Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

BitTTS: Highly Compact Text-to-Speech Using 1.58-bit Quantization and Weight Indexing

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Quantizing a TTS model's acoustic model and vocoder to ternary values and storing five weights per byte shrinks it from 25.66 MB to 4.39 MB while naturalness stays comparable to 4-bit quantization.

desk verdict BitTTS has a real, arithmetically sound compression result and a useful vocoder-bottleneck finding, but the abstract's 'similar model size' quality claim is unsupported because the only unquantized small baseline is 2.9x larger. read the letter →

arxiv 2506.03515 v1 pith:EGAWJMBE submitted 2025-06-04 eess.AS cs.LGcs.SDeess.SP

Masaya Kawamura , Takuya Hasumi , Yuma Shirahata , Ryuichi Yamamoto This is my paper
classification eess.AScs.LGcs.SDeess.SP
keywords text-to-speechquantization-awaretraining1.58-bitquantizationternaryweightsweightindexingmodelcompressionon-deviceTTSvocoder
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

BitTTS tries to establish that extreme low-precision quantization is a workable way to compress full text-to-speech systems for on-device use. The paper trains a lightweight acoustic model and a HiFi-GAN vocoder with weights restricted to the ternary set {-1, 0, 1}, then stores five consecutive ternary weights as a single 8-bit index. It reports that this reduces the model from 25.66 MB to 4.39 MB, an 83 percent cut, while synthesized speech remains comparable in naturalness to 4-bit quantized models and clearly better than a 32-bit model made small by reducing channel widths. A sympathetic reading of the experiments is that quantization, rather than architecture shrinking, is the more promising route to very small TTS models.

What carries the argument

The load-bearing mechanism is weight indexing, a storage scheme for ternary weights. After quantization-aware training rounds scaled weights to {-1,0,1}, the flattened weight vector is split into blocks of length 5; because there are $3^{5}$ = 243 possible five-value patterns, each block maps to an integer between 0 and 242 that fits in one byte. Retrieval rebuilds the weight tensor by looking up the pattern for each index. This recovers near-ideal 1.58-bit-per-weight storage on hardware that natively handles 8-bit values, and works alongside the absmax scaling and straight-through gradient estimator used during training.

What would settle it

A replication with a listening test that pits the 4.39 MB 1.58-bit model against a full-precision 32-bit model of genuinely the same 4.39 MB size would settle whether quantization helps beyond channel narrowing; if the full-precision model ties or wins, the compression advantage would not survive an equal-size comparison. A second check is whether the 1.58-bit vocoder consistently beats the 4-bit vocoder in MOS, since the paper's reported gap depends on a single training run.

Watch

Extended reading notes

Core claim

The central claim is that 1.58-bit quantization-aware training transfers from language models to TTS: a model with almost all convolutional weights in {-1,0,1} can still generate natural speech. The paper reports a naturalness MOS of 3.09 for the fully quantized 4.39 MB model with both the acoustic model and vocoder quantized, 3.30 for the 4-bit model with only the acoustic model quantized, and 1.20 for the naive 32-bit small model, with ground truth at 4.24. It also finds that quantizing the acoustic model alone keeps a 70 percent size reduction with better quality, and that the final convolutional layer before the waveform output should be left unquantized. The authors present this as evidence that quantization is effective for building small TTS models without significant quality loss.

Load-bearing premise

The quality comparison rests on treating the 32-bit channel-reduced model as the non-quantized baseline of similar size, but at 12.78 MB it is about three times larger than the 4.39 MB proposed model, leaving the 'similar model size' claim unverified.

Editorial extensions

If this is right

  • A full TTS stack can fit in less than 5 MB, making on-device deployment in car navigation and conversational agents plausible for storage-constrained settings.
  • If quality is the priority, quantizing only the acoustic model is the recommended operating point: roughly 70 percent size reduction with a MOS of 3.18 to 3.30.
  • The vocoder is the quality-critical component: quantizing it costs more naturalness than quantizing the acoustic model, and its final convolutional layer should remain full precision.
  • Weight indexing gets storage close to the 1.58-bit ideal (64.0 kB versus 63.4 kB in the paper's example layer), so further compression must come from coding the index distribution or shrinking non-quantized parts.

