arxiv: 2604.26296 · v1 · submitted 2026-04-29 · 📡 eess.AS

Recognition: unknown

SPG-Codec: Exploring the Role and Boundaries of Semantic Priors in Ultra-Low-Bitrate Neural Speech Coding

Kun Wei, Mingyu Zhao, Zhiyong Wu, Zijian Lin

Pith reviewed 2026-05-07 12:39 UTC · model grok-4.3

classification 📡 eess.AS

keywords neural speech codingsemantic priorsultra-low bitrateHuBERTWhisperword error rateintelligibilityspeech codec

0 comments

The pith

Semantic priors reduce word errors in ultra-low-bitrate speech coding but lose effect above 6 kbps

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

This paper tests the addition of frozen semantic priors from HuBERT and Whisper to neural speech codecs that struggle with intelligibility when bitrates drop to extremes. At 1.5 kbps these priors lower word error rates by up to 10 percent relative to baselines without them. The gains disappear once the bitrate exceeds 6 kbps, a limit the authors identify as a practical boundary. The study also shows that priors rich in acoustic detail preserve natural sound qualities while language-focused priors reduce errors in noisy conditions and help with new speakers, leading to a proposed method that tunes prior strength according to available bitrate.

Core claim

Integrating frozen semantic priors such as HuBERT and Whisper into neural speech coding reveals a Semantic Retirement phenomenon in which the priors reduce Word Error Rate by up to 10 percent relatively at 1.5 kbps, yet their benefits rapidly diminish beyond 6 kbps. Acoustic-rich priors better preserve prosodic and timbral details, whereas high-level linguistic priors suppress phonetic hallucinations in noisy environments by 26 percent and narrow the generalization gap for unseen speakers. A bitrate-aware regulation strategy dynamically adjusts prior strength to optimize the trade-off between semantic consistency and perceptual naturalness.

What carries the argument

The Semantic Retirement phenomenon, which identifies the bitrate threshold past which semantic priors cease to deliver measurable gains in intelligibility over standard acoustic modeling alone.

If this is right

Ultra-low-bitrate neural codecs can reach higher intelligibility by adding semantic priors when bitrate is tightly constrained.
Acoustic priors suit applications that prioritize natural prosody while linguistic priors suit noisy channels.
Bitrate-aware adjustment of prior strength provides a concrete way to balance clarity against naturalness without retraining the codec.
The same regulation approach can be applied to other generative audio tasks that operate near the acoustic-semantic transition point.

Where Pith is reading between the lines

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

The 6 kbps retirement threshold may mark a general transition point where high-level semantic information begins to conflict with low-level signal reconstruction in any generative audio system.
Hybrid systems that switch between acoustic and linguistic priors on the fly could extend the useful range of the technique beyond fixed-bitrate settings.
Lightweight versions of the regulation strategy could be tested in real-time mobile or IoT speech transmission to confirm practical gains.

Load-bearing premise

The observed retirement effect and differences between prior types are not limited to the specific test recordings, noise conditions, or baseline codecs used in the experiments.

What would settle it

Repeating the WER measurements on a fresh speech dataset with different speakers and noise types and finding either no 10 percent reduction at 1.5 kbps or continued gains past 6 kbps would falsify the claimed capacity boundary.

Figures

Figures reproduced from arXiv: 2604.26296 by Kun Wei, Mingyu Zhao, Zhiyong Wu, Zijian Lin.

**Figure 1.** Figure 1: Framework of our proposed SPG-Codec. A frozen Semantic Constraint Module (HuBERT/Whisper) guides the backbone, while the view at source ↗

**Figure 2.** Figure 2: The “Semantic Retirement” Phenomenon. Relative improvement (%) over the no-prior baseline of semantic priors (HuBERT/Whisper). Gains are substantial at 1.5 kbps but diminish above 6.0 kbps. Speaker Similarity Semantic Consistency PESQ Mel Accuracy 0.75 0.80 0.85 0.90 0.95 1.0 Prior Trade-off @ 3.0 kbps Baseline + HuBERT + Whisper view at source ↗

**Figure 3.** Figure 3: Trade-off Analysis at 3.0 kbps. Absolute scores on multiple metrics. HuBERT excels in acoustic fidelity (PESQ), while Whisper dominates in Semantic Consistency (lowest absolute WER). bits. High-Bitrate Regime (≥ 6.0 kbps). As the bitrate increases, the benefit of semantic priors diminishes rapidly. At 6.0 kbps, the relative improvement in PESQ drops to near zero (or slightly negative for Whisper), and the … view at source ↗

