Pith. sign in

REVIEW 3 major objections 5 minor 48 references

Beyond Words: Interjection Classification for Improved Human-Computer Interaction

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Audio data augmentation turns a tiny interjection dataset into a usable classifier for unseen speakers and noise.

desk verdict A genuinely new dataset and task framing for interjection classification, with an augmentation recipe that shows directional gains; but the central 'combination is best' claim is confounded by training-set size and the evaluation lacks error bars. read the letter →

arxiv 2509.03181 v1 pith:24LNMP2P submitted 2025-09-03 cs.HC cs.LG

Yaniv Goren , Yuval Cohen , Alexander Apartsin , Yehudit Aperstein This is my paper · ORCID
classification cs.HCcs.LG
keywords interjectionclassificationhuman-computerinteractiondataaugmentationaudiospeechunderstandingvoiceassistantsfeedforwardneuralnetworkMFCCfeatures
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's central claim is that a previously ignored class of speech—short interjections such as 'mmm', 'ahah', and 'oy'—can be made machine-classifiable with a modest, purpose-built dataset plus heavy artificial augmentation. The authors argue that existing speech datasets sit at word, phoneme, or event level, leaving interjections between categories, so they recorded their own collection from five speakers. Their baseline feedforward network, trained on mean audio features, performs poorly on clean held-out speakers; augmenting the training set with tempo shifts, pitch shifts, and background-noise mixtures raises accuracy substantially—most clearly when all three methods are combined, e.g., a 63.6% relative improvement for one unseen speaker and 42.9% for another. In a natural-noise test, tempo plus background noise beats clean training most strongly (0.667 vs 0.45). If true, the result matters because voice assistants currently drop these 'non-words' even though they carry agreement, hesitation, and emotional information.

What carries the argument

The carrying mechanism is the Data Augmentation System (DAS), a Pysox-based pipeline that generates synthetic training clips from clean recordings by changing tempo, shifting pitch in semitones, and mixing each clip with one of nine background scenes at controlled volume weights. DAS multiplies about thirty-two minutes of original recordings into over three hundred hours of augmented audio. The network itself is a simple fully connected feedforward architecture with three hidden layers, Adam optimizer, and softmax output; it operates on fixed-length 1.55-second clips represented by 193 mean feature values (MFCC, Mel-spectrogram, chroma-stft, spectral contrast, tonnetz). The argument assigns

What would settle it

Collect interjections from natural conversations—overlapping speech, varying prosody, multiple languages—and run the same clean-only versus augmented training comparison on them; if the augmented model's advantage disappears or accuracy stays near chance on the four interjections, the transfer claim fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes the first dataset and baseline for interjection classification, treating four language-independent interjections (nah, mmm, ahah, oy) plus a non-interjection class as a five-way classification problem. The classifier is a fully connected feedforward network whose input is a 193-dimensional vector of mean audio features per clip. The discovery claim is that data augmentation—tempo, pitch, and background noise, applied singly and in combination—significantly improves generalization to held-out speakers and to natural noise relative to clean-only training. In the first scenario the best result for each unseen speaker comes from the triple augmentation (63

Load-bearing premise

The load-bearing premise is that isolated, self-recorded clips of five speakers saying four interjections are representative of the interjections heard in real human-computer dialogue, so gains on held-out data from the same recording protocol will transfer to the HCI use case.

Editorial extensions

If this is right

  • Voice assistants could add a small interjection classifier in front of or alongside ASR, mapping interjections to actions such as confirmation, hesitation, or interruption without parsing full sentences.
  • The augmentation recipe demonstrated here gives a concrete starting point for other low-resource short-utterance tasks: clean data plus tempo and background-noise variants can exceed the accuracy of much larger clean corpora.
  • The published dataset and DAS library make interjection classification a reproducible benchmark, so future work can compare architectures and augmentation policies on the same data.
  • The noisy-environment result implies that the choice of augmentation should be environment-aware: tempo plus background noise helped most in natural noise, while pitch alone reduced accuracy compared with clean training.
  • Because improvements are measured on held-out speakers from the same recording protocol, the paper's design also establishes an evaluation protocol for this task (two train speakers, two test speakers, one validation speaker).