Reading between the lines

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

  • Beyond the paper, the non-uniform index distribution (peaks at indices 0, 121, and 242) suggests that variable-length or learned coding over five-weight blocks could shave more bytes; the paper's own Huffman coding only reduces 4.39 MB to 4.35 MB.
  • A cleaner test of the 'quantization beats shrinking' claim would compare the 4.39 MB ternary model with a 32-bit model compressed to the same 4.39 MB by channel reduction; the paper's 12.78 MB control does not settle that comparison.
  • Because only 1D convolutional layers are quantized, extending QAT to embedding and attention-like layers in the TTS stack is the natural next experiment; if those also tolerate ternary weights, even smaller footprints are plausible.
  • The RTF results show compression and speed do not align: quantized models are slightly slower than the small 32-bit model, so latency-critical deployments would need orthogonal acceleration rather than relying on this compression alone.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. BitTTS proposes a highly compact text-to-speech model by applying quantization-aware training (QAT) to as low as 1.58-bit ternary weights and storing five ternary weights as a single int8 index ("weight indexing"). The method is applied to a JETS-based acoustic model and HiFi-GAN vocoder, with the vocoder's final convolutional layer excluded from quantization. The authors report a model size reduction from 25.66 MB to 4.39 MB (83%) and a MOS of about 3.09 for the fully compressed model, claiming that this outperforms a non-quantized baseline of similar model size. The paper also compares 4-bit and 1.58-bit quantized variants, measures RTF on an Apple M1 Pro CPU, and analyzes the distribution of weight indices.

Significance. The size arithmetic is exact and the weight-indexing idea is an elegant, lossless mapping that is practical on hardware with byte-aligned storage. The paper provides a clear algorithmic description of QAT and weight indexing, and the comparison of 4-bit versus 1.58-bit quantization is informative for the TTS community. However, the headline claim of outperforming a same-size non-quantized baseline is not supported by the current experiment, as the only non-quantized small baseline is 12.78 MB, nearly three times larger than the proposed 4.39 MB model. This is a load-bearing gap that affects the abstract and the introduction. The technical contribution is otherwise sound and potentially useful for on-device TTS, but the experimental evidence needs to be aligned with the claims.

major comments (3)
  1. [Section 3.1, Table 1, Abstract] The "baseline of similar model size without quantization" is not similar in size: the 32-bit small model occupies 12.78 MB, while the proposed 1.58-bit model with weight indexing on both acoustic model and vocoder is 4.39 MB, a factor of 2.9 difference. Consequently, the abstract's claim that BitTTS outperforms a baseline of similar model size without quantization is not demonstrated by Table 1. A non-quantized 32-bit model at approximately 4.39 MB (e.g., with channel widths near 19 rather than 32) is missing; the gap between MOS 1.20 and 3.09 only shows that quantization can beat a much larger, aggressively narrowed architecture. Please either add this missing control or revise the claim to specify the exact sizes of the models being compared.
  2. [Section 3.2] The vocoder's convolutional layer closest to the waveform output is excluded from quantization, and the paper says this is because preliminary experiments showed significant quality degradation. The claims "quantizing both the acoustic model and the vocoder" and "83% reduction" should be qualified by reporting the size and parameter fraction of the excluded layer, and ideally by also reporting the MOS of a fully quantized vocoder (even if poor) to quantify the severity of the degradation. The exclusion is disclosed, but the current reporting makes the scope of the headline result ambiguous.
  3. [Section 3.3.1] The MOS comparisons use 15 raters on 30 utterances, and the reported 95% confidence intervals for the 4-bit and 1.58-bit configurations overlap (e.g., 2.96±0.11 vs 3.09±0.11 for the both-quantized models). The statement that 1.58-bit models "achieve comparable or even higher MOS" is accordingly not statistically established. Please add a significance test (e.g., paired bootstrap or Wilcoxon) or explicitly state that no significant difference was found; increasing the evaluation size would also strengthen the result.
minor comments (4)
  1. [Algorithm 1] The handling of the final block when L is not a multiple of L* is unspecified. The loop over the set in line 6 appears to iterate in descending order, and the reconstruction rules are not given for padded or unpadded blocks. Please clarify how incomplete blocks are encoded and decoded.
  2. [Section 2.2, Eq. (1)] Equation (1) defines q=2^(b-1) and clipping to [-(q-1), q], but Section 3.2 for 4-bit sets the clipping range to [-8, 7]; while this is consistent with q=8, the general notation could be made clearer to avoid confusion with the 1.58-bit ternary case.
  3. [Footnote 8] The Huffman-coding result (4.35 MB) is presented only in a footnote; if this is part of the proposed compression pipeline, it should be integrated into the main evaluation, or the footnote should be removed to avoid distracting from the main results.
  4. [Section 3.3.3] The index frequency plot (Figure 2) is informative, but the skew toward indices 0, 121, and 242 is unsurprising given the ternary weight distribution. A quantitative measure of entropy or compression gain would make the analysis of weight pattern bias more rigorous.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity found: the size reduction is arithmetic, MOS/RTF are external measurements, and weight indexing is an exact lossless encoding.

