Sub-Spectrogram Segmentation for Environmental Sound Classification via Convolutional Recurrent Neural Network and Score Level Fusion

Shan Cao; Shugong Xu; Shunqing Zhang; Tianhao Qiao; Zhichao Zhang

REVIEW 3 major objections 6 minor 24 references

Sub-Spectrogram Segmentation for Environmental Sound Classification via Convolutional Recurrent Neural Network and Score Level Fusion

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper claims that splitting a sound spectrogram into frequency sub-bands, classifying each sub-band with a convolutional recurrent network, and fusing the per-band scores raises ESC-50 accuracy to 81.9%, a 9.1-point gain over a plain…

desk verdict Competent engineering paper on ESC-50 with a plausible 81.9% result, but the number comes from test-set search, so the headline needs an independent hold-out before it can be trusted. read the letter →

arxiv 1908.05863 v1 pith:3FNJPOF2 submitted 2019-08-16 cs.SD cs.LGeess.AS

Tianhao Qiao , Shunqing Zhang , Zhichao Zhang , Shan Cao , Shugong Xu This is my paper · ORCID

classification cs.SDcs.LGeess.AS

keywords environmentalsoundclassificationsub-spectrogramsegmentationconvolutionalrecurrentneuralnetworkscorelevelfusionmixupdataaugmentationESC-50log-melspectrogram

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

The paper tries to establish that environmental sound classification improves when the log-mel spectrogram is cut into frequency sub-bands instead of being processed as one full-band image. Each sub-band is fed into the same convolutional recurrent neural network, and the resulting class scores are combined by weighted summation. Using four sub-bands with cutoffs at 3, 6, and 10 kHz, and giving more weight to the lowest band, the method reports 81.9% accuracy on ESC-50, up from 72.8% for the CNN baseline. This matters because it points to frequency-localized structure, especially in the low-frequency range, as a usable source of discriminative information for environmental sounds.

What carries the argument

The central mechanism is sub-spectrogram segmentation: partitioning the full log-mel spectrogram into several frequency-truncated sub-spectrograms, each spanning a contiguous band such as 0-3, 3-6, 6-10, and 10-22.05 kHz. Each sub-spectrogram is processed independently by the same CRNN, with convolutional layers extracting local spectro-temporal patterns and bidirectional GRU layers modeling temporal dependencies, and the per-band class probability vectors are merged by a weighted sum (score level fusion). The band boundaries and fusion weights are treated as tunable hyperparameters, selected by exhaustive search over a small grid.

What would settle it

Train the same four-branch CRNN on ESC-50 while fixing the configuration to $N_{ss}=4$, cutoffs {3,6,10} kHz, and weights {0.4,0.2,0.2,0.2} without consulting test accuracy during configuration selection, and average over the five folds. If the mean accuracy lands near the 77.5% of the unsegmented CRNN-with-mixup system rather than near 81.9%, the reported improvement is mostly a test-set selection artifact.

Watch

Extended reading notes

Core claim

The paper's central claim is that sub-spectrogram segmentation beats full-spectrogram classification on ESC-50. Instead of extracting one log-mel spectrogram over the whole frequency range, the authors split the spectrogram into $N_{ss}$ adjacent frequency bands, compute a separate spectrogram for each band, classify each with the same CRNN (eight convolutional layers, two bidirectional GRU layers, and a fully connected layer), and fuse the resulting score vectors with weights $\omega_i$. After testing $N_{ss}$ from 1 to 6, different cutoff sets, and different fusion weights, the best reported configuration is $N_{ss}=4$, cutoffs {3,6,10} kHz, weights {0.4,0.2,0.2,0.2}, giving 81.9% accuracy. The paper also claims that the low-frequency portion carries most of the discriminative information, since both finer low-frequency segmentation and higher low-frequency fusion weights improve accuracy.

Load-bearing premise

The reported 81.9% accuracy is achieved by choosing the number of bands, the frequency cutoffs, and the fusion weights using the same ESC-50 test folds that are then used to report the accuracy, so the peak may partly reflect selection on the test set rather than a generalizable rule.