Reading between the lines

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

  • If mean features suffice for these near-static interjections, frame-level temporal models (LSTM, CNN over spectrograms) may not outperform the mean-feature FNN on this dataset; testing that would separate the contribution of the input representation from the augmentation claim.
  • The four selected interjections are claimed to be language-independent, but only Hebrew and Spanish speakers were recorded; a direct extension would record the same interjections from speakers of other languages and check whether cross-language accuracy holds.
  • The two-test-speaker design leaves the clean baselines fragile; a leave-one-speaker-out evaluation across all five speakers, with variance reported, would tighten the quantitative claim.
  • The augmentation gains could be re-examined under a stricter real-use test: interjections produced in fluent dialogue, with co-articulation and overlapping speech, rather than isolated VAD-triggered clips.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces a new task and dataset for classifying four interjections (nah, mmm, ahah, oy) plus a non-interjection word class, recorded from five speakers in isolation under a VAD-based protocol. The authors build a fully connected feedforward classifier on 193 mean audio features per clip, and study how tempo, pitch, and background-noise augmentation (alone and in combination) affect held-out-speaker accuracy (Table IV) and robustness to natural background noise (Table V). They report large relative improvements, e.g., a 63.6% relative gain for speaker C and 42.9% for speaker D with the Tempo+Pitch+BGN set, and release the dataset, augmentation code, model, and evaluation scripts.

Significance. If the central claim is supported, this is a useful early benchmark resource for an under-served problem in speech-based HCI: interjections are routinely discarded by ASR, and a public dataset plus reproducible augmentation pipeline would be a real community contribution. The authors explicitly provide data and code, which is a strength. However, the empirical evidence for the main claim is weakened by a confounded experimental design (training-set size varies across augmentation conditions), by the absence of error bars or significance tests, and by the intramural, small-scale evaluation. The paper's value is therefore preliminary; the core idea is plausible but the current experiments do not yet establish that the reported gains are attributable to the augmentation methods rather than to dataset size or run-to-run variation.

major comments (3)
  1. [§V.A, Table III, and §VI (Table IV)] The central claim that the triple augmentation combination is best is confounded. The seven augmented sets differ simultaneously in the number of original samples per class (120, 120, 100, 60, 40, 40, 10) and in the total generated samples per class (2,520 to 9,320). The Tempo+Pitch+BGN row, which yields the best results in Table IV, is also the row with the largest training set (9,320 samples/class). Thus the reported 63.6% and 42.9% relative gains may be due to more training data rather than to a synergistic effect of the three methods. A fair test should match total training-set size and original-sample count across conditions, or otherwise isolate each augmentation method and its combinations at a fixed data budget. As written, the sentence 'the combination of the augmentation methods gives better results' is not supported by controlled evidence.
  2. [§VI, Tables IV and V] All reported accuracies are single runs with no confidence intervals, repeated seeds, or significance tests. Speaker C's clean baseline of 0.286 is close to the 0.2 chance level, so small random fluctuations could change the ranking of methods. In Scenario 2, Table V is described as 'average accuracy over all subjects' but no variance is reported. The repeated use of 'significantly improves' in the abstract and Section VI is therefore not justified by the statistics presented. The authors should report multiple runs (e.g., different seeds), mean±std, and ideally a paired test across conditions.
  3. [§IV.B, §IV.D, and Tables IV–V] There is a mismatch between the stated evaluation metric and the reported results: Section IV.D says the model is evaluated using the F1 score, but Tables IV and V report 'accuracy' and the text speaks of accuracy throughout. Please clarify which metric is used. Additionally, the feature representation collapses each clip into 193 mean feature values, discarding temporal structure; for interjections, prosodic and temporal contours are likely informative. No ablation or frame-level/temporal model is provided to show that this loss is acceptable. This does not overturn the augmentation result, but it limits the strength of the 'baseline' claim for the task.
minor comments (5)
  1. [§V.A vs. Table III] The text says 'For each speaker, 120 different audio samples were used from our original dataset' to create the augmented sets, but Table III lists different original-sample counts per class (120, 120, 100, 60, 40, 40, 10). Please reconcile this inconsistency, as it is important for reproducibility.
  2. [§IV.C, Eq. (1)] Equation (1) appears garbled in the PDF ('S_mix = W_orig * S_orig + W_BGN * S_BGN' is clear from context, but the rendered formula is not). Also, the weighting ranges mentioned in §V.A (W_orig in 0.83–0.93, W_BGN in 0.07–0.17) are not tied to the specific conditions in Table III.
  3. [§IV.D / Tables IV–V] The text says the F1 score is used, but tables report accuracy. Align the text and tables.
  4. [§IV.B] The paper says each folder contains 'approximately 800 fixed-length recordings', but Table II gives per-speaker per-word counts of 850, 550, 300, 300, 300. Clarify how filtering and padding produce the final folder counts.
  5. [Throughout] Minor typos: 'Chorma-stft' should be 'Chroma-stft'; 'speake' in the reader's note aside, in the manuscript 'an -data augmentation' appears in the introduction. Also, reference [41] for Auditok points to a U.S. patent, which may be incorrect; the software/tool should be cited.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the reported accuracies are direct measurements on held-out data; the central claims are not restatements of the paper's inputs.