full rationale

BitTTS's central claims are self-contained. The 83% size reduction is a direct arithmetic consequence of (25.66-4.39)/25.66 x 100 as stated in footnote 7, using the model sizes reported in Table 1. The weight-indexing scheme is an exact integer packing: because 3^5 = 243 <= 256, each block of five ternary weights maps bijectively to one int8 index, so reconstruction is lossless and the storage formula is definitional rather than circular. The quantization scale beta in Eq. (5) is computed from the model weights themselves, not fitted to the MOS target, and the clipping/rounding operations are fixed mathematical transforms. MOS values are subjective listening-test scores on held-out test utterances, and RTF is a measured runtime on an Apple M1 Pro; neither is derived from the paper's assumptions. The only self-citation, [9], appears in the introduction as an example of prior lightweight TTS and is not load-bearing for the proposed method. The weakness in the 'similar model size' comparison -- the unquantized small baseline is 12.78 MB rather than 4.39 MB -- is a comparison-fairness issue, not a circular-reasoning issue, because the reported quality numbers are measured rather than constructed from the model-size equations.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

BitTTS introduces no new particles, forces, or theoretical entities. The central claims rest on design hyperparameters (L*=5, b, p) and on assumptions inherited from BitNet-style QAT plus the author-defined baselines; the vocoder output layer exclusion is a disclosed post-hoc adjustment. None of these are fitted to the evaluation target in a circular sense, but they limit generalization.

free parameters (3)
  • weight-index block size L* = 5
    Hand-selected so that 3^5=243 ternary patterns fit in one 8-bit integer; this choice directly enables the advertised 83 percent size reduction.
  • quantization bit width b = 1.58 and 4
    These precision levels are chosen as experimental conditions to compare trade-offs; they are not optimized or derived.
  • input activation precision p = 8
    Set in Section 2.2 for input scaling; no experiment justifies this particular value.
assumptions (5)
  • domain assumption Straight-through gradient estimation works for ternary quantization in a TTS training loop.
    Taken from BitNet b1.58 and prior binary-network work; the paper relies on it but does not test alternatives.
  • domain assumption Quantizing only 1D convolutional layers captures most of the achievable compression.
    Section 3.2 states more than 90 percent of weights belong to 1D convolutions; if a different architecture spreads weights differently, the size results change.
  • ad hoc to paper The 32-bit channel-reduced model is a fair non-quantized baseline.
    Defined in Section 3.1 with channels set to 32; its 12.78 MB size is not matched to the 4.39 MB proposed model, so the abstract's 'similar model size' comparison is not actually controlled.
  • ad hoc to paper The vocoder output layer must be excluded from quantization.
    Section 3.2 excludes it after preliminary experiments; this post-hoc choice changes the quality and size results.
  • domain assumption MOS from 15 raters and 30 utterances is a reliable estimate of naturalness.
    A common small-scale speech evaluation, but confidence intervals overlap across several conditions, so fine-grained rankings are uncertain.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BitTTS: Highly Compact Text-to-Speech Using 1.58-bit Quantization and Weight Indexing." pith.science (2026). https://pith.science/paper/EGAWJMBE