**Figure 6.** Figure 6: Impact of Regulation Weight α. A “Sweet Spot” is observed around α = 0.05 − 0.1. Excessive weight (α = 0.2) leads to an over-constrained system. E. Generalization to Unseen Speakers To assess whether our method merely overfits to the training distribution or learns generalizable features, we evaluated performance on the challenging LibriSpeech test-other set. As shown in Table I, the baseline model suffers… view at source ↗

**Figure 5.** Figure 5: Hallucination Rate Analysis. Absolute hallucination rates under a separate detection protocol. Whisper effectively suppresses phonetic hallucinations, reducing the absolute rate by 26% in clean conditions at 1.5 kbps. its representations are highly invariant to speaker and noise but rich in linguistic content. This makes Whisper the ideal choice for maximizing intelligibility in extreme compression scenar… view at source ↗

read the original abstract

Conventional neural speech codecs suffer from severe intelligibility degradation at ultra-low bitrates, where the bottleneck transitions from acoustic distortion to semantic loss. To address this issue, this paper conducts a systematic investigation into the role and fundamental limits of integrating frozen semantic priors -- specifically HuBERT and Whisper -- into neural speech coding. We introduce and quantitatively validate a novel Semantic Retirement phenomenon: while semantic constraints reduce the Word Error Rate (WER) by up to ~10% relatively at 1.5 kbps, their benefits rapidly diminish beyond 6 kbps, indicating a practical capacity boundary. We further uncover a clear trade-off between different prior types: acoustic-rich priors (HuBERT) better preserve prosodic and timbral details, whereas high-level linguistic priors (Whisper) effectively suppress phonetic hallucinations in noisy environments (reducing hallucination rates by 26 percent) and substantially narrow the generalization gap for unseen speakers. Building on these findings, we propose a bitrate-aware regulation strategy that dynamically adjusts prior strength to optimize the trade-off between semantic consistency and perceptual naturalness. Extensive experimental evaluations confirm that our approach achieves competitive intelligibility and noise robustness compared to existing baselines, offering a principled pathway toward ultra-low-bitrate generative speech coding.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The core contribution is an empirical map of when semantic priors stop paying off in ultra-low-bitrate speech coding, around 6 kbps, plus a practical adjustment method.

read the letter

The main thing to know about this paper is that it identifies a bitrate threshold around 6 kbps where semantic priors stop helping much in neural speech coding. Below that, especially at 1.5 kbps, using frozen HuBERT or Whisper features cuts word error rate by about 10 percent relative to baselines without them. Above 6 kbps the gains shrink fast, which they term Semantic Retirement. They also find that HuBERT keeps better prosody and timbre while Whisper does a better job stopping phonetic hallucinations in noise, cutting those rates by 26 percent, and it helps with unseen speakers. What the work does well is run head-to-head tests of the two prior types and then build a bitrate-aware regulation method on top of the observations. That gives a clear picture of when each prior is useful and how to blend them without overdoing it at higher rates. The experimental claims are grounded in comparisons to existing codecs, which is the right way to show the incremental value. The soft spots are mostly around generalization. The 6 kbps crossover and the specific trade-offs could be sensitive to the training data, the test conditions, or exactly how the semantic features are fused into the codec architecture. If those choices are narrow, the boundary might not hold up as a universal limit. The abstract does not show error bars or full ablation tables, so the stability of the reported improvements is hard to gauge from what is here. The free parameters in the regulation strategy also need checking to see if they require per-dataset tuning. This paper is aimed at researchers in neural audio compression and low-bitrate communication systems. Anyone working on generative speech models at constrained rates will find the empirical boundaries useful to know. It deserves a serious referee because the questions it asks are practical and the results are falsifiable with standard metrics like WER and hallucination counts. The authors engage honestly with the limits of semantic priors rather than claiming they solve everything. I recommend sending it to peer review.

Referee Report

2 major / 1 minor

Summary. The paper investigates integrating frozen semantic priors from HuBERT and Whisper into neural speech codecs at ultra-low bitrates. It introduces the Semantic Retirement phenomenon, claiming that semantic constraints yield up to ~10% relative WER reductions at 1.5 kbps but benefits diminish beyond 6 kbps. It reports trade-offs (HuBERT better for prosody/timbre, Whisper for reducing hallucinations by 26% and narrowing speaker generalization gaps) and proposes a bitrate-aware regulation strategy for dynamic prior strength adjustment, with experiments showing competitive intelligibility and noise robustness versus baselines.

