Multiple-target deep learning for LSTM-RNN based speech enhancement
The DOI matches its Crossref record; 27 of the 27 references it cites were found in a registry or an index.
Abstract
In this study, we explore long short-term memory recurrent neural networks (LSTM-RNNs) for speech enhancement. First, a regression LSTM-RNN approach for a direct mapping from the noisy to clean speech features is presented and verified to be more effective than deep neural network (DNN) based regression techniques in modeling long-term acoustic context. Then, a comprehensive comparison between the proposed direct mapping based LSTM-RNN and ideal ratio mask (IRM) based LSTM-RNNs is conducted. We observe that the direct mapping framework achieves better speech intelligibility at low signal-to-noise ratios (SNRs) while the IRM approach shows its superiority at high SNRs. Accordingly, to fully utilize this complementarity, a novel multiple-target joint learning approach is designed. The experiments under unseen noises show that the proposed framework can consistently and significantly improve the objective measures for both speech quality and intelligibility.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Hochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318 · cited 2×
“Consequently, long short-term memory recurrent neural network (LSTM-RNN) [11] introduces the concepts of memory cell and a series of gates to dynamically control the information flow, which well solves the vanishing gradient problem.”
BackgroundMethodA fast learning algorithm for deep belief nets
Hinton, G. E., Osindero, S., & Teh, Y.-W. (2006) · Neural Computation · doi:10.1162/neco.2006.18.7.1527 · Cited by 16,336 · cited 1×
“Recently, with the emergence of deep learning techniques [5], deep neural networks (DNNs) also demonstrate the powerful modeling capability and achieve better performances than the traditional methods in speech enhancement [6, 7].”
BackgroundGreff, K., Srivastava, R. K., Koutník, J., Steunebrink, B. R., & Schmidhuber, J. (2015) · arXiv · doi:10.48550/arxiv.1503.04069 · Cited by 7,073
Greff, K., Srivastava, R. K., Koutnik, J., Steunebrink, B. R., & Schmidhuber, J. (2017) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2016.2582924 · Cited by 6,955 · cited 1×
“As for the non-linear activation functions, the logistic sigmoid function is used in each gate and g is a hyperbolic tangent [21].”
MethodOn the difficulty of training recurrent neural networks
Pascanu, R., Mikolov, T., & Bengio, Y. (2012) · arXiv · doi:10.48550/arxiv.1211.5063 · Cited by 4,310 · cited 3×
“To alleviate the problem of vanishing and exploding gradients in conventional RNN [10], a design of the LSTM block is proposed to control the information flow in [11].”
BackgroundMethodSpeech enhancement using a minimum mean-square error log-spectral amplitude estimator
Ephraim, Y., & Malah, D. (1985) · IEEE Transactions on Acoustics, Speech, and Signal Processing · doi:10.1109/tassp.1985.1164550 · Cited by 3,723 · cited 1×
“In conventional algorithms, such as spectral subtraction [1] and MMSE-based spectral amplitude estimator [2, 3], the settings are unsupervised based on mathematical assumptions about speech and noise, and often introduces artifacts (e.g., musical noise) and performance limitations in enhanced…”
BackgroundRix, A. W., Beerends, J. G., Hollier, M. P., & Hekstra, A. P. (2002) · 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221) · doi:10.1109/icassp.2001.941023 · Cited by 3,149 · cited 1×
“The quality and intelligibility of the enhanced speech were measured by perceptual evaluation of speech quality (PESQ) [27], and short-time objective intelligibility (STOI) [28], respectively.”
MethodLong short-term memory recurrent neural network architectures for large scale acoustic modeling
Sak, H., Senior, A., & Beaufays, F. (2014) · Interspeech 2014 · doi:10.21437/interspeech.2014-80 · Cited by 2,973 · cited 1×
“Although the temporal information can be incorporated into DNN training via frame expansion [7, 8], there are still limitations in modeling long-term acoustic contexts because the relationship between the neighbouring frames is not explicitly modeled.”
BackgroundSpeech enhancement using a minimum-mean square error short-time spectral amplitude estimator
Ephraim, Y., & Malah, D. (1984) · IEEE Transactions on Acoustics, Speech, and Signal Processing · doi:10.1109/tassp.1984.1164453 · Cited by 2,863 · cited 1×
“In conventional algorithms, such as spectral subtraction [1] and MMSE-based spectral amplitude estimator [2, 3], the settings are unsupervised based on mathematical assumptions about speech and noise, and often introduces artifacts (e.g., musical noise) and performance limitations in enhanced…”
BackgroundAn Algorithm for Intelligibility Prediction of Time–Frequency Weighted Noisy Speech
Taal, C. H., Hendriks, R. C., Heusdens, R., & Jensen, J. (2011) · IEEE Transactions on Audio, Speech, and Language Processing · doi:10.1109/tasl.2011.2114881 · Cited by 2,254 · cited 1×
“The quality and intelligibility of the enhanced speech were measured by perceptual evaluation of speech quality (PESQ) [27], and short-time objective intelligibility (STOI) [28], respectively.”