@misc{pith2026250603515,
  author       = {Pith},
  title        = {Pith review of: BitTTS: Highly Compact Text-to-Speech Using 1.58-bit Quantization and Weight Indexing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EGAWJMBE}},
  note         = {Machine review of arXiv:2506.03515}
}
read the original abstract

This paper proposes a highly compact, lightweight text-to-speech (TTS) model for on-device applications. To reduce the model size, the proposed model introduces two techniques. First, we introduce quantization-aware training (QAT), which quantizes model parameters during training to as low as 1.58-bit. In this case, most of 32-bit model parameters are quantized to ternary values {-1, 0, 1}. Second, we propose a method named weight indexing. In this method, we save a group of 1.58-bit weights as a single int8 index. This allows for efficient storage of model parameters, even on hardware that treats values in units of 8-bit. Experimental results demonstrate that the proposed method achieved 83 % reduction in model size, while outperforming the baseline of similar model size without quantization in synthesis quality.

Figures

Figures reproduced from arXiv: 2506.03515 by the authors.

Figure 1
Figure 1. BitTTS quantizes the weight of TTS. The proposed method reduces the model size by storing indices instead of weights. natural language processing [19, 20] but also speech recogni￾tion [17, 26] due to its higher accuracy [14, 16] than the static quantization such as PTQ. In this paper, we applied a popular QAT method called fake QAT [16]. 2.2. Quantization formulation The weights of a neural network layer are denoted… view at source ↗
Figure 2
Figure 2. Frequency of indices when applying weight indexing to the 1.58-bit TTS model. 3.3.3. Analysis of weight patterns [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Evolution Strategy-Based Calibration for Low-Bit Quantization of Speech Models

    cs.SD 2026-03 conditional novelty 5.0 of 10

    Using CMA-ES to jointly optimize activation quantization scales keeps speech-model accuracy near full precision under full INT8 and INT4 quantization.

Reference graph

Works this paper leans on

43 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    With the advancement of these TTS models, they are increasingly being integrated into mobile applications, such as car navigation systems and conver- sational bots, among others

    Introduction Many high-quality neural text-to-speech (TTS) models have been extensively researched [1–5]. With the advancement of these TTS models, they are increasingly being integrated into mobile applications, such as car navigation systems and conver- sational bots, among others. In these applications, TTS models are typically required to generate spe...

    work page

  2. [2]

    BitTTS: Highly Compact Text-to-Speech Using 1.58-bit Quantization and Weight Indexing

    Method 2.1. Quantization-aware training QAT [16] is a technique to quantize models dynamically during training. This quantization approach is widely used in not only 2We use 5 here because 243 patterns can be efficiently represented by the 256 patterns supported by an 8-bit integer. In other words, we store1.58×5 = 7.9bit into an 8-bit integer. 3https://m...

    work page Pith review arXiv 2025

  3. [3]

    Experimental conditions We conducted experiments to evaluate the effectiveness of quantization in TTS and the proposed methods

    Experiments 3.1. Experimental conditions We conducted experiments to evaluate the effectiveness of quantization in TTS and the proposed methods. We used LibriTTS-R [29], a high-quality multi-speaker corpus for ex- periments. This corpus comprises 585 hours of speech data at a 24 kHz sampling rate from 2,456 English speakers. We divided the corpus into a t...

    work page

  4. [4]

    Additionally, we introduced a method called weight indexing to further re- duce model size