Significance. If the empirical curves and trade-offs hold under broader validation, the work offers a useful empirical map of semantic prior limits in speech coding, highlighting a practical capacity boundary and a regulation mechanism that could inform ultra-low-bitrate codec design. The systematic comparison of prior types and the quantitative validation of retirement effects constitute a constructive contribution, particularly the focus on hallucination suppression and unseen-speaker generalization.

major comments (2)

[Abstract] Abstract: the central claims of ~10% relative WER reduction at 1.5 kbps, 26% hallucination-rate drop, and the 6 kbps retirement boundary are presented without error bars, statistical significance tests, or ablation tables; this makes it difficult to assess whether the observed crossover is robust or sensitive to the specific test sets, noise conditions, and baseline codecs used.
[Method] The bitrate-aware regulation strategy relies on free parameters for prior strength (as noted in the axiom ledger); the manuscript should clarify whether these are tuned per bitrate or learned, and include sensitivity analysis, because the strategy is load-bearing for the claim that it optimizes the semantic-perceptual trade-off.

minor comments (1)

[Introduction] The term 'Semantic Retirement' is introduced as a novel phenomenon; a brief comparison to related concepts in prior literature on semantic bottlenecks in codecs would improve context.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review of our manuscript. We address each major comment point by point below, providing clarifications based on the existing experimental evidence and indicating revisions where the presentation can be strengthened without altering the core findings.

read point-by-point responses

Referee: [Abstract] Abstract: the central claims of ~10% relative WER reduction at 1.5 kbps, 26% hallucination-rate drop, and the 6 kbps retirement boundary are presented without error bars, statistical significance tests, or ablation tables; this makes it difficult to assess whether the observed crossover is robust or sensitive to the specific test sets, noise conditions, and baseline codecs used.

Authors: We acknowledge that the abstract, due to its length constraints, presents the key quantitative results in summary form without error bars or full statistical details. The full manuscript reports these in Section 4, including standard deviations from repeated evaluations across multiple seeds, paired statistical tests confirming significance of the WER and hallucination improvements (p < 0.01), and ablation studies varying test sets, noise levels, and baselines. The 6 kbps retirement boundary is shown to be consistent in Figures 3–5. In the revised version, we have added a brief qualifier in the abstract directing readers to these supporting analyses and included an explicit reference to the robustness checks, while preserving the abstract's conciseness. revision: partial
Referee: [Method] The bitrate-aware regulation strategy relies on free parameters for prior strength (as noted in the axiom ledger); the manuscript should clarify whether these are tuned per bitrate or learned, and include sensitivity analysis, because the strategy is load-bearing for the claim that it optimizes the semantic-perceptual trade-off.

Authors: The free parameters for prior strength are tuned per bitrate on a held-out validation set to balance semantic consistency and perceptual quality, as indicated in the axiom ledger; they are not learned jointly with the codec. We have revised the Method section to state this tuning procedure explicitly and added a sensitivity analysis subsection. This analysis sweeps the prior-strength hyperparameter over a range of values at each bitrate and reports the resulting WER, PESQ, and hallucination metrics, confirming stable gains around the selected operating points without sensitivity to small perturbations. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical investigation with independent experimental results

full rationale

The paper is structured as an empirical study that integrates frozen semantic priors (HuBERT/Whisper) into a neural codec and reports measured outcomes such as relative WER reductions at specific bitrates, hallucination rate drops, and trade-offs between prior types. No equations, predictions, or first-principles derivations are presented that reduce reported gains or the 'Semantic Retirement' boundary to quantities defined by the same fitted parameters or self-citations. The bitrate-aware regulation strategy is proposed based on observed experimental curves rather than any self-referential construction. All central claims rest on baseline comparisons and test-set evaluations that remain externally falsifiable and do not collapse into the inputs by definition.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claims rest on the domain assumption that frozen HuBERT and Whisper embeddings supply independent semantic information usable by a codec, plus the empirical observation that benefits saturate; the regulation strategy introduces adjustable parameters whose values are not derived from first principles.

free parameters (1)

bitrate-aware prior strength parameters
The dynamic adjustment strategy requires parameters that control how strongly the semantic prior is applied at each bitrate; these are chosen or fitted to optimize the observed trade-off.

