Pith. sign in

REVIEW 4 major objections 4 minor 35 references

Faster IndexTTS-2 demonstrates that converting an autoregressive zero-shot TTS model to TensorRT/TensorRT-LLM yields up to 3.6x end-to-end speedup and 5.0x GPT speedup with minimal quality degradation, while adding streaming and batching.

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 · deepseek-v4-flash

2026-08-01 08:36 UTC pith:FT3HFXQJ

load-bearing objection Solid engineering paper with real speedups, but the headline numbers likely overstate the gain over a properly tuned PyTorch baseline; the architectural work for TensorRT-LLM is still worth refereeing. the 4 major comments →

arxiv 2607.21042 v1 pith:FT3HFXQJ submitted 2026-07-23 cs.AI

Faster IndexTTS-2: Accelerating and Streaming Autoregressive Zero-Shot Text-to-Speech Synthesis on GPUs

Muyang Du , Shuang Yu , Junjie Lai This is my paper
classification cs.AI
keywords text-to-speechinference accelerationTensorRT-LLMautoregressivezero-shot TTSstreaming synthesisGPU inferenceIndexTTS-2
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Faster IndexTTS-2 aims to show that the quality of a state-of-the-art autoregressive zero-shot TTS system does not have to be traded away for speed: by reimplementing the model's four stages as GPU inference engines and adapting an LLM-serving runtime to the speech token stream, the authors report up to 3.6x end-to-end speedup (5.0x on the autoregressive GPT) with only minor changes in word error rate, speaker similarity, and naturalness. It also introduces streaming synthesis with ~600 ms time-to-first-audio and batched inference across all stages, features the original PyTorch model lacked. The paper positions this as a reusable recipe, so a sympathetic reader is intended to take away that autoregressive TTS can be made interactive and production-viable on a single GPU.

Core claim

The central claim: IndexTTS-2, an autoregressive TTS model whose inference barely reached real-time, becomes production-deployable by converting all components to TensorRT and its GPT to TensorRT-LLM. The key adaptations are prompt-tuning injection of conditioning embeddings, merged text/codec vocabularies with shifted IDs, custom position IDs, and an extra output for per-step GPT hidden states. Measurements show up to 3.60x end-to-end and 5.0x GPT speedups over the PyTorch FP32 baseline, with WER +0.1, SIM-o -0.01, UTMOS -0.1, plus ~600ms streaming TTFA and batched throughput gains.

What carries the argument

The load-bearing mechanism is the TensorRT/TensorRT-LLM conversion pipeline adapted to IndexTTS-2: preprocessing, DiT, and vocoder become TensorRT engines via an ONNX workflow with dynamic shapes, while the decoder-only GPT is served by TensorRT-LLM and modified to accept IndexTTS-2's mixed inputs. Four adaptations carry the argument: storing 32 speaker latents, emotion, and speed embeddings in a prompt-embedding table; merging text and semantic-codec embedding tables with shifted token IDs; building custom position IDs (conditioning gets empty positions, text and semantics get separate ranges); and registering an extra output for per-step hidden states. This machinery is what lets a general

Load-bearing premise

The headline speedups assume the PyTorch FP32 baseline is a fairly optimized comparison point; if that baseline was run without standard optimizations such as CUDA graphs or tuned kernels, the reported 3.6x and 5.0x gains would overstate Faster IndexTTS-2's advantage over a well-tuned PyTorch implementation.

What would settle it

Re-run the original IndexTTS-2 in PyTorch FP32 with standard optimizations (torch.compile, CUDA graphs, fused kernels, tuned batch sizes) on the same A100 and compare end-to-end latency and GPT latency against Faster IndexTTS-2's published numbers; if the gap narrows markedly, the claimed speedups reflect baseline weakness, not intrinsic superiority. Also, listen for audible seams in 1 s streaming chunks at overlap settings beyond 5 frames.