    Conclusions In this paper, we proposed a highly compact TTS model with 1.58-bit quantization for on-device applications. Additionally, we introduced a method called weight indexing to further re- duce model size. Since QAT has not been explored in the TTS field, we also evaluated the trade-off between model size and synthesized speech quality. Experimenta...

    work page

  5. [5]

    Statistical parametric speech synthesis using deep neural networks,

    H. Zen, A. Senior, and M. Schuster, “Statistical parametric speech synthesis using deep neural networks,” inProc. ICASSP, 2013, pp. 7962–7966

    work page 2013

  6. [6]

    A review of deep learning based speech synthesis,

    Y . Ning, S. He, Z. Wu, C. Xing, and L.-J. Zhang, “A review of deep learning based speech synthesis,”Appl. Sci., vol. 9, no. 19, 2019

    work page 2019

  7. [7]

    A survey on neural speech synthesis,

    X. Tan, T. Qin, F. Soong, and T.-Y . Liu, “A survey on neural speech synthesis,”arXiv preprint arXiv:2106.15561, 2021

    arXiv 2021

  8. [8]

    An overview of affective speech synthesis and conversion in the deep learning era,

    A. Triantafyllopoulos, B. W. Schuller, G. I ˙ymen, M. Sezgin, X. He, Z. Yang, P. Tzirakis, S. Liu, S. Mertes, E. Andr ´e, R. Fu, and J. Tao, “An overview of affective speech synthesis and conversion in the deep learning era,”Proc. of the IEEE, vol. 111, no. 10, pp. 1355–1381, 2023

    work page 2023

Show all 43 references
  1. [9]

    A review of deep learning techniques for speech processing,

    A. Mehrish, N. Majumder, R. Bharadwaj, R. Mihalcea, and S. Po- ria, “A review of deep learning techniques for speech processing,” Information Fusion, vol. 99, p. 101869, 2023

    2023

  2. [10]

    Lightspeech: Lightweight and fast text to speech with neural architecture search,

    R. Luo, X. Tan, R. Wang, T. Qin, J. Li, S. Zhao, E. Chen, and T.-Y . Liu, “Lightspeech: Lightweight and fast text to speech with neural architecture search,” inProc. ICASSP, 2021, pp. 5699–5703

    2021

  3. [11]

    NIX- TTS: Lightweight and end-to-end text-to-speech via module-wise distillation,

    R. Chevi, R. E. Prasojo, A. F. Aji, A. Tjandra, and S. Sakti, “NIX- TTS: Lightweight and end-to-end text-to-speech via module-wise distillation,” inProc. SLT, 2023, pp. 970–976

    2023

  4. [12]

    ConvNeXt-TTS and ConvNeXt-VC: ConvNeXt-based fast end-to-end sequence- to-sequence text-to-speech and voice conversion,

    T. Okamoto, Y . Ohtani, T. Toda, and H. Kawai, “ConvNeXt-TTS and ConvNeXt-VC: ConvNeXt-based fast end-to-end sequence- to-sequence text-to-speech and voice conversion,” inProc. ICASSP, 2024, pp. 12 456–12 460

    2024

  5. [13]

    Lightweight and high-fidelity end-to-end text-to-speech with multi-band generation and inverse short-time fourier transform,

    M. Kawamura, Y . Shirahata, R. Yamamoto, and K. Tachibana, “Lightweight and high-fidelity end-to-end text-to-speech with multi-band generation and inverse short-time fourier transform,” inProc. ICASSP, 2023, pp. 1–5

    2023

  6. [14]

    SpeedySpeech: Efficient neural speech synthesis,

    J. Vainer and O. Du ˇsek, “SpeedySpeech: Efficient neural speech synthesis,” inProc. Interspeech, 2020, pp. 3575–3579

    2020

  7. [15]

    ClariNet: Parallel wave genera- tion in end-to-end text-to-speech,

    W. Ping, K. Peng, and J. Chen, “ClariNet: Parallel wave genera- tion in end-to-end text-to-speech,” inProc. ICLR, 2019

    2019

  8. [16]