MethodVarga, A., & Steeneken, H. J. M. (1993) · Speech Communication · doi:10.1016/0167-6393(93)90095-3 · Cited by 2,105 · cited 1×
“To evaluate on un-seen noise types, three other noise types, namely Buccaneer1, Destroyer engine and HF channel from the NOISEX-92 corpus [26], were adopted.”
MethodOn Training Targets for Supervised Speech Separation
Wang, Y., Narayanan, A., & Wang, D. (2014) · IEEE/ACM Transactions on Audio, Speech, and Language Processing · doi:10.1109/taslp.2014.2352935 · Cited by 1,067 · cited 1×
“Then compared with the previous LSTM-RNN approaches [13, 14] to directly or indirectly learn the ideal ratio mask (IRM) [15, 16, 17] of the time-frequency (T-F) bins as the targets, we observe that the LSTM-RNN speech enhancement models with different learning targets could exhibit complementary…”
ResultAn Experimental Study on Speech Enhancement Based on Deep Neural Networks
Xu, Y., Du, J., Dai, L.-R., & Lee, C.-H. (2014) · IEEE Signal Processing Letters · doi:10.1109/lsp.2013.2291240 · Cited by 924 · cited 3×
“Although the temporal information can be incorporated into DNN training via frame expansion [7, 8], there are still limitations in modeling long-term acoustic contexts because the relationship between the neighbouring frames is not explicitly modeled.”
BackgroundSelf-citationSpeech Enhancement: Theory and Practice
Loizou, P. C. (2013) · CRC Press eBooks · doi:10.1201/b14529 · Cited by 759
Phase-sensitive and recognition-boosted speech separation using deep recurrent neural networks
Erdogan, H., Hershey, J. R., Watanabe, S., & Le Roux, J. (2015) · 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp.2015.7178061 · Cited by 646 · cited 1×
“Then compared with the previous LSTM-RNN approaches [13, 14] to directly or indirectly learn the ideal ratio mask (IRM) [15, 16, 17] of the time-frequency (T-F) bins as the targets, we observe that the LSTM-RNN speech enhancement models with different learning targets could exhibit complementary…”
ResultSpeech Enhancement with LSTM Recurrent Neural Networks and its Application to Noise-Robust ASR
Weninger, F., Erdogan, H., Watanabe, S., Vincent, E., Le Roux, J., Hershey, J. R., & Schuller, B. (2015) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-22482-4_11 · Cited by 602 · cited 1×
“Finally, an indirect mapping approach similar to [13, 24], denoted as LSTM-IM , learns the IRM target via MMSE be-tween the masked and reference clean LPS features: where ˆ z IRM ( t, f ) is the estimated IRM from LSTM-RNN with the logarithm operation and the noisy LPS features x LPS ( t, f ) to…”
MethodCaruana, R. (1998) · Learning to Learn · doi:10.1007/978-1-4615-5529-2_5 · Cited by 577
On Ideal Binary Mask As the Computational Goal of Auditory Scene Analysis
Wang, D. (2006) · Speech Separation by Humans and Machines · doi:10.1007/0-387-22794-6_12 · Cited by 556
Ideal ratio mask estimation using deep neural networks for robust speech recognition
Narayanan, A., & Wang, D. (2013) · 2013 IEEE International Conference on Acoustics, Speech and Signal Processing · doi:10.1109/icassp.2013.6639038 · Cited by 526 · cited 1×
“Then compared with the previous LSTM-RNN approaches [13, 14] to directly or indirectly learn the ideal ratio mask (IRM) [15, 16, 17] of the time-frequency (T-F) bins as the targets, we observe that the LSTM-RNN speech enhancement models with different learning targets could exhibit complementary…”
ResultTowards Scaling Up Classification-Based Speech Separation
Wang, Y., & Wang, D. (2013) · IEEE Transactions on Audio, Speech, and Language Processing · doi:10.1109/tasl.2013.2250961 · Cited by 430 · cited 1×
“Recently, with the emergence of deep learning techniques [5], deep neural networks (DNNs) also demonstrate the powerful modeling capability and achieve better performances than the traditional methods in speech enhancement [6, 7].”