Editorial extensions

If this is right

If the central claim holds, frequency-splitting the spectrogram into four bands with low-frequency emphasis is a direct accuracy lever for ESC, worth 9.1 percentage points over the plain CNN baseline on ESC-50.
The optimal configuration uses more sub-bands and more fusion weight at low frequencies, implying that coarse high-frequency information still helps but contributes less than low-frequency detail.
Score level fusion alone adds 2.5 to 3.7 percentage points over uniform weight assignment in the tested two-, three-, and four-band cases.
The CRNN architecture and mixup augmentation each provide separate gains before segmentation is added, so the final result is the accumulation of four compatible components.
The method's message is that the choice of spectrogram segmentation and fusion weights can be as important as the network architecture for ESC accuracy.

Reading between the lines

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

The paper leaves implicit that the segmentation boundaries and fusion weights could be learned by the network; a natural extension is an attention mechanism over frequency bands, and one test is whether it rediscovers the {3,6,10} kHz split.
Because the reported configuration was selected using the ESC-50 test folds themselves, my inference is that the 9.1-point gain is an upper-bound estimate; a nested or held-out evaluation would likely show a smaller but still positive gain.
If low-frequency dominance is a genuine property of environmental sounds, the same sub-band recipe should transfer to other ESC datasets or to acoustic scene classification, but the paper only evaluates ESC-50, so this remains an open extension.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Dear [name],

Quick take: this is a competent engineering paper that combines known pieces — sub-spectrogram segmentation (from SubspectralNet), CRNN, mixup, and score fusion — and reports a decent number on ESC-50. The result is plausible, but the headline accuracy is picked after searching over configurations on the test folds, so treat 81.9% as an upper bound, not an unbiased estimate.

What's genuinely useful: the paper is clearly written, the architecture table is complete, and the ablation in Table VII shows each ingredient adds something. The finding that splitting the spectrogram into four bands with more weight on low frequencies helps is sensible and matches earlier work. I also like that they compare against their own CNN baseline instead of just citing other people's numbers. The experiment details (STFT parameters, mel filters, optimizer, mixup alpha) are specific enough that someone could reproduce it.

The soft spots are real. In Sections III-B and III-D, the number of segments, the band edges, and the fusion weights are chosen by exhaustive search over the ESC-50 test accuracy, and the same test set is then used to report the final 81.9%. There is no held-out split or nested cross-validation, so the reported number is the max over many configurations rather than a fair estimate on new data. That inflates the claim and makes '9.1% improvement over baseline' less meaningful than it looks. Second, there is no comparison with contemporaneous ESC-50 methods — only their own baseline — so we don't know whether 81.9% is actually competitive. Third, no variance or multiple runs are reported, so the difference between Nss=4 and Nss=3 (81.9 vs 81.6) may be noise.

None of this is fatal. The method is sensible and the core idea — low-frequency bands carry most of the discriminative information — is backed by their experiments. The paper would be acceptable as a workshop or conference paper if the authors add a proper held-out validation, report mean and standard deviation over folds, and compare with at least a couple of published ESC-50 methods. As it stands, it's a solid engineering report whose headline number needs independent confirmation.

My recommendation: send it to peer review; the topic is relevant and there is enough substance to be worth refereeing, but the authors should be asked to fix the evaluation protocol before acceptance.

Best.

Referee Report

3 major / 6 minor

Summary. The paper proposes an environmental sound classification (ESC) framework that splits the Logmel spectrogram into sub-bands, classifies each sub-spectrogram with a convolutional recurrent neural network (CRNN) trained with mixup augmentation, and fuses the per-sub-band class scores with learned weights. On the ESC-50 dataset the authors report 81.9% accuracy, an absolute improvement of 9.1% over their own CNN baseline. The claimed contributions are the sub-spectrogram segmentation strategy, the CRNN architecture, and score-level fusion.

