Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Finite scalar quantization gives neural audio codecs a built-in redundancy that keeps speech intelligible when up to 10% of transmitted bits flip, where conventional residual vector quantization codecs fall apart after 1%.

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 ↗

FSQ-based audio codecs outperform RVQ-based codecs in simulated bit-error transmission, preserving intelligibility at bit-flip rates up to 10%.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Interesting FSQ redundancy result, but the bit-flip comparison is confounded by codebook count and bitrate; the distillation experiment is the cleaner result. the 4 major comments →

arxiv 2509.09550 v2 pith:IS5RL3UQ submitted 2025-09-11 cs.SD cs.LG

Finite Scalar Quantization Enables Redundant and Transmission-Robust Neural Audio Compression at Low Bit-rates

Harry Julian , Rachel Beeson , Lohith Konathala , Johanna Ulin , Jiameng Gao This is my paper
classification cs.SD cs.LG
keywords finite scalar quantizationresidual vector quantizationneural audio codecbit-flip robustnessnoisy channel transmissionaudio compressionspeech intelligibilityencoder distillation
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.

The reading

Neural audio codecs compress speech into discrete code sequences, but most use residual vector quantization (RVQ), which maps each code index to an arbitrary learned vector so a single bit error can be catastrophic. This paper introduces NeuCodec, a codec that uses finite scalar quantization (FSQ), and argues that FSQ's fixed grid of evenly spaced levels gives code sequences an intrinsic locality: flipping a bit moves the index to a neighboring, acoustically similar point. Two experiments support this. First, two very different encoders trained against the same frozen FSQ quantizer and decoder produce code sequences that match only 2% element-wise, yet reconstruct with similar quality, suggesting the encoding is redundant. Second, when code bits are flipped at rates from 0.1% to 50%, FSQ codecs hold their intelligibility through roughly 10% bit errors, while RVQ codecs degrade sharply past 1%.

Core claim

The paper's central discovery is that FSQ is not merely a simpler training alternative to RVQ; it changes the error structure of the compressed bitstream. Because FSQ enumerates a grid and maps neighboring scalar levels to neighboring integer indices, a bit flip causes a small, bounded shift in embedding space. The paper demonstrates that two encoders sharing an FSQ bottleneck and decoder can learn nearly orthogonal code sequences—2% element-wise agreement—while achieving comparable reconstruction, with 93% of code predictions correct or off by one level. In transmission simulation, FSQ codecs maintain stable STOI and reconstruction metrics under bit flips up to about 10%, whereas RVQ codecs

What carries the argument

The central mechanism is Finite Scalar Quantization (FSQ), which projects encoder outputs into a low-dimensional bounded box, quantizes each dimension independently to a fixed number of evenly spaced levels, and enumerates all combinations into a single codebook. This fixed-grid structure creates a codebook in which index distance correlates with geometric distance in the quantization space: neighboring codes point to neighboring cells. The paper's two experiments rely on that locality: the distillation experiment uses a frozen FSQ bottleneck and decoder to show code-sequence diversity, and the perturbation experiment uses the codebook's bit representation to show graceful degradation.

Load-bearing premise

The claim stands or falls on the assumption that the bit-flip comparison isolates the quantizer, despite FSQ and RVQ codecs differing in architecture, frame rate, codebook size, and total bitrate—and the paper does not specify the FSQ enumeration order or effective bitrates.

What would settle it

Repeat Section 5's bit-flip test at P_flip=0.01 with an FSQ and an RVQ codec matched on architecture, code rate, frame rate, codebook size, and total bitrate, and look at STOI and PESQ. If the RVQ codec does not show the sharp decline reported for conventional RVQ codecs, the claimed robustness advantage is an artifact of the comparison rather than a property of the quantizer.

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

If this is right

  • FSQ codecs can be transmitted over noisy channels with less error-correction overhead, because a 1–10% bit-error rate degrades speech gracefully instead of destroying it.
  • Downstream models trained on FSQ token sequences should be more tolerant of token substitution errors, since nearby code indices point to nearby acoustic content.
  • Encoder distillation with a frozen FSQ bottleneck and decoder can shrink the encoder by 15x with 6x speedup and no loss in quality, enabling encode-heavy applications.
  • The single-codebook FSQ design removes the need for auxiliary losses and hierarchical modeling, simplifying training and deployment of low-bitrate codecs.

