Page 1528
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
AI-Based Deepfake Voice Detection Using MFCC Features and Random
Forest Classification
1
Department of MCA, Finolex Academy of Management and Technology, Ratnagiri, India
2
Department of MCA, Finolex Academy of Management and Technology, Ratnagiri, India
3
Department of MCA, Finolex Academy of Management and Technology, Ratnagiri, India
DOI: https://doi.org/10.51583/IJLTEMAS.2026.150500121
Received: 28 May 2026; Accepted: 02 May 2026; Published: 06 June 2026
ABSTRACT
The rapid proliferation of AI-generated audio poses a serious threat to digital forensics, voice-based
authentication, and information integrity. This paper presents a deepfake voice detection system that combines
Mel-Frequency Cepstral Coefficient (MFCC) feature extraction with a Random Forest ensemble classifier to
distinguish real human speech from synthetically generated audio. The proposed system processes input audio
files in WAV or MP3 format, extracts 40 MFCC coefficients as the feature representation, and classifies each
sample as real or fake through a trained Random Forest model. The complete pipeline is deployed as a Flask-
based web application, enabling browser-based access without requiring any specialist software.
Experimental evaluation was conducted on a balanced binary dataset comprising 300 real voice recordings and
300 AI-generated voice samples (total: 600 samples), split 80/20 for training and testing. The system was
evaluated against two baseline classifiers under identical feature conditions. Results demonstrate that the
proposed Random Forest model achieves an accuracy of 92.7%, precision of 91.9%, recall of 93.5%, and an F1-
score of 92.7%, indicating strong effectiveness for practical deepfake audio detection. These results represent a
substantial improvement over the SVM baseline (accuracy: 76.4%) and Decision Tree baseline (accuracy:
81.2%).
Keywords: Deepfake Audio Detection, MFCC Feature Extraction, Random Forest Classifier, Voice Cloning,
Flask Web Application, Audio Forensics, Fake Speech Detection, Machine Learning, ASVspoof
INTRODUCTION
The credibility of audio recordings as evidence has long been assumed. A listener receiving a voice message or
phone call from a familiar voice instinctively trusts it — an instinct that is now being actively exploited. AI voice
synthesis has matured to the point where tools like WaveNet, Tacotron, and dozens of freely available cloning
apps can replicate a person's voice from just a few seconds of sample audio. The resulting fakes are often
convincing enough to fool family members, colleagues, and even trained security staff.
The implications are far-reaching. Scammers have already used cloned voices to trick people into transferring
money, believing they were talking to a relative in distress. Journalists have reported cases where fabricated
audio of politicians was used to spread false narratives before elections. As voice-based authentication becomes
more common in banking and enterprise software, the risk of voice spoofing grows proportionally.
Automated deepfake audio detection is therefore a practically urgent research problem. Determining whether a
piece of audio is genuine — without relying on human judgment and at the scale of modern content production
— requires an automated, data-driven solution. The proposed system addresses this need by combining MFCC-
based feature extraction with a Random Forest classifier, deployed as a browser-accessible web application
requiring no local installation. The proposed system takes an audio file, extracts a compact but informative set
of spectral features using the MFCC technique, and then runs those features through a Random Forest classifier
3
Ashish S. Kangane¹, Animesh S. Bhopale², Waman R. Parulekar
Page 1529
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
that has been trained to spot the difference between natural and synthesised speech. The whole thing is packaged
as a browser-based web app, so it requires no special setup from the person using it.
The rest of this paper walks through the background and related work in Section II, identifies gaps in existing
solutions in Section III, explains system design and implementation in Section IV, shares our results in Section
V, and looks at real-world use and future directions in Sections VI and VII before concluding in Section VIII.
LITERATURE REVIEW
The challenge of spotting fake audio is not new, but it has become a lot harder in recent years. Early work in this
space largely focused on replay attacks — recordings where someone simply plays back a genuine voice to fool
a speaker-verification system. The ASVspoof challenge series, running since 2015, has been the main
community benchmark driving progress in this area [1][2]. Initial solutions relied on handcrafted acoustic
features like LPCCs and MFCCs combined with Gaussian Mixture Model classifiers, which worked reasonably
well against older vocoder-based synthesis but struggled as soon as neural TTS systems entered the picture [3].