Significance. If the accuracy claim were properly validated, the paper would make a modest but useful contribution: it applies a domain-motivated preprocessing step (sub-band division) to ESC, reports the architecture in sufficient detail to be reproduced, and shows that placing more weight on low-frequency sub-bands can help. The experimental tables are detailed enough to follow the pipeline. However, as submitted the central number is not an unbiased estimate: the segmentation points, number of segments, and fusion weights are selected by exhaustive search on the same test accuracy that is later reported, and no validation split, repeated cross-validation, or variance estimate is given. The significance of the work therefore depends on a validation experiment that is currently missing.

major comments (3)

[Section III-B and III-D, Tables I and VI] The optimal number of sub-spectrograms Nss, the segmentation points {fi}, and the fusion weights {omega_i} are chosen by maximizing classification accuracy on the ESC-50 test set, and the same test set is then used to report the final 81.9% figure. This makes the headline result the maximum over a searched grid rather than an estimate of performance on unseen data. The paper needs an independent validation protocol: for example, select all configuration parameters on a held-out split or via nested cross-validation, and only then measure final accuracy on a test fold that was not used for model or hyperparameter selection.
[Section IV and Table V] The experimental section does not specify how the ESC-50 folds are used, how many runs are averaged, or which random seeds are used. Because the accuracy differences among configurations in Table VI are as small as 0.3 percentage points (e.g., 81.6% vs 81.9%), the claimed optimality of Nss=4 and {3,6,10} kHz could easily be within run-to-run noise. Report mean and standard deviation over multiple runs, and state the exact train/validation/test split protocol.
[Abstract, Section IV-B, and Table VII] The comparison is only against the authors' own CNN baseline (72.8%). The claim of '9.1% improvement over traditional baseline schemes' is therefore narrower than the wording suggests, and it does not position the method against contemporaneous ESC-50 results, including the CNN baseline from the original ESC-50 paper and later published methods. At minimum, cite and compare with published ESC-50 accuracy numbers under the same fold protocol.

minor comments (6)

[Introduction] The introduction states that 'Logmel with CNN' reaches 73.2% on ESC-50, while Table III reports the CNN baseline as 72.8%; clarify which network and setting each number refers to.
[Figure 3] The vertical axis label contains the typo 'Accucacy'; the caption should also define the blue and orange lines more explicitly.
[Section III-C] In the sentence describing mixup, 'adogbark Logmel spectrogram and acryingbaby Logmel spectrogram' is missing spaces; this is a typo.
[Section II-A] The notation uses n both as a frame index and as an STFT frequency index in Eqs. (1) and (2); a consistent index convention would improve readability.
[Table II] The table gives output shapes, but it would be useful to state the input spectrogram size and the exact number of GRU directions (e.g., bi-directional with concatenation or summation).
[General] The paper does not mention whether code or trained models are available; providing them would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the 81.9% result is an empirical maximum over a transparent configuration search, not a derivation from its own inputs.

full rationale

The paper's central claim is an empirical accuracy measurement on ESC-50, not a quantity derived from its own assumptions. The only questionable procedure is that the configuration Nss=4, {fi}={3,6,10} kHz, and {wi}={0.4,0.2,0.2,0.2} is selected by exhaustive search over the same test accuracy that is then reported, as seen in Section IV-A ('Finally we assign different {wi} to each case for testing the performance of the model several times') and Section IV-B ('we obtained the highest accuracy of 81.9% with the combination of these methods'). This is a test-set selection / generalization caveat, and it may make the reported accuracy optimistically biased as an estimate of performance on new audio, but it is not circularity under the stated criteria: the accuracy is measured, not mathematically forced by the definition of the configuration, and the paper never presents the 81.9% figure as an independently predicted value. The self-citation [21] is used only as a pointer to mixup for environmental sound, with the original mixup method [24] also cited; it is not load-bearing for the reported accuracy. No equation in the paper defines the output in terms of the input or reduces any prediction to a fitted parameter. The derivation chain, such as it is, consists of standard feature extraction, a CRNN architecture, and score fusion, all evaluated empirically. Therefore no significant circularity is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper's central claim is an empirical performance number on a public dataset. It does not derive from first principles. The main free parameters are the segmentation and fusion settings, which were selected by maximizing test-set accuracy. The key untested assumption is that this selection procedure yields a generalizable configuration. No new physical or algorithmic entities are introduced.

free parameters (5)