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

If this is right

  • At FP16 or W8A16, the accelerated pipeline reaches RTF 0.21-0.24 on A100, roughly 4-5x faster than real-time, enabling interactive voice agents.
  • Chunked streaming with a 2 s chunk size cuts TTFA to about 600 ms while keeping WER, SIM-o, and UTMOS close to non-streaming values, with 1 s chunks trading a little naturalness for lower latency.
  • Batching raises throughput up to ~1.8x at batch size 8-16, with diminishing returns beyond 8 and higher TTFA in streaming mode at large batches.
  • Weight-only quantization of the GPT to W8A16 preserves quality; W4A16 is usable with slightly larger degradation, providing memory-flexible deployment options.
  • The four GPT adaptations are transferable to other autoregressive speech models, making the method a general template rather than a one-off.

Where Pith is reading between the lines

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

  • Editorial inference: because the paper's own data show PyTorch FP16 running slower than FP32, the PyTorch FP32 baseline is probably not an optimized reference; a well-tuned PyTorch baseline might narrow the reported 3.6x/5.0x gap, although the absolute RTF values would remain meaningful.
  • Editorial inference: the prompt-tuning trick for injecting speaker/emotion/speed latents suggests autoregressive TTS can be served through the same runtime as LLMs, which could lower the engineering cost of voice assistants that generate text and speech tokens together.
  • Editorial inference: since streaming degradation appears mainly in UTMOS (naturalness) rather than WER or speaker similarity at small chunk sizes, audible seams rather than content errors are the limiting factor; improved boundary blending could make 1 s chunks nearly lossless.

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 / 4 minor

Summary. The paper presents Faster IndexTTS-2, a production-oriented acceleration of the autoregressive TTS model IndexTTS-2 using NVIDIA TensorRT and TensorRT-LLM. The authors describe adaptations of TensorRT-LLM to handle IndexTTS-2's mixed text/semantic-token inputs, conditioning prompts, custom position IDs, and per-step hidden-state outputs. They report up to 5.0× speedup on the autoregressive GPT and up to 3.6× end-to-end speedup versus the original PyTorch implementation, with minimal degradation in WER, speaker similarity, and UTMOS. They also enable streaming synthesis with chunked decoding and batched inference across all pipeline stages. Experiments on Seed-TTS English and Chinese test sets compare latency, RTF, TTFA, throughput, and quality across precisions, streaming configurations, and batch sizes.

Significance. If the speedup and quality-preservation claims are substantiated, the work has significant practical value: it turns a barely-real-time research model into a deployable system with streaming and batching support, and it provides a reusable methodology for adapting TensorRT-LLM to autoregressive speech models. The engineering contributions—prompt-table conditioning, merged embedding tables, custom position-ID handling, hidden-state output registration, and chunked streaming with overlap—are concrete and potentially transferable. The paper also offers a relatively thorough benchmark across languages, precisions, chunk/overlap settings, and batch sizes. However, the central quantitative claims rest on a comparison against an unoptimized PyTorch baseline, and the internal consistency of several reported latency/quality numbers needs to be established.