axioms (1)

domain assumption Frozen pre-trained semantic models (HuBERT, Whisper) provide useful, task-agnostic priors for speech coding without retraining
The entire investigation presupposes that these models' representations remain beneficial when inserted into a codec pipeline.

invented entities (1)

Semantic Retirement phenomenon no independent evidence
purpose: To name and quantify the point at which additional semantic constraints cease to improve intelligibility
New descriptive term introduced to organize the observed saturation behavior; no independent falsifiable prediction outside the reported experiments.

pith-pipeline@v0.9.0 · 5530 in / 1395 out tokens · 40339 ms · 2026-05-07T12:39:33.873044+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 13 canonical work pages · 2 internal anchors

[1]

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

Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae, “HiFi-GAN: Genera- tive adversarial networks for efficient and high fidelity speech synthesis,” inAdvances in Neural Information Processing Systems (NeurIPS), 2020, vol. 33, pp. 17022–17033

2020
[2]

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
[3]

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 (TMLR), 2023

2023
[4]

arXiv preprint arXiv:2305.02765 , year=

Dongchao Yang, Songxiang Liu, Rongjie Huang, Jinchuan Tian, Chao Weng, and Yuexian Zou, “HiFi-Codec: Group-residual vector quanti- zation for high fidelity audio codec,”arXiv preprint arXiv:2305.02765, 2023

work page arXiv 2023
[5]

Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers

Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu et al., “Neural codec language models are zero-shot text to speech synthesizers,”arXiv preprint arXiv:2301.02111, 2023

work page internal anchor Pith review arXiv 2023
[6]

AudioLM: A language modeling approach to audio generation,

Zal ´an Borsos, Rapha ¨el Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi et al., “AudioLM: A language modeling approach to audio generation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 2523–2533, 2023

2023
[7]

A survey on speech large language models for understanding,

Jing Peng, Yucheng Wang, Bohan Li, Yiwei Guo, Hankun Wang, Yangui Fang et al., “A survey on speech large language models for understanding,”arXiv preprint arXiv:2410.18908, 2025

work page arXiv 2025
[8]

V ALL-E 2: Neural codec language models are human parity zero-shot text to speech synthesizers

Sanyuan Chen, Shujie Liu, Long Zhou, Yanqing Liu, Xu Tan, Jinyu Li et al., “V ALL-E 2: Neural codec language models are human parity zero-shot text to speech synthesizers,”arXiv preprint arXiv:2406.05370, 2024

work page arXiv 2024
[9]

Moshi: a speech-text foundation model for real-time dialogue

Alexandre D ´efossez, Laurent Mazar ´e, Manu Orsini, Am ´elie Royer, Patrick P ´erez, Herv ´e J ´egou et al., “Moshi: A speech-text foundation model for real-time dialogue,”arXiv preprint arXiv:2410.00037, 2024

work page internal anchor Pith review arXiv 2024
[10]

BigCodec: Pushing the limits of low-bitrate neural speech codec.arXiv preprint arXiv:2409.05377,

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

work page arXiv 2024
[11]

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

Yushen Chen, Zhikang Niu, Ziyang Ma, Keqi Deng, Chunhui Wang, Jian Zhao et al., “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 (ACL), Vienna, Austria, 2025, pp. 6255–6271

2025
[12]

Codec does matter: Exploring the semantic shortcoming of codec for audio language model,

Zhen Ye, Peiwen Sun, Jiahe Lei, Hongzhan Lin, Xu Tan, Zheqi Dai et al., “Codec does matter: Exploring the semantic shortcoming of codec for audio language model,”arXiv preprint arXiv:2408.17175, 2024

work page arXiv 2024
[13]

SpeechTokenizer: Unified speech tokenizer for speech large language models,

Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and Xipeng Qiu, “SpeechTokenizer: Unified speech tokenizer for speech large language models,” inInternational Conference on Learning Representations (ICLR), 2024

2024
[14]

RepCodec: A speech representation codec for speech tokenization,

Zhichao Huang, Chutong Meng, and Tom Ko, “RepCodec: A speech representation codec for speech tokenization,” inProceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), Bangkok, Thailand, 2024, pp. 5777–5790

2024
[15]