Nss (number of sub-spectrogram segments) = 4
Selected by exhaustive search over Nss from 1 to 6 to maximize ESC-50 accuracy (Table I).
Segmentation points {fi} = {3, 6, 10} kHz
Chosen from several candidate sets in Table VI as the one giving the highest test accuracy.
Fusion weights {ωi} = {0.4, 0.2, 0.2, 0.2}
Selected by searching weight combinations on the test set, as shown in Fig. 3 and Table VI.
Mixup hyperparameter α = 0.2
Set by the authors; governs the strength of data augmentation and affects the training distribution (Table V).
Network architecture hyperparameters = See Table II (filter counts, GRU units, pooling)
Manually designed; no sensitivity analysis is provided, so their influence on the reported accuracy is unquantified.

assumptions (4)

domain assumption ESC-50 labels and audio clips are reliable enough to support the reported classification comparison
All conclusions rest on the public benchmark; the paper does not independently audit the dataset.
domain assumption Time-frequency representation via Logmel spectrogram retains enough discriminative information for classification after frequency truncation
The segmentation method assumes that each sub-band is still informative enough to be classified separately (Section III-B).
ad hoc to paper The exhaustively searched configuration (Nss, {fi}, {ωi}) generalizes to unseen data
This is the main untested premise. Selecting hyperparameters on the test set biases the 81.9% accuracy upward; no validation set is used.
standard math Standard definitions of STFT and mel filter bank as in Eqs. (1)-(2)
The feature extraction relies on conventional signal processing definitions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Sub-Spectrogram Segmentation for Environmental Sound Classification via Convolutional Recurrent Neural Network and Score Level Fusion." pith.science (2026). https://pith.science/paper/3FNJPOF2

@misc{pith2026190805863,
  author       = {Pith},
  title        = {Pith review of: Sub-Spectrogram Segmentation for Environmental Sound Classification via Convolutional Recurrent Neural Network and Score Level Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3FNJPOF2}},
  note         = {Machine review of arXiv:1908.05863}
}

read the original abstract

Environmental Sound Classification (ESC) is an important and challenging problem, and feature representation is a critical and even decisive factor in ESC. Feature representation ability directly affects the accuracy of sound classification. Therefore, the ESC performance is heavily dependent on the effectiveness of representative features extracted from the environmental sounds. In this paper, we propose a subspectrogram segmentation based ESC classification framework. In addition, we adopt the proposed Convolutional Recurrent Neural Network (CRNN) and score level fusion to jointly improve the classification accuracy. Extensive truncation schemes are evaluated to find the optimal number and the corresponding band ranges of sub-spectrograms. Based on the numerical experiments, the proposed framework can achieve 81.9% ESC classification accuracy on the public dataset ESC-50, which provides 9.1% accuracy improvement over traditional baseline schemes.

Figures

Figures reproduced from arXiv: 1908.05863 by the authors.

**Figure 1.** Logmel spectrograms of crying baby, rain, helicopter and dog bark for entire freaquency band, where the horizontal axis denotes the time dimension and the vertical axis denotes the frequency dimension. B. DNN Generally speaking, DNN is realized by connecting multiple layers of neurons to form a more powerful neural network, which typically includes MLP, CNN and recurrent neural network (RNN) structures. MLP is a cla… view at source ↗

**Figure 2.** Compare the baseline system and the proposed Sub-Spectrogram Segmentation system. In this figure, the first branch is the baseline system, which [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Compare using and not using sub-spectrogram segmentation on CRNN [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 20 canonical work pages

[1]

Acoustic scene classiﬁcation: Classifying environments from the sounds they produce,

D. Barchiesi, D. Giannoulis, D. Stowell, and M. D. Plumbley, “Acoustic scene classiﬁcation: Classifying environments from the sounds they produce,” IEEE Signal Process. Magazine , vol. 32, no. 3, pp. 16–34, 2015

work page 2015
[2]

Structure and function of auditory cortex: music and speech,

R. J. Zatorre, P. Belin, and V . B. Penhune, “Structure and function of auditory cortex: music and speech,” Trends in cognitive sciences, vol. 6, no. 1, pp. 37–46, 2002

work page 2002
[3]

Environmental sound recognition: A survey,

S. Chachada and C.-C. J. Kuo, “Environmental sound recognition: A survey,” APSIPA Trans. Signal, Inf. Process. , vol. 3, 2014

work page 2014
[4]