major comments (4)
  1. [§III-B, Table I] The speedup claims (5.0× GPT, 3.6× end-to-end) are computed relative to a PyTorch baseline whose configuration is not disclosed. The paper does not state whether the baseline uses standard inference optimizations such as torch.compile, CUDA graphs, or autocast FP16/BF16. The observation that PyTorch FP16 is over 20% slower than FP32 on an A100 is a strong indication that the baseline is not tuned; on modern PyTorch with CUDA graphs and FP16 autocast, transformer inference typically becomes faster, not slower. A well-tuned PyTorch baseline would likely reduce the reported speedups. The authors should either benchmark against a reasonably optimized PyTorch baseline and report the resulting ratios, or explicitly frame the comparison as 'vs. the original unoptimized IndexTTS-2 implementation' and temper the headline claims accordingly.
  2. [§III-B, text vs. Table I] The quality-degradation figures in the text are inconsistent with Table I. The text states WER increases by 0.13 and 0.07 percentage points on English and Chinese, but Table I shows differences of 0.19 (1.84→2.03) and 0.08 (1.01→1.09) between FP32 baseline and Faster FP16. Furthermore, no variance, confidence intervals, or significance tests are reported for WER, SIM-o, or UTMOS. Given the test-set sizes (~1,088 and ~2,020 utterances), a WER difference of 0.19 or even 0.13 could plausibly arise from noise. The claim of 'minimal degradation' is load-bearing and needs statistical support (e.g., bootstrap confidence intervals or paired tests).
  3. [§III-C/D, Tables II and III] Several latency values are internally inconsistent across tables. For Faster FP16 non-streaming English, Table I reports overall latency 874.5 ms and WER 2.03; Table II reports the same latency but WER 1.97. More seriously, Table III (BS=1, non-streaming) reports latency 1097.0 ms and RTF 0.2232, which cannot be reconciled with Table I's 874.5 ms and RTF 0.24 for the same configuration. Similarly, streaming BS=1 in Table III gives 1492.9 ms total latency, whereas Table II's 2s-chunk streaming entry gives 1180.4 ms. The paper does not specify the test set or utterance lengths used for Table III. If Table III uses a different subset or includes additional overhead, that must be stated. As written, these discrepancies undermine the reliability of the efficiency measurements.
  4. [§II-C, GPT acceleration] The TensorRT-LLM adaptations are described only at a high level. For a paper whose stated contribution includes a 'reusable methodology', the implementation details are insufficient: how exactly are the virtual input IDs and prompt embedding table configured? How are the merged text/codec embedding tables and shifted token IDs handled at runtime? How are the custom position IDs constructed, and how is the additional hidden-state output tensor registered in the engine? Without these details, the methodology is not reproducible and its transferability to other autoregressive TTS models cannot be assessed. Please add a more detailed description or pseudocode.
minor comments (4)
  1. [Table I] Typo: 'V ocoder' should be 'Vocoder' in the header. Also, the notation W8A16 and W4A16 is used without definition; please define in the caption or text that W8A16 means 8-bit weight-only quantization with 16-bit activations.
  2. [Fig. 2] The axis labels indicate 'Chunk size (codec frames)' and 'Overlap size (codec frames)' but the numeric tick values are 1–6 and 25–100, respectively. This is confusing; the mapping from codec frames to seconds (e.g., 100 frames ≈ 2s) should be made explicit, and the figure should use consistent units.
  3. [§II-D, streaming] The phrase 'Hann window cross-fading' is fine, but the implementation of overlapping chunks (e.g., alignment of codec frames to audio samples) is not described. A brief note would improve reproducibility.
  4. [§III-B, baseline] The statement 'simply reducing precision in PyTorch does not guarantee faster inference' is a useful observation, but it would be more informative if the paper reported whether the baseline used autocast and whether any CUDA graph or compile was attempted. A short discussion of the baseline's eager-mode nature would help contextualize the speedup numbers.

Circularity Check

0 steps flagged

No significant circularity: speedups are measured against external baselines, and self-citations are not load-bearing.

full rationale

The paper's claimed contribution is an engineering acceleration of IndexTTS-2 via TensorRT and TensorRT-LLM (Sections II.C-II.D), evaluated by direct latency/throughput/quality measurement on the Seed-TTS benchmark (Section III). The speedup figures in Table I are ratios of measured latencies (PyTorch FP32 vs TRT/TRT-LLM), not quantities derived from fitted parameters or self-referential equations. No self-definitional, fitted-input-called-prediction, ansatz-smuggling, or renaming pattern applies. The self-citations [8] and [13] are prior works by overlapping authors, but they are used only as related-work examples in the introduction, not as evidence for the speedup or as a uniqueness theorem; hence they are not load-bearing. The only concern one could raise is that the PyTorch FP32 baseline may not be optimally tuned (the paper itself notes PyTorch FP16 is over 20% slower than FP32, indicating a suboptimal baseline configuration), but under the reviewing rules that is a correctness/fairness concern about baseline optimization, not circularity: the comparison is still an external empirical benchmark rather than a reduction of the claim to its own inputs. Consequently the resulting circularity score is minimal.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

No fitted parameters are introduced; the system is a compilation and orchestration of existing components. The central assumptions are the validity of the benchmark procedure and the fairness of the PyTorch baseline comparison.