Where Pith is reading between the lines

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

  • Editorial inference: the robustness likely depends on the enumeration order of FSQ's scalar levels; a poorly chosen index mapping could turn a single-bit flip into a jump across the quantization grid, so robustness should be treated as a design property, not an automatic consequence of FSQ.
  • Editorial inference: the paper's 'baked-in redundancy' could be exploited as a form of error-correcting code—neighboring-code information could let a receiver detect and correct errors without explicit channel coding.
  • Editorial inference: the 93% near-neighbor statistic suggests downstream token-based models may benefit from fuzzy token matching, treating off-by-one codes as soft targets rather than hard errors.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper introduces NeuCodec, an FSQ-based neural audio codec, and claims that FSQ quantizers produce code sequences with baked-in redundancy and transmission robustness. Two experiments are presented: (i) an encoder-distillation experiment in which two different encoders, sharing the same frozen FSQ quantizer and decoder, produce code sequences that disagree at 98% of positions yet yield comparable reconstruction quality; and (ii) a bit-flip transmission simulation on Librispeech test-clean comparing FSQ codecs (NeuCodec, distilled NeuCodec, StableCodec) with RVQ codecs (DAC, Encodec). The authors conclude that FSQ codecs are markedly more robust to channel bit errors than RVQ codecs, and they propose mechanistic explanations based on grid locality and codebook redundancy.

Significance. If the transmission-robustness claim were established, it would be a practically valuable property: FSQ is simpler to train than RVQ, uses a single codebook, and robustness to channel errors matters for low-bitrate speech transmission. The distillation experiment is a well-posed controlled setup (same quantizer and decoder) and the confusion-matrix analysis is informative. The bit-flip simulation is a reasonable first-order channel model. However, the central comparative claim is not currently supported: Section 5 confounds quantizer type with total bitrate, number of codebooks, frame rate, architecture, and training data, and the experimental reporting lacks confidence intervals and essential mapping details. A controlled comparison is needed before the paper's headline conclusion can be accepted.

major comments (4)
  1. [Section 5, Table 3, Fig. 3] The FSQ-vs-RVQ comparison does not control for the number of transmitted bits per frame or per second. NeuCodec transmits one 16-bit FSQ index per frame, while DAC uses 6 RVQ codebooks and Encodec uses 12. If the RVQ indices are 10-bit (codebook size 1024), Encodec transmits 120 bits/frame versus NeuCodec's 16 bits/frame. At the same per-bit flip probability P_flip, RVQ codecs therefore absorb many more expected corruptions per second (e.g., 7.5x for Encodec at equal frame rates). The observed 'sharp decline' of RVQ curves may be an artifact of larger absolute channel error load rather than of the quantizer type. The authors should report total bitrates and frame rates, and either equalize expected corrupted bits/sec across codecs, equalize total bitrate, or use a matched architecture with FSQ and RVQ bottlenecks.
  2. [Section 5, Eq. (1)] The bit-level experiment is not reproducible because the paper does not specify the FSQ enumeration order used to convert quantized tuples to integer indices, nor the per-dimension level counts n_i for NeuCodec (only the total codebook size 2^16 and projection dimension 8). The effect of flipping a bit in the binary representation of an integer index depends entirely on the chosen enumeration (e.g., mixed-radix ordering, row-major order, Gray coding). A 'neighboring code' in the quantizer grid (Section 4.2) is not necessarily a single-bit-flip neighbor under the unspecified integer-to-bit mapping. Please state the exact enumeration and the n_i values, and consider analyzing sensitivity to the enumeration.
  3. [Section 5, Fig. 3] No confidence intervals or significance tests are reported for the bit-flip simulation. Since bit flips are stochastic and the underlying test set is finite, a single run leaves the observed differences between codecs within sampling noise. Report means and error bars over multiple random seeds (and ideally over bootstrap resamples of utterances) for each P_flip, and also show the P_flip=0 baselines so that 'stable performance' is measured relative to each codec's own noiseless quality.
  4. [Section 5, Table 3] The conclusion attributes the robustness gap to 'FSQ versus RVQ', but the compared systems differ in many other respects: StableCodec is a 1-billion-parameter transformer, DAC and Encodec are convolutional; training data, model capacity, codebook size, number of codebooks, and operating bitrate all differ. Without controlling for these factors, the experiment cannot isolate the quantizer type. At minimum, the authors should add an ablation that keeps architecture, decoder, bitrate, and frame rate fixed and varies only FSQ vs RVQ, or explicitly discuss why the uncontrolled differences cannot explain the observed ordering.
minor comments (5)
  1. [Section 4.2] The text says code sequences are 'vastly different' (2% exact match) but also that 93% of level predictions are correct or within one level. These statements should be reconciled: the codes are mostly close in the implicit codebooks, not arbitrary. Defining 'redundancy' operationally would strengthen the claim.
  2. [Table 2] The numerical entries for WER, CER, STOI, PESQ appear to run together (e.g., '6352.3 0.90.90 2.06'). Please reformat the table so each column is legible and the model rows are unambiguous.
  3. [Section 5] 'using the maximum size of the codebook' should probably be 'using the codebook size'. Please specify the codebook sizes used for DAC and Encodec as well as for the FSQ models.
  4. [Section 3 / Table 3] Clarify which decoder (16 kHz or 24 kHz) is used in the bit-flip experiment and what frame rate each codec operates at. The current Table 3 lists only the audio sampling rate, not the code frame rate.
  5. [Section 7] Typo: 'futre' should be 'future'.