full rationale

This paper makes no mathematical derivation claims. The central assertions—that a new interjection dataset was collected, that a feedforward baseline achieves the stated accuracies, and that pitch/tempo/background-noise augmentation is associated with higher accuracy on held-out speakers and on natural noise—are supported by direct experimental measurements reported in Tables IV and V. The evaluation is intramural (the dataset, task, and baseline are all introduced by the authors), but that is an external-validity limitation, not circularity: the test data are held-out speakers and natural-noise recordings not used in training, and the augmented training sets are not defined in terms of the test accuracies. No fitted constant is subsequently relabeled as a prediction; no equation equates an input to an output by construction; and no load-bearing step depends on a self-citation, as the reference list contains no prior work by the present authors. The main methodological risk—that the augmented-set comparisons vary simultaneously in original-sample count and total sample size (Table III), so the 'combination is best' conclusion may be confounded by dataset size—is an experimental design issue, not a circular-reasoning issue, and does not satisfy the standard for flagging circularity. Accordingly, no circular steps are identified and the score is 0.

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

The central numbers rest on hand-chosen augmentation ranges, unreported network widths, a fixed clip-length rule with arbitrary 0.45 and 1.55 s cutoffs, and a 193-mean-feature representation chosen without ablation. The axioms are mostly domain assumptions about label preservation and the representativeness of the self-collected data.