Environmental sound recog- nition with time–frequency audio features,

S. Chu, S. Narayanan, and C.-C. J. Kuo, “Environmental sound recog- nition with time–frequency audio features,” IEEE Trans. Audio, Speech, Language Process., vol. 17, no. 6, pp. 1142–1158, 2009

work page 2009
[5]

Comparison of techniques for environmental sound recognition,

M. Cowling and R. Sitte, “Comparison of techniques for environmental sound recognition,” Pattern Recogn. Lett. , vol. 24, no. 15, pp. 2895– 2907, 2003

work page 2003
[6]

Short-time energy, magnitude, zero crossing rate and autocorrelation measurement for discriminating voiced and unvoiced segments of speech signals,

M. Jalil, F. A. Butt, and A. Malik, “Short-time energy, magnitude, zero crossing rate and autocorrelation measurement for discriminating voiced and unvoiced segments of speech signals,” in Proc. Int. Conf. TAEECE, 2013, pp. 208–212

work page 2013
[7]

Sound event recognition in unstructured environments us- ing spectrogram image processing,

J. W. Dennis, “Sound event recognition in unstructured environments us- ing spectrogram image processing,” Nanyang Technological University, Singapore, 2014

work page 2014
[8]

L. R. Rabiner, B.-H. Juang, and J. C. Rutledge, Fundamentals of speech recognition. Prentice-Hall, 1993, vol. 14

work page 1993

Show all 24 references

[9]

Environmental sound classiﬁcation with convolutional neural networks,

K. J. Piczak, “Environmental sound classiﬁcation with convolutional neural networks,” in Proc. 25th Int. Workshop Mach. Learning Signal Process., 2015, pp. 1–6

2015
[10]

A fuzzy k-nearest neighbor algorithm,

J. M. Keller, M. R. Gray, and J. A. Givens, “A fuzzy k-nearest neighbor algorithm,” IEEE Trans. on Systems, Man, and Cybernetics , no. 4, pp. 580–585, 1985

1985
[11]

Scholkopf and A

B. Scholkopf and A. J. Smola, Learning with kernels: support vector machines, regularization, optimization, and beyond . MIT press, 2001

2001
[12]

Random forests,

L. Breiman, “Random forests,” Machine learning , vol. 45, no. 1, pp. 5–32, 2001

2001
[13]

Audio based event detection for multimedia surveillance,

P. K. Atrey, N. C. Maddage, and M. S. Kankanhalli, “Audio based event detection for multimedia surveillance,” in 2006 IEEE Int. Conf. Acoustics Speech and Signal Process. Proceedings, vol. 5. IEEE, 2006, pp. V–V

2006
[14]

Feature learning with matrix factorization applied to acoustic scene classiﬁcation,

V . Bisot, R. Serizel, S. Essid, and G. Richard, “Feature learning with matrix factorization applied to acoustic scene classiﬁcation,” IEEE/ACM Trans. Audio, Speech, Language Process., vol. 25, no. 6, pp. 1216–1229, 2017

2017
[15]

Learning from between-class examples for deep sound recognition,

Y . Tokozume, Y . Ushiku, and T. Harada, “Learning from between-class examples for deep sound recognition,” In ICLR, 2018

2018
[16]

Robust sound event classiﬁcation using deep neural networks,

I. McLoughlin, H. Zhang, Z. Xie, Y . Song, and W. Xiao, “Robust sound event classiﬁcation using deep neural networks,” IEEE/ACM Trans. Audio, Speech, Language Process. , vol. 23, no. 3, pp. 540–552, 2015

2015
[17]

Classifying environmental sounds using image recognition networks,

V . Boddapati, A. Petef, J. Rasmusson, and L. Lundberg, “Classifying environmental sounds using image recognition networks,” Procedia computer science, vol. 112, pp. 2048–2056, 2017

2017
[18]

Dilated convolution neural network with leakyrelu for environmental sound classiﬁcation,

X. Zhang, Y . Zou, and W. Shi, “Dilated convolution neural network with leakyrelu for environmental sound classiﬁcation,” in Proc. 22nd Int. Conf. Digital Signal Process. , 2017, pp. 1–5