Circularity Check

0 steps flagged

No significant circularity: the paper's central robustness claim is an empirical comparison against external baselines, not a derivation forced by construction or by self-citation.

full rationale

The paper's two main claims are (1) that two encoders can learn different code sequences with comparable reconstruction quality, and (2) that FSQ-based codecs are more robust to bit flips than RVQ codecs. Both are empirical findings, not consequences of fitted parameters or self-citation. The distillation experiment (Sec. 4) trains a student encoder with a loss that includes MSE between pre-quantization representations, but the observed 2% code-level agreement and 0.73 cosine similarity are measured outcomes, not quantities fixed by the training objective. The bit-flip experiment (Sec. 5) evaluates NeuCodec, Distill-NeuCodec, StableCodec, DAC, and Encodec by transmitting their code sequences through a simulated binary symmetric channel and measuring STOI, PESQ, SI-SDR, and mel-MSE. The robustness conclusion could have failed: RVQ codecs could have remained robust, or FSQ codecs could have degraded sharply. No parameter was fitted to the bit-flip outcomes, and no 'prediction' is a renamed fit. The mechanism discussion in Sec. 6 is post hoc interpretation rather than a derivation. The paper does not rely on load-bearing self-citations; key references (FSQ, Encodec, DAC, StableCodec, XCodec2) are external prior work. The comparison does contain a potential confound—RVQ codecs transmit more bits per frame (e.g., 12 codebooks for Encodec vs. 1 for NeuCodec) so identical per-bit flip probabilities produce different expected numbers of corrupted codes—but this is a validity threat to the causal attribution, not circularity. A confounded experiment is still an experiment whose outcome is not guaranteed by the inputs.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard FSQ background, a simplified noise-channel model, fixed-width integer coding, and the assumption that observed robustness differences are not caused by uncontrolled codec differences. No new physical entities are introduced. The main free choices are the FSQ codebook configuration, the StableCodec quantizer levels, and the unreported distillation loss weights.

free parameters (3)
  • NeuCodec FSQ codebook configuration = 2^16 levels, projection dim 8
    The robustness and redundancy claims are demonstrated only for this codebook size and shape; results may not hold for smaller codebooks.
  • StableCodec quantizer levels = {8, 8, 8, 8, 4, 4}
    Authors set these levels post hoc to make a 2^16 power-of-two codebook so it fits the bit-flip experiment; this choice is not derived from the model.
  • Distillation loss weights = not reported
    Training uses a weighted sum of five losses, but the weights are not given, so the distillation experiment is under-specified.
axioms (5)
  • standard math Codebook size of FSQ is the product of per-dimension levels (Eq. 1).
    Background to FSQ; not derived in the paper.
  • domain assumption An independent bit-flip binary symmetric channel is a valid model for transmission noise.
    Section 5; the entire perturbation experiment relies on this model.
  • domain assumption Fixed-width integer coding of code indices is lossless and representative.
    Section 5; the bit-to-integer mapping is not fully specified.
  • domain assumption The frozen decoder and quantizer in the distillation experiment are reliable probes of code-space redundancy.
    Section 4; similar reconstruction from different codes is interpreted as redundancy in the encoding.
  • domain assumption Observed robustness differences are attributable to FSQ versus RVQ rather than architecture, data, or bitrate confounds.
    Section 5; these variables are not controlled across codecs.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Finite Scalar Quantization Enables Redundant and Transmission-Robust Neural Audio Compression at Low Bit-rates." pith.science (2026). https://pith.science/paper/IS5RL3UQ