free parameters (4)
  • Augmentation ranges: tempo factor, pitch semitones, BGN weights = tempo 0.86-1.14; pitch -2.4..2.4 semitones; W_orig 0.83-0.93, W_BGN 0.07-0.17
    Chosen by the authors to preserve label semantics (Section V.A); no experiment justifies these ranges, and pitch is later found harmful in scenario 2.
  • Network hyperparameters (learning rate 0.009, 3 hidden layers, layer widths not reported) = lr=0.009, 3 hidden layers
    Selected after 'extensive experimentation' (Section V.B); widths of hidden layers are never stated, so a key free parameter is unreported.
  • Clip length 1.55 s and minimum 0.45 s = 1.55 s max, 0.45 s min
    Chosen by hand to accommodate long interjections and exclude noise snippets (Section IV.B); affects the feature matrix dimensions.
  • Feature-set composition (193 mean features) = 193 mean MFCC/mel/chroma/contrast/tonnetz columns
    The exact Librosa parameters (n_mfcc, n_fft, hop length, window) are not given; averaging over time is a modeling choice that removes temporal structure.
assumptions (5)
  • domain assumption MFCC and related spectral features averaged over time retain enough discriminative information for interjection classification
    Section IV.B reduces features to per-file mean values, discarding the temporal contour; no ablation supports this.
  • domain assumption The four interjections are language-independent and semantically clear across speakers
    Section IV.A asserts this; 'oy' is labeled 'mainly Jewish' and speakers are Hebrew/Spanish, so language independence is untested.
  • domain assumption Applying tempo, pitch, and noise transforms does not change the interjection label
    Section V.A states semantic validity must be maintained but provides no listening or validation check.
  • ad hoc to paper The recorded clean samples plus VAD segmentation yield correctly labeled clips
    Section IV.A: Auditok-based recording with length cutoffs; label correctness is asserted, no annotation verification or inter-annotator agreement reported.
  • standard math Standard math: FFT/DCT as implemented by Librosa, Adam optimizer, softmax cross-entropy
    Used throughout Sections IV.D and V; standard tools, no formal verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Words: Interjection Classification for Improved Human-Computer Interaction." pith.science (2026). https://pith.science/paper/24LNMP2P

@misc{pith2026250903181,
  author       = {Pith},
  title        = {Pith review of: Beyond Words: Interjection Classification for Improved Human-Computer Interaction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/24LNMP2P}},
  note         = {Machine review of arXiv:2509.03181}
}
read the original abstract

In the realm of human-computer interaction, fostering a natural dialogue between humans and machines is paramount. A key, often overlooked, component of this dialogue is the use of interjections such as "mmm" and "hmm". Despite their frequent use to express agreement, hesitation, or requests for information, these interjections are typically dismissed as "non-words" by Automatic Speech Recognition (ASR) engines. Addressing this gap, we introduce a novel task dedicated to interjection classification, a pioneer in the field to our knowledge. This task is challenging due to the short duration of interjection signals and significant inter- and intra-speaker variability. In this work, we present and publish a dataset of interjection signals collected specifically for interjection classification. We employ this dataset to train and evaluate a baseline deep learning model. To enhance performance, we augment the training dataset using techniques such as tempo and pitch transformation, which significantly improve classification accuracy, making models more robust. The interjection dataset, a Python library for the augmentation pipeline, baseline model, and evaluation scripts, are available to the research community.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 37 canonical work pages

  1. [1]

    Interjections and emotion (with special reference to “surprise

    Goddard, Cliff. “Interjections and emotion (with special reference to “surprise” and “disgust”).” Emotion review, Vol. 6,.no. 1, pp. 53-63. 2014

    work page 2014

  2. [2]

    Interjections in English: Neglected but Important Aspect of Foreign Language Learning and Teaching

    Hışmanoğlu, Murat. “Interjections in English: Neglected but Important Aspect of Foreign Language Learning and Teaching.” Journal of Theory & Practice in Education (JTPE), Vol. 6. No.1 2010

    work page 2010

  3. [3]

    On the margins of language: Ideophones, interjections and dependencies in linguistic theory

    Dingemanse, Mark. "On the margins of language: Ideophones, interjections and dependencies in linguistic theory." Dependencies in language Vol. 15,: pp. 195-203. 2017

    work page 2017

  4. [4]

    Interjections as pragmatic markers

    Norrick, Neal R. "Interjections as pragmatic markers." Journal of pragmatics, vol. 41, no. 5, pp. 866-891. 2009

    work page 2009

  5. [5]

    Alexa, Siri, Cortana, and more: an introduction to voice assistants

    M. B. Hoy, “Alexa, Siri, Cortana, and more: an introduction to voice assistants”, Medical reference services quarterly, vol.37, no. 1, pp. 81-88, 2018 , doi: 10.1080/02763869.2018.1404391

    arXiv 2018

  6. [6]

    User Experience Comparison of Intelligent Personal Assistants: Alexa, Google Assistant, Siri and Cortana