axioms (3)
  • domain assumption TensorRT and TensorRT-LLM correctly implement the ONNX export and the custom modifications, so the measured speedups reflect framework efficiency rather than bugs that change outputs.
    The paper assumes the accelerated engines are functionally equivalent to the original PyTorch models; correctness is only indirectly checked via quality metrics.
  • domain assumption The Seed-TTS test sets and the automatic metrics (WER, SIM-o, UTMOS) are valid proxies for speech synthesis quality in production.
    The 'minimal degradation' claim relies entirely on these metrics, which are single-run objective scores without human evaluation.
  • domain assumption The baseline IndexTTS-2 PyTorch implementation is the standard, unmodified reference.
    The paper treats the PyTorch FP32 numbers in Table I as the original model's performance, without detailing what optimizations are present or absent in that baseline.

pith-pipeline@v1.3.0-alltime-deepseek · 8767 in / 7786 out tokens · 72909 ms · 2026-08-01T08:36:23.018660+00:00 · methodology

0 comments
read the original abstract

Autoregressive text-to-speech models achieve strong naturalness but suffer from slow inference due to sequential token generation, limiting their deployment in production applications that require low latency. IndexTTS-2 is a state-of-the-art autoregressive TTS model consisting of a GPT, a flow-matching Diffusion Transformer, and a vocoder. Despite its high synthesis quality, its inference speed barely reaches real-time without streaming or batching support. We present Faster IndexTTS-2, which accelerates all neural network components of IndexTTS-2 for production deployment on GPUs using NVIDIA TensorRT and TensorRT-LLM. Faster IndexTTS-2 also enables streaming synthesis for latency-sensitive interactive applications, and batched inference across all components to maximize GPU utilization. Experiments on the Seed-TTS benchmark for both English and Chinese demonstrate up to 5.0$\times$ speedup on the autoregressive GPT and 3.6$\times$ end-to-end, with minimal degradation in word error rate, speaker similarity, and naturalness. Our methodology provides a practical reference for efficiently accelerating similar autoregressive speech models on GPUs.

Figures

Figures reproduced from arXiv: 2607.21042 by Junjie Lai, Muyang Du, Shuang Yu.

Figure 1
Figure 1. Figure 1: Overview of Faster IndexTTS-2. In this paper, we present Faster IndexTTS-2, a fully accelerated IndexTTS-2 on GPUs for production deployment using TensorRT and TensorRT-LLM. Our main contributions are as follows: • We accelerate all neural network components of IndexTTS-2 using TensorRT and TensorRT-LLM, achieving up to 3.6× end-to-end speedup and up to 5.0× speedup on the autoregressive GPT. • We present … view at source ↗
Figure 2
Figure 2. Figure 2: Streaming synthesis quality and latency of Faster IndexTTS-2 in FP16 across chunk sizes and overlap sizes on the Seed-TTS test sets. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

35 extracted references · 11 linked inside Pith

  1. [1]

    IndexTTS2: A breakthrough in emotionally expressive and duration- controlled auto-regressive zero-shot text-to-speech,

    S. Zhou, Y . Zhou, Y . He, X. Zhou, J. Wang, W. Deng, and J. Shu, “IndexTTS2: A breakthrough in emotionally expressive and duration- controlled auto-regressive zero-shot text-to-speech,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 41, 2026, pp. 35 139–35 148

    2026

  2. [2]

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

    Z. Du, Q. Chen, S. Zhang, K. Hu, H. Lu, Y . Yang, H. Hu, S. Zheng, Y . Gu, Z. Maet al., “CosyV oice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens,”arXiv preprint arXiv:2407.05407, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    CosyV oice 2: Scalable streaming speech synthesis with large language models,

    Z. Du, Y . Wang, Q. Chen, X. Shi, X. Lv, T. Zhao, Z. Gao, Y . Yang, C. Gao, H. Wanget al., “CosyV oice 2: Scalable streaming speech synthesis with large language models,”arXiv preprint arXiv:2412.10117, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Fish-speech: Leveraging large language models for advanced multilin- gual text-to-speech synthesis,