@misc{pith2026250909550,
  author       = {Pith},
  title        = {Pith review of: Finite Scalar Quantization Enables Redundant and Transmission-Robust Neural Audio Compression at Low Bit-rates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IS5RL3UQ}},
  note         = {Machine review of arXiv:2509.09550}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Neural Audio Codecs (NACs) have become increasingly adopted in speech processing tasks due to their excellent rate-distortion performance and compatibility with Large Language Models (LLMs) as discrete feature representations for audio generation. While most existing codecs rely on Residual Vector Quantization (RVQ), Finite Scalar Quantization (FSQ) has recently emerged as a compelling alternative that simplifies training and natively supports single codebooks. We introduce NeuCodec, an FSQ-based NAC, and show that FSQ encodes baked-in redundancy which produces an encoding which is robust when transmitted through noisy channels. First, through an encoder distillation experiment, we show that two different encoders can learn to encode identical audio into vastly different code sequences whilst maintaining comparable reconstruction quality with the same quantizer and decoder. Second, we demonstrate that FSQ has vastly superior bit-level perturbation robustness by comparing the performance of RVQ and FSQ codecs when simulating the transmission of code sequences through a noisy channel.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. CrossAccent-TTS: Cross-Lingual Accent-Intensity Controllable Text-to-Speech via Disentangled Speaker and Accent Representations

    eess.AS 2026-06 unverdicted novelty 5.0

    CrossAccent-TTS adds an Accent Intensity Controller to disentangled representations for controllable accent strength in cross-lingual TTS on Indic and L2 datasets.

Reference graph

Works this paper leans on

39 extracted references · 6 linked inside Pith · cited by 1 Pith paper

  1. [1]

    INTRODUCTION Recently, Neural Audio Codecs (NACs) have gained widespread us- age in speech processing, due to their ability to compress speech into ultra-low bitrate discrete code sequences whilst maintaining high perceptual quality when reconstructing these sequences back into waveforms [1]. The autoencoding task used to train NACs embeds a compressed la...

  2. [2]

    BACKGROUND RVQ discretizes an embedding space through first performing Vector Quantization [10] over a finite codebook, after which discretization errors (e.g. the distance between the scalar vector and the nearest neighbor codeword embedding) are obtained and discretized again, a process that continues for a predetermined number of codebooks. This means ...

    Pith/arXiv arXiv 2025

  3. [3]

    The encoder takes raw waveforms as an input and consists of a pre- trained frozen semantic encoder and a trainable acoustic encoder

    NEUCODEC NeuCodec is primarily based on XCodec2 [5], an ultra-low bitrate audio codec designed for downstream modeling in LLM-based TTS. The encoder takes raw waveforms as an input and consists of a pre- trained frozen semantic encoder and a trainable acoustic encoder. The semantic encoder is Wav2Vec2-BERT-large [13] which was pre- trained on 4.5 million ...

  4. [4]

    ENCODER DISTILLA TION XCodec2 was originally designed as a feature representation for TTS. Its asymmetric configuration of an encoder that largely out- sizes the decoder in parameter size and compute complexity, enables a trade-off of enhanced compression performance and slow encod- ing speeds with fast decoding speeds at inference time. As a trained TTS ...

  5. [5]

    How catastrophic the perturbation of a single bit is to the received signal depends on the encoding of said signal

    BIT-LEVEL PERTURBA TION EXPERIMENT When signals are transmitted through a medium, the data that is sent may be different from what is received due to signal interference or noise. How catastrophic the perturbation of a single bit is to the received signal depends on the encoding of said signal. The code indices of our quantized encoder outputs can be view...

  6. [6]

    Encoder/decoder parameter breakdown and performance comparison on CMU-Arctic subset

    DISCUSSION As shown in our distillation experiment, when encoder outputs and code sequences change, the reconstruction quality can remain the Table 2. Encoder/decoder parameter breakdown and performance comparison on CMU-Arctic subset. Encoder Decoder Acoustic (M) Semantic (M) Total (M) WER (%) CER (%) STOI PESQ encRTF NeuCodec 16kHz 35 600 6352.3 0.90.90...