    A. Berdasco, G. López, I. Diaz, L. Quesada, , L. A. Guerrero, “User Experience Comparison of Intelligent Personal Assistants: Alexa, Google Assistant, Siri and Cortana”. In: Multi-disciplinary Digital Publishing Institute Proceedings, vol. 31, no. 1, Art no. 51, 2019, doi: https://doi.org/10.3390/proceedings2019031051

    work page doi:10.3390/proceedings2019031051 2019

  7. [7]

    Talk to me: Exploring user interactions with the Amazon Alexa

    I. Lopatovska, K. Rink, I. Knight, K. Raines, K. Cosenza, H. Williams, A. Martinez, “Talk to me: Exploring user interactions with the Amazon Alexa”. Journal of Librarianship and In - formation Science, vol. 51, no. 4, pp. 984-997, 2019, doi: https://doi.org/10.1177/0961000618759414

    work page doi:10.1177/0961000618759414 2019

  8. [8]

    The State of Speech in HCI: Trends, Themes and Challenges

    L.Clark, P. Doyle, D. Garaialde, E. Gilmartin, S. Schlögl, J. Edlund, & R. B. Cowan, “The State of Speech in HCI: Trends, Themes and Challenges”. Interacting with Computers, vol. 31, no. 4, pp. 349-371, 2019 . https://doi.org/10.1093/iwc/iwz016

    work page doi:10.1093/iwc/iwz016 2019

Show all 48 references
  1. [9]

    Analysis of acoustic-prosodic features related to paralinguistic information carried by interjections in dialogue speech

    Ishi, Carlos T., Hiroshi Ishiguro, and Norihiro Hagita. "Analysis of acoustic-prosodic features related to paralinguistic information carried by interjections in dialogue speech." Twelfth Annual Conference of the International Speech Communication Association. 2011

    2011

  2. [10]

    Speech Recognition: Keyword Spotting Through Image Recognition

    S. K. Gouda, S. Kanetkar, D. Harrison, & M. K. Warmuth, “Speech Recognition: Keyword Spotting Through Image Recognition”. arXiv preprint arXiv:1803.03759. 2018 . https:doi.org/10.48550/arXiv.1803.03759

    2018 doi arXiv

  3. [11]

    Speech emotion recognition using deep learning techniques: A review

    R. A. Khalil, E. Jones, M. I. Babar, T. Jan, M. H. Zafar & T. Alhussain, “Speech emotion recognition using deep learning techniques: A review”. IEEE Access, vol. 7, pp. 117327-117345, 2019, doi: 10.1109/ACCESS.2019.2936124

    2019

  4. [12]

    Speech commands: A dataset for limited - vocabulary speech recognition

    P. Warden, “Speech commands: A dataset for limited - vocabulary speech recognition”. arXiv preprint, arXiv:1804.03209. 2018 . https://doi.org/10.48550/arXiv.1804.03209

    2018 doi arXiv

  5. [13]

    E2E-DASR: End-to-end deep learning-based dysarthric automatic speech recognition

    A. Almadhor, R. Irfan, J. Gao, N. Saleem, H. T. Rauf, & S. Kadry, “E2E-DASR: End-to-end deep learning-based dysarthric automatic speech recognition”. Expert Systems with Applications, vol. 222, Art. no. 119797, 2023 , https://doi.org/10.1016/j.eswa.2023.119797

    2023

  6. [14]

    Introduction to acoustic event and scene analysis

    K. Imoto, “Introduction to acoustic event and scene analysis”. Acoustical Science and Technology, vol. 39, no. 3, pp. 182-188, 2018. https://doi.org/10.1250/ast.39.182

    2018 doi

  7. [15]

    Benchmarking performance of object detection under image distortions in an un- controlled environment

    A. Beghdadi, M. Mallem, L. Beji, “Benchmarking performance of object detection under image distortions in an un- controlled environment”. In 2022 IEEE International Conference on Image Processing (ICIP), Bordeaux, France, pp. 2071-2075, October 2022. doi: 10.1109/ICIP46576.202...