    S. Liao, Y . Wang, T. Li, Y . Cheng, R. Zhang, R. Zhou, and Y . Xing, “Fish-speech: Leveraging large language models for advanced multilin- gual text-to-speech synthesis,”arXiv preprint arXiv:2411.01156, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Spark-TTS: An efficient LLM-based text- to-speech model with single-stream decoupled speech tokens,

    X. Wang, M. Jiang, Z. Ma, Z. Zhang, S. Liu, L. Li, Z. Liang, Q. Zheng, R. Wang, X. Fenget al., “Spark-TTS: An efficient LLM-based text- to-speech model with single-stream decoupled speech tokens,”arXiv preprint arXiv:2503.01710, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Qwen3-TTS technical report,

    H. Hu, X. Zhu, T. He, D. Guo, B. Zhang, X. Wang, Z. Guo, Z. Jiang, H. Hao, Z. Guoet al., “Qwen3-TTS technical report,”arXiv preprint arXiv:2601.15621, 2026

    Pith/arXiv arXiv 2026

  7. [7]

    FireRedTTS: A foundation text-to-speech framework for industry-level generative speech applications,

    H.-H. Guo, Y . Hu, K. Liu, F.-Y . Shen, X. Tang, Y .-C. Wu, F.-L. Xie, K. Xie, and K.-T. Xu, “FireRedTTS: A foundation text-to-speech framework for industry-level generative speech applications,”arXiv preprint arXiv:2409.03283, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Streaming T5-based text-to-speech synthesis with limited lookahead,

    M. Du, J. Roche, and J. Lai, “Streaming T5-based text-to-speech synthesis with limited lookahead,”arXiv preprint arXiv:2606.21882, 2026

    Pith/arXiv arXiv 2026

  9. [9]

    F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching,

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. Zhao, K. Yu, and X. Chen, “F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, 2025, pp. 6255–6271

    2025

  10. [10]

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

    Y . Wang, H. Zhan, L. Liu, R. Zeng, H. Guo, J. Zheng, Q. Zhang, X. Zhang, S. Zhang, and Z. Wu, “MaskGCT: Zero-shot text-to-speech with masked generative codec transformer,” inProc. ICLR, 2025

    2025

  11. [11]

    E2 TTS: Embarrassingly easy fully non-autoregressive zero-shot TTS,

    S. E. Eskimez, X. Wang, M. Thakker, C. Li, C.-H. Tsai, Z. Xiao, H. Yang, Z. Zhu, M. Tang, X. Tanet al., “E2 TTS: Embarrassingly easy fully non-autoregressive zero-shot TTS,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 682–689

    2024

  12. [12]

    ZipV oice: Fast and high-quality zero-shot text-to-speech with flow matching,

    H. Zhu, W. Kang, Z. Yao, L. Guo, F. Kuang, Z. Li, W. Zhuang, L. Lin, and D. Povey, “ZipV oice: Fast and high-quality zero-shot text-to-speech with flow matching,” in2025 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2025, pp. 1–8

    2025

  13. [13]

    InstantSpeech: Instant synchronous text- to-speech synthesis for LLM-driven voice chatbots,

    M. Du, C. Liu, and J. Lai, “InstantSpeech: Instant synchronous text- to-speech synthesis for LLM-driven voice chatbots,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  14. [14]

    NVIDIA TensorRT,

    NVIDIA, “NVIDIA TensorRT,” https://developer.nvidia.com/tensorrt

  15. [15]

    TensorRT-LLM,

    NVIDIA Corporation, “TensorRT-LLM,” https://github.com/NVIDIA/ TensorRT-LLM

  16. [16]

    Efficient memory management for large language model serving with PagedAttention,

    W. Kwon, Z. Li, S. Zhuang, Y . Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica, “Efficient memory management for large language model serving with PagedAttention,” inProceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023

    2023

  17. [17]