    Learning trans- ferable architectures for scalable image recognition,

    B. Zoph, V . Vasudevan, J. Shlens, and Q. V . Le, “Learning trans- ferable architectures for scalable image recognition,” inProc. CVPR, 2018

    2018

  9. [17]

    Neural architec- ture optimization,

    R. Luo, F. Tian, T. Qin, E. Chen, and T.-Y . Liu, “Neural architec- ture optimization,” inProc. NeurIPS, vol. 31, 2018

    2018

  10. [18]

    A survey of quantization methods for efficient neural network inference,

    A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, and K. Keutzer, “A survey of quantization methods for efficient neural network inference,”arXiv preprint arXiv:2103.13630, 2021

    2021 arXiv

  11. [19]

    A white paper on neural network quantization,

    M. Nagel, M. Fournarakis, R. A. Amjad, Y . Bondarenko, M. van Baalen, and T. Blankevoort, “A white paper on neural network quantization,”arXiv preprint arXiv:2106.08295, 2021

    2021 arXiv

  12. [20]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference,

    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” in Proc. CVPR, 2018, pp. 2704–2713

    2018

  13. [21]

    4-bit conformer with native quantization aware training for speech recognition,

    S. Ding, P. Meadowlark, Y . He, L. Lew, S. Agrawal, and O. Ry- bakov, “4-bit conformer with native quantization aware training for speech recognition,” inProc. Interspeech, 2022, pp. 1711– 1715

    2022

  14. [22]

    Sub-8-bit quantization aware training for 8-bit neural network accelerator with on-device speech recog- nition,

    K. Zhen, H. D. Nguyen, R. Chinta, N. Susanj, A. Mouchtaris, T. Afzal, and A. Rastrow, “Sub-8-bit quantization aware training for 8-bit neural network accelerator with on-device speech recog- nition,” inProc. Interspeech, 2022, pp. 3033–3037

    2022

  15. [23]

    Bitnet: Scaling 1-bit transformers for large language models,

    H. Wang, S. Ma, L. Dong, S. Huang, H. Wang, L. Ma, F. Yang, R. Wang, Y . Wu, and F. Wei, “Bitnet: Scaling 1-bit transformers for large language models,”arXiv preprint arXiv:2310.11453, 2023

    2023 arXiv

  16. [24]

    The era of 1-bit llms: All large lan- guage models are in 1.58 bits,

    S. Ma, H. Wang, L. Ma, L. Wang, W. Wang, S. Huang, L. Dong, R. Wang, J. Xue, and F. Wei, “The era of 1-bit llms: All large lan- guage models are in 1.58 bits,”arXiv preprint arXiv:2402.17764, 2024

    2024 arXiv

  17. [25]

    Xnor- net: Imagenet classification using binary convolutional neural net- works,

    M. Rastegari, V . Ordonez, J. Redmon, and A. Farhadi, “Xnor- net: Imagenet classification using binary convolutional neural net- works,” inProc. ECCV, 2016, pp. 525–542

    2016

  18. [26]

    Xnor-net++: Improved binary neural networks,

    A. Bulat and G. Tzimiropoulos, “Xnor-net++: Improved binary neural networks,” inProc. BMVC, 2019

    2019

  19. [27]

    Bi- nary neural networks: A survey,

    H. Qin, R. Gong, X. Liu, X. Bai, J. Song, and N. Sebe, “Bi- nary neural networks: A survey,”Pattern Recognition, vol. 105, p. 107281, 2020

    2020

  20. [28]

    Binarized neural networks: Training deep neural networks with weights and activations constrained to +1 or -1,

    M. Courbariaux, I. Hubara, D. Soudry, R. El-Yaniv, and Y . Ben- gio, “Binarized neural networks: Training deep neural networks with weights and activations constrained to +1 or -1,”arXiv preprint arXiv:1602.02830, 2016