    2022

  8. [16]

    Voices obscured in com-plex environmental settings (voices) corpus

    C. Richey, M. A, Barrios, Z. Armstrong, C. Bartels, H. Franco, M. Graciarena, P. Gamble, “Voices obscured in com-plex environmental settings (voices) corpus”. arXiv preprint arXiv:1804.05053, 2018 . https://doi.org/10.48550/arXiv.1804.05053

    2018 doi arXiv

  9. [17]

    A Review of Music Analysis Techniques

    K. R. Kulkarni, & S. R. R. Naik. “A Review of Music Analysis Techniques”, International Research Journal of Engineering and Technology, vol. 5, no. 3, pp. 4102-4108, March

  10. [19]

    Semi-Supervised Multiple Source Localization Using Relative Harmonic Coefficients Under Noisy and Reverberant Environments,

    Y. Hu, P. N. Samarasinghe, S. Gannot and T. D. Abhayapala, "Semi-Supervised Multiple Source Localization Using Relative Harmonic Coefficients Under Noisy and Reverberant Environments," in IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 3108-3123, ...

    2020

  11. [20]

    Sound Event Detection in Urban Audio with Single and Multi -Rate Pcen,

    C. Ick and B. McFee, "Sound Event Detection in Urban Audio with Single and Multi -Rate Pcen," IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2021), Toronto, Canada, pp. 880-884, June 2021, doi: 10.1109/ICASSP39728.2021.9414697

    2021

  12. [21]

    Analysis of speech recognition techniques

    P. Gadekar, M. H. Kaldane, D. Pawar, O. Jadhav, & A. Patil, “Analysis of speech recognition techniques”, International Journal of Advance Research, Ideas and Innovations in Technology, vol

  13. [22]

    MFCC-based recurrent neural network for auto-matic clinical depression recognition and assessment from speech

    E. Rejaibi, A. Komaty, F. Meriaudeau, S. Agrebi & A. Othmani, “MFCC-based recurrent neural network for auto-matic clinical depression recognition and assessment from speech”, Biomedical Signal Processing and Control, vol. 71, part A, Art. no. 103107, 2022, doi: 10.1016/j.bspc....

    2022

  14. [23]

    No" - Informal no mmm mhm, uh-hu

    to recover the most probable utterance by modelling the conditional probability of the nth word, using the (n -1) earlier words. Linguistic and pronunciation dictionaries are often used to improve the decoding performance. An acoustic model [23] is a fundamental part of the AS...

  15. [24]

    No. 2. pp. 1129-1132, 2019. https://www.academia.edu/download/59152470/V5I2- 158420190506-30272-1e4jqre.pdf

    2019

  16. [25]

    Bagged support vector machines for emotion recognition from speech

    A. Bhavan, P. Chauhan, & R. R. Shah, “Bagged support vector machines for emotion recognition from speech”. Knowledge-Based Systems, vol. 184, Art. no. 104886, 2019. https://doi.org/10.1016/j.knosys.2019.104886

    2019

  17. [26]

    Automatic speech recognition: Systematic literature review

    S. Alharbi, M. Alrazgan, A. Alrashed, T. Alnomasi, R. Almojel, R. Alharbi, & M. Almojil, “Automatic speech recognition: Systematic literature review”, IEEE Access, vol. 9, pp. 131858-131876. 2021 doi: 10.1109/ACCESS.2021.3112535

    2021

  18. [27]

    A systematic review of hidden Markov models and their applications

    B. Mor, S. Garhwal, & A. Kumar, “A systematic review of hidden Markov models and their applications”. Archives of computational methods in engineering, vol.28, pp. 1429 -1448, 2021. https://doi.org/10.1007/s11831-020-09422-4

    2021 doi

  19. [28]