SemantiCodec: An ultra low bitrate semantic audio codec for general sound,

Haohe Liu, Xuenan Xu, Yi Yuan, Mengyue Wu, Wenwu Wang, and Mark D. Plumbley, “SemantiCodec: An ultra low bitrate semantic audio codec for general sound,”IEEE Journal of Selected Topics in Signal Processing, vol. 18, no. 8, pp. 1448–1461, 2024

2024
[16]

SNAC: Multi-scale neural audio codec,

Hubert Siuzdak, Florian Gr ¨otschla, and Luca A. Lanzend ¨orfer, “SNAC: Multi-scale neural audio codec,” inAudio Imagination: NeurIPS 2024 Workshop AI-Driven Speech, Music, and Sound Generation, 2024, arXiv:2410.14411

work page arXiv 2024
[17]

Jong Wook Kim, Justin Salamon, Peter Li, and Juan Pablo Bello

Shengpeng Ji, Ziyue Jiang, Wen Wang, Yifu Chen, Minghui Fang, Jia- long Zuo et al., “WavTokenizer: An efficient acoustic discrete codec to- kenizer for audio language modeling,”arXiv preprint arXiv:2408.16532, 2024, Accepted to ICLR 2025

work page arXiv 2024
[18]

DualCodec: A low-frame-rate, semantically- enhanced neural audio codec for speech generation,

Jiaqi Li, Xiaolong Lin, Zhekai Li, Shixi Huang, Yuancheng Wang, Chaoren Wang et al., “DualCodec: A low-frame-rate, semantically- enhanced neural audio codec for speech generation,” inProceedings of Interspeech, 2025, arXiv:2505.13000

work page arXiv 2025
[19]

HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,

Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakho- tia, Ruslan Salakhutdinov, and Abdelrahman Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3451–3460, 2021

2021
[20]

Robust speech recognition via large- scale weak supervision,

Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever, “Robust speech recognition via large- scale weak supervision,” inInternational Conference on Machine Learning (ICML), 2023, pp. 28492–28518

2023
[21]

Neural discrete representation learning,

Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu, “Neural discrete representation learning,” inAdvances in Neural Information Processing Systems (NeurIPS), 2017, vol. 30, pp. 6306–6315

2017
[22]

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 (NeurIPS), 2023, vol. 36

2023
[23]

Fewer-token neural speech codec with time-invariant codes,

Yong Ren, Tao Wang, Jiangyan Yi, Le Xu, Jianhua Tao, Chu Yuan Zhang et al., “Fewer-token neural speech codec with time-invariant codes,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 12737–12741

2024
[24]

SACodec: Asymmetric quantization with semantic anchoring for low-bitrate high-fidelity neural speech codecs,

Zhongren Dong, Bin Wang, Jing Han, Haotian Guo, Xiaojun Mo, Yimin Cao et al., “SACodec: Asymmetric quantization with semantic anchoring for low-bitrate high-fidelity neural speech codecs,”arXiv preprint arXiv:2512.20944, 2025

work page arXiv 2025
[25]

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 et al., “CosyV oice 3: Towards in-the-wild speech generation via scaling-up and post-training,”arXiv preprint arXiv:2505.17589, 2025

work page arXiv 2025
[26]

E2 TTS: embarrassingly easy fully non-autoregressive zero-shot TTS

Sefik Emre Eskimez, Xiaofei Wang, Manthan Thakker, Canrun Li, Chung-Hsien Tsai, Zhen Xiao et al., “E2 TTS: Embarrassingly easy fully non-autoregressive zero-shot tts,”arXiv preprint arXiv:2406.18009, 2024

work page arXiv 2024
[27]

wav2vec 2.0: A framework for self-supervised learning of speech representations,

Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” inAdvances in Neural Information Processing Systems (NeurIPS), 2020, vol. 33, pp. 12449–12460

2020
[28]

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

Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen et 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
[29]

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,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 5206–5210

2015
[30]

Perceptual evaluation of speech quality (PESQ): A new method for speech quality assessment of telephone networks and codecs,

Antony W. Rix, John G. Beerends, Michael P. Hollier, and Andries P. Hekstra, “Perceptual evaluation of speech quality (PESQ): A new method for speech quality assessment of telephone networks and codecs,” inIEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2001, vol. 2, pp. 749–752

2001
[31]

An algorithm for intelligibility prediction of time-frequency weighted noisy speech,

Cees H. Taal, Richard C. Hendriks, Richard Heusdens, and Jesper Jensen, “An algorithm for intelligibility prediction of time-frequency weighted noisy speech,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 19, no. 7, pp. 2125–2136, 2011

2011