listener preference. InProceedings of the 33rd ACM International Conference on Multimedia, pages 7006–7015. ACM, 2025. 27

    2025

  72. [73]

    Audio editing in the era of foundation models: A survey

    Changhao Pan, Yifei Fan, Fan Zhuo, Yifu Chen, Wenxiang Guo, Yu Zhang, Ruiqi Li, Zhiyuan Zhu, Rui Yang, Shengpeng Ji, Chenyuhao Wen, Jiayang Xu, Ke Lei, Xiaoda Yang, Jingyu Lu, and Zhou Zhao. Audio editing in the era of foundation models: A survey. arXiv preprint arXiv:2606.23139, 2026. URL https: //arxiv.org/abs/2606.23139

    Pith/arXiv arXiv 2026

  73. [74]

    Comprehensive benchmarking of long-form speech generation in diverse scenarios.arXiv preprint arXiv:2605.28618, 2026

    Changhao Pan, Rui Yang, Han Wang, Zhuan Zhou, Xuming He, Wenxiang Guo, Ziyue Jiang, Ruiqi Li, Yu Zhang, Chenyuhao Wen, Ke Lei, Xiang Yin, Jingyu Lu, Zhiyuan Zhu, and Zhou Zhao. Comprehensive benchmarking of long-form speech generation in diverse scenarios.arXiv preprint arXiv:2605.28618, 2026

    Pith/arXiv arXiv 2026

  74. [75]

    Librispeech: an asr corpus based on public domain audio books

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206–5210. IEEE, 2015

    2015

  75. [76]

    Same: A semantically-aligned music autoencoder.arXiv preprint arXiv:2605.18613, 2026

    Julian D Parker, Zach Evans, CJ Carr, Zachary Zukowski, Josiah Taylor, Matthew Rice, and Jordi Pons. Same: A semantically-aligned music autoencoder.arXiv preprint arXiv:2605.18613, 2026

    Pith/arXiv arXiv 2026

  76. [77]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195–4205, 2023

    2023

  77. [78]

    Vibevoice technical report

    Zhiliang Peng, Jianwei Yu, Wenhui Wang, Yaoyao Chang, Yutao Sun, Li Dong, Yi Zhu, Weijiang Xu, Hangbo Bao, Zehua Wang, Shaohan Huang, Yan Xia, and Furu Wei. Vibevoice technical report. arXiv preprint arXiv:2508.19205, 2025

    Pith/arXiv arXiv 2025

  78. [79]

    Qwen3-TTS technical report.arXiv preprint arXiv:2601.15621, 2026

    Qwen Team. Qwen3-TTS technical report.arXiv preprint arXiv:2601.15621, 2026

    Pith/arXiv arXiv 2026

  79. [80]

    Nemo forced aligner and its application to word alignment for subtitle generation

    Elena Rastorgueva, Vitaly Lavrukhin, and Boris Ginsburg. Nemo forced aligner and its application to word alignment for subtitle generation. InInterspeech, pages 5257–5258, 2023

    2023

  80. [81]

    Chandan K. A. Reddy, Vishak Gopal, and Ross Cutler. Dnsmos: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors. InICASSP 2021 - 2021 IEEE InternationalConference on Acoustics, Speech and Signal Processing (ICASSP), pages 6493–6497, 2021

    2021

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.