  7. [7]

    CONCLUSION In conclusion, we found that FSQ biases NACs to learn discrete au- dio encodings that have in-built redundancy and a code-level pertur- bation robustness that could be advantageous for designing futre low bit-rate neural compressors that are resilient to noise in transmis- sion. Future work should assess (1) the usefulness of this property in l...

  8. [8]

    Recent advances in discrete speech tokens: A review,

    Yiwei Guo et al., “Recent advances in discrete speech tokens: A review,”arXiv preprint arXiv:2502.06490, 2025

    arXiv 2025

  9. [9]

    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

  10. [10]

    Codec-asr: Training performant auto- matic speech recognition systems with discrete speech repre- sentations,

    Kunal Dhawan et al, “Codec-asr: Training performant auto- matic speech recognition systems with discrete speech repre- sentations,” inInterspeech 2024, 2024

    2024

  11. [11]

    Moshi: a speech-text foun- dation model for real-time dialogue,

    Alexandre D ´efossez et al., “Moshi: a speech-text foun- dation model for real-time dialogue,”arXiv preprint arXiv:2410.00037, 2024. Fig. 3. Perturbation Robustness Across conditions for all NACs

    Pith/arXiv arXiv 2024

  12. [12]

    Llasa: Scaling train-time and inference-time compute for llama-based speech synthesis,

    Zhen Ye et al., “Llasa: Scaling train-time and inference-time compute for llama-based speech synthesis,”arXiv preprint arXiv:2502.04128, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Soundstream: An end-to- end neural audio codec,

    Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi, “Soundstream: An end-to- end neural audio codec,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 495–507, 2022

    2022

  14. [14]

    High-fidelity audio compression with improved rvqgan,

    Rithesh Kumar, Prem Seetharaman, Alejandro Luebs, Ishaan Kumar, and Kundan Kumar, “High-fidelity audio compression with improved rvqgan,” inAdvances in Neural Information Processing Systems, 2023, vol. 36

    2023

  15. [15]

    Neural codec language models are zero- shot text to speech synthesizers,

    Sanyuan Chen et al., “Neural codec language models are zero- shot text to speech synthesizers,”IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 705–718, 2025

    2025

  16. [16]

    Finite scalar quantization: VQ-V AE made simple,

    Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen, “Finite scalar quantization: VQ-V AE made simple,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  17. [17]

    Neural discrete representation learning,

    Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu, “Neural discrete representation learning,” inAdvances in Neu- ral Information Processing Systems, 2017, vol. 30

    2017

  18. [18]

    High fidelity neural audio compression,

    Alexandre D ´efossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi, “High fidelity neural audio compression,”Transactions on Machine Learning Research, 2023

    2023

  19. [19]

    Scaling transformers for low-bitrate high-quality speech coding,

    Julian D Parker, Anton Smirnov, Jordi Pons, CJ Carr, Zack Zukowski, Zach Evans, and Xubo Liu, “Scaling transformers for low-bitrate high-quality speech coding,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  20. [20]

    Seamless: Multilingual ex- pressive and streaming speech translation,

    Seamless Communication et al., “Seamless: Multilingual ex- pressive and streaming speech translation,”arXiv preprint arXiv:2313.05187, 2023

    arXiv 2023

  21. [21]

    Bigcodec: Pushing the limits of low-bitrate neu- ral speech codec,

    Detai Xin, Xu Tan, Shinnosuke Takamichi, and Hiroshi Saruwatari, “Bigcodec: Pushing the limits of low-bitrate neu- ral speech codec,”arXiv preprint arXiv:2409.05377, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Neural net- works fail to learn periodic functions and how to fix it,

    Liu Ziyin, Tilman Hartwig, and Masahito Ueda, “Neural net- works fail to learn periodic functions and how to fix it,” in Advances in Neural Information Processing Systems, 2020, vol. 33

    2020

  23. [23]

    V ocos: Closing the gap between time- domain and fourier-based neural vocoders for high-quality au- dio synthesis,

    Hubert Siuzdak, “V ocos: Closing the gap between time- domain and fourier-based neural vocoders for high-quality au- dio synthesis,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  24. [24]

    Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation,

    Haorui He et al., “Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation,” in 2024 IEEE Spoken Language Technology Workshop (SLT), 2024, pp. 885–890

    2024

  25. [25]

    Mls: A large-scale multilingual dataset for speech research,

    Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Syn- naeve, and Ronan Collobert, “Mls: A large-scale multilingual dataset for speech research,” inInterspeech 2020, 2020