    SGLang: Efficient execution of structured language model programs,

    L. Zheng, L. Yin, Z. Xie, C. Sun, J. Huang, C. H. Yu, S. Cao, C. Kozyrakis, I. Stoica, J. E. Gonzalez, C. Barrett, and Y . Sheng, “SGLang: Efficient execution of structured language model programs,” inAdvances in Neural Information Processing Systems 37, 2024

    2024

  18. [18]

    ONNX Runtime,

    Microsoft, “ONNX Runtime,” https://onnxruntime.ai

  19. [19]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskeveret al., “Language models are unsupervised multitask learners,”OpenAI blog, 2019

    2019

  20. [20]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution-augmented transformer for speech recognition,”arXiv preprint arXiv:2005.08100, 2020

    Pith/arXiv arXiv 2005

  21. [21]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inProc. of ICLR, 2023

    2023

  22. [22]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4195–4205

    2023

  23. [23]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,”arXiv preprint arXiv:2207.12598, 2022

    Pith/arXiv arXiv 2022

  24. [24]

    BigVGAN: A universal neural vocoder with large-scale training,

    S.-g. Lee, W. Ping, B. Ginsburg, B. Catanzaro, and S. Yoon, “BigVGAN: A universal neural vocoder with large-scale training,”arXiv preprint arXiv:2206.04658, 2022

    Pith/arXiv arXiv 2022

  25. [25]

    W2v-BERT: Combining contrastive learning and masked lan- guage modeling for self-supervised speech pre-training,

    Y .-A. Chung, Y . Zhang, W. Han, C.-C. Chiu, J. Qin, R. Pang, and Y . Wu, “W2v-BERT: Combining contrastive learning and masked lan- guage modeling for self-supervised speech pre-training,” in2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2021, pp. 244–250

    2021

  26. [26]

    Neural discrete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,” inProc. NeurIPS, vol. 30, 2017

    2017

  27. [27]

    CAM++: A fast and efficient network for speaker verification using context-aware masking,

    H. Wang, S. Zheng, Y . Chen, L. Cheng, and Q. Chen, “CAM++: A fast and efficient network for speaker verification using context-aware masking,” inProc. of INTERSPEECH, 2023, pp. 5301–5305

    2023

  28. [28]

    Seed-TTS: A family of high-quality versatile speech generation models,

    P. Anastassiou, J. Chen, J. Chen, Y . Chen, Z. Chen, Z. Chen, J. Cong, L. Deng, C. Ding, L. Gaoet al., “Seed-TTS: A family of high-quality versatile speech generation models,”arXiv preprint arXiv:2406.02430, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Henretty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” inProceedings of the Twelfth Language Resources and Evaluation Conference, 2020, pp. 4218–4222

    2020

  30. [30]

    DiDiSpeech: A large scale mandarin speech corpus,

    T. Guo, C. Wen, D. Jiang, N. Luo, R. Zhang, S. Zhao, W. Li, C. Gong, W. Zou, K. Hanet al., “DiDiSpeech: A large scale mandarin speech corpus,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6968–6972

    2021

  31. [31]

    Robust speech recognition via large-scale weak super- vision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak super- vision,” inProc. of ICML, vol. 202, 2023, pp. 28 492–28 518

    2023

  32. [32]

    FunASR: A fundamental end-to-end speech recognition toolkit,

    Z. Gao, Z. Li, J. Wang, H. Luo, X. Shi, M. Chen, Y . Li, L. Zuo, Z. Du, and S. Zhang, “FunASR: A fundamental end-to-end speech recognition toolkit,” inProc. of INTERSPEECH, 2023, pp. 1593–1597

    2023

  33. [33]

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

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA-TDNN: emphasized channel attention, propagation and aggregation in TDNN based speaker verification,” inProc. of INTERSPEECH, 2020, pp. 3830– 3834

    2020

  34. [34]

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

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet 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

    2022

  35. [35]

    UTMOS: UTokyo-SaruLab system for V oiceMOS chal- lenge 2022,

    T. Saeki, D. Xin, W. Nakata, T. Koriyama, S. Takamichi, and H. Saruwatari, “UTMOS: UTokyo-SaruLab system for V oiceMOS chal- lenge 2022,” inProc. of INTERSPEECH, 2022

    2022