    Dynamic time warping (DTW) algorithm in speech: a review

    M. Yadav, & A. Alam, “Dynamic time warping (DTW) algorithm in speech: a review”, Int. J. Res. Electron. Comput. Eng, vol. 6., no. 1, pp. 524-528, 2018

    2018

  20. [29]

    Deep spoken key-word spotting: An overview

    I. López-Espejo, Z. H. Tan, J. H. Hansen, & J. Jensen, “Deep spoken key-word spotting: An overview”, IEEE Access, vol. 10, pp. 4169-4199, 2021 . doi: 10.1109/ACCESS.2021.3139508

    2021

  21. [30]

    Audio adversarial examples generation with recurrent neural networks

    K. H., Chang, P. H. H. Huang-Yu, Y. Jin, & T. C. Wang, “Audio adversarial examples generation with recurrent neural networks”. In 2020 25th Asia and South Pacific Design Automation Conference (ASP-DAC), Beijing, China, pp. 488 - 493, Janu-ary 2020. doi: 10.1109/ASP-DAC47756.20...

    2020

  22. [31]

    A deep neural network-driven feature learning method for poly -phonic acoustic event detection from real-life recordings

    M. Mulimani, A. B. Kademani, & S. G. Koolagudi, “A deep neural network-driven feature learning method for poly -phonic acoustic event detection from real-life recordings”. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Ba...

    2020

  23. [32]

    Recent advances in end -to-end automatic speech recognition. APSIPA

    J. Li, “Recent advances in end -to-end automatic speech recognition. APSIPA”, Transactions on Signal and Information Processing, vol. 11, no. 1, pp.1-61 ,2022

    2022

  24. [33]

    Survey on machine learning in speech emotion recognition and vision systems using a recurrent neural network (RNN)

    S. P. Yadav, S. Zaidi, A. Mishra, & V. Yadav, “Survey on machine learning in speech emotion recognition and vision systems using a recurrent neural network (RNN)”. Archives of Computational Methods in Engineering, vol. 29, no. 3, pp. 1753- 1770, 2022. https://doi.org/10.1007/s...

    2022 doi

  25. [34]

    Specaugment: A simple data aug-mentation method for automatic speech recognition

    D. S. Park, W. Chan, Y. Zhang, C. C. Chiu, B. Zoph, E. D. Cubuk, & Q. V. Le, “Specaugment: A simple data aug-mentation method for automatic speech recognition”. arXiv preprint arXiv:1904.08779, 2019 . https://doi.org/10.48550/arXiv.1904.08779

    1904 doi arXiv

  26. [35]

    Improved Meta Learning for Low Resource Speech Recognition

    S. Singh, R. Wang, & F. Hou, “Improved Meta Learning for Low Resource Speech Recognition”. In 2022 IEEE Interna-tional Conference on Acoustics, Speech and Signal Processing (ICASSP 2022), Singapore, Singapore, pp. 4798-4802, May 2022. doi: 10.1109/ICASSP43922.2022.9746899

    2022

  27. [36]

    Improved Vocal Tract Length Perturbation for a State -of-the-Art End -to-End Speech Recognition System

    C. Kim, M. Shin, A. Garg, & D. Gowda, “Improved Vocal Tract Length Perturbation for a State -of-the-Art End -to-End Speech Recognition System”. In Interspeech, pp. 739 -743, September 2019. doi: 10.21437/Interspeech.2019-3227

    2019 doi

  28. [37]

    The analysis of interjection types and emotional state used in “Sam & Cat

    K. B. Nabila, & O. T. Ena, “The analysis of interjection types and emotional state used in “Sam & Cat” sitcom”. Rainbow Journal of Literature, Linguistics and Culture Studies, vol. 11, no. 2, 9-15. 2022 . https://doi.org/10.15294/rainbow.v11i2.55333

    2022 doi

  29. [38]

    Deep Convolutional Neural Networks and Data Augmentation for Environmental Sound Classification,

    J. Salamon and J. P. Bello, "Deep Convolutional Neural Networks and Data Augmentation for Environmental Sound Classification," in IEEE Signal Processing Letters, vol. 24, no. 3, pp. 279-283, March 2017. doi: 10.1109/LSP.2017.2657381