    2020

  26. [26]

    Libritts-r: A restored multi-speaker text-to-speech corpus,

    Yuma Koizumi et al., “Libritts-r: A restored multi-speaker text-to-speech corpus,” inInterspeech 2023, 2023, pp. 5496– 5500

    2023

  27. [27]

    FLEURS-R: A Restored Multilingual Speech Corpus for Generation Tasks,

    Min Ma et al., “FLEURS-R: A Restored Multilingual Speech Corpus for Generation Tasks,” inInterspeech 2024, 2024, pp. 1835–1839

    2024

  28. [28]

    Common voice: A massively- multilingual speech corpus,

    Rosana et al. Ardila, “Common voice: A massively- multilingual speech corpus,” inProceedings of the Twelfth Language Resources and Evaluation Conference, Marseille, France, May 2020, pp. 4218–4222, European Language Re- sources Association

    2020

  29. [29]

    Hui-audio- corpus-german: A high quality tts dataset,

    Pascal Puchtler, Johannes Wirth, and Ren ´e Peinl, “Hui-audio- corpus-german: A high quality tts dataset,” inKI 2021: Ad- vances in Artificial Intelligence, Cham, 2021, pp. 204–216, Springer International Publishing

    2021

  30. [30]

    L3ac: Towards a lightweight and lossless audio codec,

    Linwei Zhai, Han Ding, Cui Zhao, fei wang, Ge Wang, Wang Zhi, and Wei Xi, “L3ac: Towards a lightweight and lossless audio codec,”arXiv preprint arXiv:2504.04949, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Distilhu- bert: Speech representation learning by layer-wise distillation of hidden-unit bert,

    Heng-Jui Chang, Shu-wen Yang, and Hung-yi Lee, “Distilhu- bert: Speech representation learning by layer-wise distillation of hidden-unit bert,” inICASSP 2022 - 2022 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 7087–7091

    2022

  32. [32]

    TS3-Codec: Transformer-Based Simple Streaming Single Codec,

    Haibin Wu, Naoyuki Kanda, Sefik Emre Eskimez, and Jinyu Li, “TS3-Codec: Transformer-Based Simple Streaming Single Codec,” inInterspeech 2025, 2025, pp. 604–608

    2025

  33. [33]

    Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,

    Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae, “Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,” inAdvances in Neural Information Process- ing Systems, 2020, vol. 33, pp. 17022–17033

    2020

  34. [34]

    CMU ARCTIC databases for speech synthesis,

    John Kominek and Alan W. Black, “CMU ARCTIC databases for speech synthesis,” Technical Report CMU-LTI-03-177, Language Technologies Institute, Carnegie Mellon University, 2003

    2003

  35. [35]

    Robust speech recog- nition via large-scale weak supervision,

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine Mcleavey, and Ilya Sutskever, “Robust speech recog- nition via large-scale weak supervision,” 2023, vol. 202 ofPro- ceedings of Machine Learning Research, pp. 28492–28518

    2023

  36. [36]

    A short-time objective intelligibility measure for time-frequency weighted noisy speech,

    Cees H. Taal, Richard C. Hendriks, Richard Heusdens, and Jes- per Jensen, “A short-time objective intelligibility measure for time-frequency weighted noisy speech,” inProceedings of the IEEE International Conference on Acoustics, Speech, and Sig- nal Processing (ICASSP), 2010, pp. 4214–4217

    2010

  37. [37]

    Perceptual evaluation of speech quality (PESQ)—a new method for speech quality assessment of tele- phone networks and codecs,

    Antony W. Rix, John G. Beerends, Martin P. Hollier, and An- dries P. Hekstra, “Perceptual evaluation of speech quality (PESQ)—a new method for speech quality assessment of tele- phone networks and codecs,”Proceedings of the IEEE Interna- tional Conference on Acoustics, Speech, and Signal Processing (ICASSP), vol. 2, pp. 749–752, 2001

    2001

  38. [38]

    Librispeech: An asr corpus based on public do- main audio books,

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur, “Librispeech: An asr corpus based on public do- main audio books,” in2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 5206–5210

    2015

  39. [39]

    Sdr – half-baked or well done?,

    Jonathan Le Roux, Scott Wisdom, Hakan Erdogan, and John R. Hershey, “Sdr – half-baked or well done?,” inICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019, pp. 626–630

    2019

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