BackgroundSupervised and Unsupervised Speech Enhancement Using Nonnegative Matrix Factorization
Mohammadiha, N., Smaragdis, P., & Leijon, A. (2013) · IEEE Transactions on Audio, Speech, and Language Processing · doi:10.1109/tasl.2013.2270369 · Cited by 389 · cited 1×
“Non-negative matrix factorization (NMF) [4] is one typical example where speech and noise bases are obtained separately from training speech and noise data.”
BackgroundAn Introduction to Computational Networks and the Computational Network Toolkit
Yu, D., Eversole, A., Seltzer, M., Yao, K., Kuchaiev, O., Zhang, Y., Seide, F. T. B., Huang, Z., Guenter, B. K., Wang, H., et al. (2014) · Cited by 365
Discriminatively trained recurrent neural networks for single-channel speech separation
Weninger, F., Hershey, J. R., Le Roux, J., & Schuller, B. (2014) · 2014 IEEE Global Conference on Signal and Information Processing (GlobalSIP) · doi:10.1109/globalsip.2014.7032183 · Cited by 283 · cited 3×
“Finally, an indirect mapping approach similar to [13, 24], denoted as LSTM-IM , learns the IRM target via MMSE be-tween the masked and reference clean LPS features: where ˆ z IRM ( t, f ) is the estimated IRM from LSTM-RNN with the logarithm operation and the noisy LPS features x LPS ( t, f ) to…”
MethodResultMulti-task learning in deep neural networks for improved phoneme recognition
Seltzer, M. L., & Droppo, J. (2013) · 2013 IEEE International Conference on Acoustics, Speech and Signal Processing · doi:10.1109/icassp.2013.6639012 · Cited by 222 · cited 2×
“This is similar to multi-objective learning demonstrated recently in [19, 20].”
BackgroundResultLearning Sequential Structure in Simple Recurrent Networks
Servan‐Schreiber, D., Cleeremans, A., & McClelland, J. L. (1988) · Dépôt institutionnel de l'Université libre de Bruxelles (Université Libre de Bruxelles) · Cited by 175
Single-channel speech separation with memory-enhanced recurrent neural networks
Weninger, F. J., Eyben, F., & Schuller, B. W. (2014) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp.2014.6854294 · Cited by 136 · cited 1×
“In comparison to the DNN approach, LSTM-RNN was proposed in [12, 13] to yield a superior performance of noise reduction at low signal-to-noise ratios (SNRs).”
ResultXu, Y., Du, J., Huang, Z., Dai, L.-R., & Lee, C. (2015) · Interspeech (USB) · doi:10.21437/interspeech.2015-358 · Cited by 116 · cited 3×
“This is similar to multi-objective learning demonstrated recently in [19, 20].”
BackgroundMethodResultSelf-citation
Cited by
Supervised Speech Separation Based on Deep Learning: An Overview
Wang, D., & Chen, J. (2018) · IEEE/ACM Transactions on Audio, Speech, and Language Processing · doi:10.1109/taslp.2018.2842159 · Cited by 1,524
Speech Recognition Using Deep Neural Networks: A Systematic Review
Nassif, A. B., Shahin, I., Attili, I., Azzeh, M., & Shaalan, K. (2019) · IEEE Access · doi:10.1109/access.2019.2896880 · Cited by 1,152
Ephrat, A., Mosseri, I., Lang, O., Dekel, T., Wilson, K., Hassidim, A., Freeman, W. T., & Rubinstein, M. (2018) · ACM Transactions on Graphics · doi:10.1145/3197517.3201357 · Cited by 620
An Overview of Deep-Learning-Based Audio-Visual Speech Enhancement and Separation
Michelsanti, D., Tan, Z.-H., Zhang, S.-X., Xu, Y., Yu, M., Yu, D., & Jensen, J. (2021) · IEEE/ACM Transactions on Audio, Speech, and Language Processing · doi:10.1109/taslp.2021.3066303 · Cited by 316
Fu, S.-W., Wang, T.-W., Tsao, Y., Lu, X., & Kawai, H. (2018) · IEEE/ACM Transactions on Audio, Speech, and Language Processing · doi:10.1109/taslp.2018.2821903 · Cited by 314
Metadata
- Authors
- Sun, L., Du, J., Dai, L.-R., & Lee, C.-H.
- Venue
- 2017 Hands-free Speech Communications and Microphone Arrays (HSCMA)
- Publisher
- IEEE
- Published
- 2017
- Pages
- 136-140
- OpenAlex
- W2605589342
- Cited by
- 207
- References
- 27
- FWCI
- 17.63
- Type
- conference paper
- Language
- en
- Open access
- closed
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.