    2017

  30. [39]

    Interjections oh! wow! gosh! ah! god! as a means of expressing emotions of personages in the novel ‘can you keep a secret’

    R. A. Ulfatovna, “Interjections oh! wow! gosh! ah! god! as a means of expressing emotions of personages in the novel ‘can you keep a secret’.” by S. Kinsel-la. ACADEMICIA: An In-ternational Multidisciplinary Research Journal, vol. 11, no. 1, pp. 865-870, 2021. doi: 10.5958/224...

    2021

  31. [40]

    Fear emotion classification in speech by acoustic and behavioral cues

    S. A. Yoon, G. Son & S. Kwon, “Fear emotion classification in speech by acoustic and behavioral cues”. Multimedia Tools and Applications, vol. 78, no. 2, pp. 2345-2366, 2019 . https://doi.org/10.1007/s11042-018-6329-2

    2019 doi

  32. [41]

    Spontaneous Conversational Speech Synthesis from Found Data

    É. Székely, G. E. Henter, J. Beskow, & J. Gustafson, “Spontaneous Conversational Speech Synthesis from Found Data”. In Interspeech pp. 4435-4439, September 2019. doi: 10.21437/Interspeech.2019-2836

    2019 doi

  33. [42]

    A Large-Scale User Study of an Alexa Prize Chatbot: Effect of TTS Dynamism on Perceived Quality of Social Dialog

    M.Cohn, C. Y. Chen, & Z. Yu, “A Large-Scale User Study of an Alexa Prize Chatbot: Effect of TTS Dynamism on Perceived Quality of Social Dialog”. In Proceedings of the 20th Annual SIGdial Meeting on Discourse and Dialogue, pp. 293 -306, September 2019. doi: 10.18653/v1/W19-5935

    2019 doi

  34. [43]

    Speech emotion recognition system with librosa

    Babu, P. Ashok, V. Siva Nagaraju, and Rajeev Ratna Vallabhuni. "Speech emotion recognition system with librosa." 2021 10th IEEE International Conference on Communication Systems and Network Technologies (CSNT). IEEE, 2021

    2021

  35. [44]

    U.S. Patent No. 9,997,172

    D. A. Barreda, J. E Lainez., D.Sharma, , & P. Naylor, “U.S. Patent No. 9,997,172”. Washington, DC: U.S. Patent and Trademark Office. 2018 . https://patents.google.com/patent/US9997172B2/en

    2018

  36. [45]

    librosa: Audio and music signal analysis in python

    McFee, Brian, et al. "librosa: Audio and music signal analysis in python."Proceedings of the 14th python in science conference. Vol. 8. 2015

    2015

  37. [46]

    Guest Editorial: Special Issue on Affective Speech and Language Synthesis, Generation, and Conversion

    Amiriparian, S., Schuller, B. W., Asghar, N., Zen, H., & Burkhardt, F.. "Guest Editorial: Special Issue on Affective Speech and Language Synthesis, Generation, and Conversion." IEEE Transactions on Affective Computing Vol. 14,no. 1, pp. 3-5. 2023

    2023

  38. [47]

    Pysox : Leveraging the audio signal processing power of sox in python

    Bittner, Rachel, Eric Humphrey, and Juan Bello. "Pysox : Leveraging the audio signal processing power of sox in python." Proceedings of the international society for music information retrieval conference late breaking and demo papers. 2016

    2016

  39. [48]

    Speaker identification using a hybrid cnn -mfcc approach

    Ashar, Aweem, Muhammad Shahid Bhatti, and Usama Mushtaq. "Speaker identification using a hybrid cnn -mfcc approach." 2020 International Conference on Emerging Trends in Smart Technologies (ICETST). IEEE, 2020

    2020

  40. [2018]

    https://www.academia.edu/download/56794754/IRJET - V5I3962.pdf

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.