2017
[19]

Subspectralnet-using sub- spectrogram based convolutional neural networks for acoustic scene classiﬁcation,

S. S. R. Phaye, E. Benetos, and Y . Wang, “Subspectralnet-using sub- spectrogram based convolutional neural networks for acoustic scene classiﬁcation,” arXiv preprint arXiv:1810.12642 , 2018

2018 arXiv
[20]

Empirical evaluation of gated recurrent neural networks on sequence modeling,

J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,” arXiv preprint arXiv:1412.3555, 2014

2014 arXiv
[21]

Deep convolutional neural network with mixup for environmental sound classiﬁcation,

Z. Zhang, S. Xu, S. Cao, and S. Zhang, “Deep convolutional neural network with mixup for environmental sound classiﬁcation,” in Chinese Conf. PRCV. Springer, 2018, pp. 356–367

2018
[22]

V oice recognition algorithms using mel frequency cepstral coefﬁcient (mfcc) and dynamic time warping (dtw) techniques,

E. I. Muda L, Begam M, “V oice recognition algorithms using mel frequency cepstral coefﬁcient (mfcc) and dynamic time warping (dtw) techniques,” arXiv preprint arXiv:1003.4083 , 2010

2010 arXiv
[23]

Esc: Dataset for environmental sound classiﬁcation,

K. J. Piczak, “Esc: Dataset for environmental sound classiﬁcation,” in Proc. 23rd ACM Int. Conf. Multimedia , 2015, pp. 1015–1018

2015
[24]

Mixup: Beyond empirical risk minimization,

H. Zhang, M. Cisse, Y . N. Dauphin, and D. Lopez-Paz, “Mixup: Beyond empirical risk minimization,” arXiv preprint arXiv:1710.09412 , 2017

2017 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Acoustic scene classiﬁcation: Classifying environments from the sounds they produce,

D. Barchiesi, D. Giannoulis, D. Stowell, and M. D. Plumbley, “Acoustic scene classiﬁcation: Classifying environments from the sounds they produce,” IEEE Signal Process. Magazine , vol. 32, no. 3, pp. 16–34, 2015

work page 2015

[2] [2]

Structure and function of auditory cortex: music and speech,

R. J. Zatorre, P. Belin, and V . B. Penhune, “Structure and function of auditory cortex: music and speech,” Trends in cognitive sciences, vol. 6, no. 1, pp. 37–46, 2002

work page 2002

[3] [3]

Environmental sound recognition: A survey,

S. Chachada and C.-C. J. Kuo, “Environmental sound recognition: A survey,” APSIPA Trans. Signal, Inf. Process. , vol. 3, 2014

work page 2014

[4] [4]

Environmental sound recog- nition with time–frequency audio features,

S. Chu, S. Narayanan, and C.-C. J. Kuo, “Environmental sound recog- nition with time–frequency audio features,” IEEE Trans. Audio, Speech, Language Process., vol. 17, no. 6, pp. 1142–1158, 2009

work page 2009

[5] [5]

Comparison of techniques for environmental sound recognition,

M. Cowling and R. Sitte, “Comparison of techniques for environmental sound recognition,” Pattern Recogn. Lett. , vol. 24, no. 15, pp. 2895– 2907, 2003

work page 2003

[6] [6]

Short-time energy, magnitude, zero crossing rate and autocorrelation measurement for discriminating voiced and unvoiced segments of speech signals,

M. Jalil, F. A. Butt, and A. Malik, “Short-time energy, magnitude, zero crossing rate and autocorrelation measurement for discriminating voiced and unvoiced segments of speech signals,” in Proc. Int. Conf. TAEECE, 2013, pp. 208–212

work page 2013

[7] [7]

Sound event recognition in unstructured environments us- ing spectrogram image processing,

J. W. Dennis, “Sound event recognition in unstructured environments us- ing spectrogram image processing,” Nanyang Technological University, Singapore, 2014

work page 2014

[8] [8]

L. R. Rabiner, B.-H. Juang, and J. C. Rutledge, Fundamentals of speech recognition. Prentice-Hall, 1993, vol. 14

work page 1993