    2016 arXiv

  21. [29]

    Basic binary convolution unit for binarized image restoration network,

    B. Xia, Y . Zhang, Y . Wang, Y . Tian, W. Yang, R. Timofte, and L. Van Gool, “Basic binary convolution unit for binarized image restoration network,” inProc. ICLR, 2023

    2023

  22. [30]

    2-bit conformer quantization for automatic speech recog- nition,

    O. Rybakov, P. Meadowlark, S. Ding, D. Qiu, J. Li, D. Rim, and Y . He, “2-bit conformer quantization for automatic speech recog- nition,” inProc. Interspeech, 2023, pp. 4908–4912

    2023

  23. [31]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv:1607.06450, 2016

    2016 arXiv

  24. [32]

    Estimating or propa- gating gradients through stochastic neurons for conditional com- putation,

    Y . Bengio, N. L ´eonard, and A. Courville, “Estimating or propa- gating gradients through stochastic neurons for conditional com- putation,”arXiv preprint arXiv:1308.3432, 2013

    2013 arXiv

  25. [33]

    LibriTTS-R: A re- stored multi-speaker text-to-speech corpus,

    Y . Koizumi, H. Zen, S. Karita, Y . Ding, K. Yatabe, N. Morioka, M. Bacchiani, Y . Zhang, W. Han, and A. Bapna, “LibriTTS-R: A re- stored multi-speaker text-to-speech corpus,” inProc. Interspeech, 2023, pp. 5496–5500

    2023

  26. [34]

    Montreal forced aligner: Trainable text-speech align- ment using kaldi

    M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Son- deregger, “Montreal forced aligner: Trainable text-speech align- ment using kaldi.” inProc. Interspeech, vol. 2017, 2017, pp. 498– 502

    2017

  27. [35]

    JETS: Jointly training FastSpeech2 and HiFi-GAN for end to end text to speech,

    D. Lim, S. Jung, and E. Kim, “JETS: Jointly training FastSpeech2 and HiFi-GAN for end to end text to speech,” inProc. Interspeech, 2022, pp. 21–25

    2022

  28. [36]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” inProc. NeurIPS, 2017, pp. 5998–6008

    2017

  29. [37]

    HiFi-GAN: Generative adversarial networks for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “HiFi-GAN: Generative adversarial networks for efficient and high fidelity speech synthesis,” inProc. NeurIPS, vol. 33, 2020, pp. 17 022–17 033

    2020

  30. [38]

    Mixture density networks,

    C. M. Bishop, “Mixture density networks,”Aston University, Birmingham UK, 1994

    1994

  31. [39]

    Phone-level prosody modelling with GMM- based MDN for diverse and controllable speech synthesis,

    C. Du and K. Yu, “Phone-level prosody modelling with GMM- based MDN for diverse and controllable speech synthesis,” IEEE/ACM Trans. on Audio, Speech, and Lang. Process., vol. 30, pp. 190–201, 2022

    2022

  32. [40]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” inProc. ICLR, 2019

    2019

  33. [41]

    An exponential learning rate schedule for deep learning,

    Z. Li and S. Arora, “An exponential learning rate schedule for deep learning,”arXiv preprint arXiv:1910.07454, 2019

    1910 arXiv

  34. [42]

    ESPnet-TTS: Uni- fied, reproducible, and integratable open source end-to-end text- to-speech toolkit,

    T. Hayashi, R. Yamamoto, K. Inoue, T. Yoshimura, S. Watanabe, T. Toda, K. Takeda, Y . Zhang, and X. Tan, “ESPnet-TTS: Uni- fied, reproducible, and integratable open source end-to-end text- to-speech toolkit,” inProc. ICASSP, 2020, pp. 7654–7658

    2020

  35. [43]

    A method for the construction of minimum- redundancy codes,

    D. A. Huffman, “A method for the construction of minimum- redundancy codes,”Proc. IRE, vol. 40, no. 9, pp. 1098–1101, 1952

    1952

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.