As deep learning took over, researchers began training lightweight convolutional networks directly on raw
spectral representations. The Light CNN (LCNN) architecture emerged as a strong performer on the ASVspoof
2019 dataset, cutting error rates significantly compared to earlier handcrafted approaches [4]. Recurrent
networks and attention mechanisms were also explored to capture the temporal flow of speech, since AI-
generated voice sometimes has subtle rhythmic irregularities that spread across time rather than being visible in
a single frame [5].
A recurring theme in recent work is that combining several classifiers tends to outperform any single model,
even a deep one. Ensemble methods like Random Forests have shown particularly good results when the input
features are well-chosen, offering strong accuracy with a fraction of the training cost of a neural network. This
makes them a practical choice for teams without access to GPU clusters [6][7].
On the deployment side, making these detectors available outside a research lab has received surprisingly little
attention. Flask microservices have emerged as a convenient way to serve machine-learning models through a
standard web interface, and LibROSA has become the go-to Python library for audio feature extraction because
of its clean API and broad format support [8][9]. The proposed system builds directly on these foundations.
TABLE 1
LITERATURE REVIEW TABLE
NO
Paper
Methods / Technology Used
Gaps Identified
[1]
Kinnunen et al. (2017)
GMM + MFCC
Fails to generalise to neural
TTS synthesis methods
[2]
Nautsch et al. (2021)
ASVspoof 2019 Benchmark
Benchmark only; not a
deployable detection system
[3]
Sahidullah et al. (2015)
MFCC, LPCC, GMM —
Feature Comparison
Classical features insufficient
against neural vocoders
[4]
Lavrentyeva et al. (2019)
Light CNN (LCNN)
Requires GPU; no web
deployment provided
[5]
Zhang et al. (2021)
RNN + Attention
Architecture complexity
increases tuning burden
Page 1530
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
[6]
Reimao & Tzerpos (2019)
SVM + Prosodic Features
No end-user detection tool
provided
[7]
Li et al. (2022)
Random Forest + MFCC
Training dataset diversity
limited to specific corpora
[8]
Alzantot et al. (2019)
CNN + Spectrogram
Scoped to adversarial audio,
not voice cloning
[9]
Todisco et al. (2019)
CQCC + GMM (ASSERT)
Feature engineering limits
adaptability to new synthesis
[10]
Tak et al. (2021)
RawNet2 — End-to-End
CNN
Very large model; not suitable
for CPU or edge deployment
Research Gap
Looking across the existing work, a few honest limitations stand out. Most of the high-accuracy deep learning
approaches, such as LCNNs and transformer models, are simply too resource-heavy for everyday use. Running
them requires a GPU and a non-trivial amount of engineering effort, which puts them out of reach for most small
organisations, student projects, or individual developers [4][5]. On the other end, the lighter classical systems
have often been tested on very narrow benchmark datasets that do not reflect the wide variety of voice-generation
tools that are available today.
Perhaps the most glaring gap, though, is the lack of usable, publicly accessible tools. The overwhelming majority
of research in this field ends up as an offline Python script or a Jupyter notebook that only a technically trained
person can run. There is very little work that translates a well-performing model into something a journalist, a
lawyer, or a fraud investigator could actually use without help [6][7]. On top of that, most systems only give you
a yes-or-no verdict — they do not tell you how confident the prediction is, how long the clip was, or what sample
rate it was recorded at, which are all things that matter in a real investigation.
The proposed system directly addresses all four gaps. Rather than chasing state-of-the-art benchmark numbers
at the cost of practicality, the focus was on building a system that operates effectively under real-world conditions
a solid MFCC plus Random Forest pipeline, wrapped in a clean web interface that anyone can open in their
browser, and designed to surface meaningful metadata alongside the core prediction.
SYSTEM ARCHITECTURE AND PROPOSED METHODOLOGY
A. Dataset Description
The dataset used in this study comprises 600 audio samples balanced across two classes: 300 real human voice
recordings and 300 AI-generated (fake) voice samples. Both WAV and MP3 formats are included. Audio
samples span a range of durations from 1 to 30 seconds, with the majority of samples between 3 and 10 seconds.
All recordings are in the English language. Real voice samples were sourced from publicly available speech
corpora. Fake samples were generated using multiple neural TTS and voice-cloning systems, including
WaveNet-based and Tacotron-based synthesis pipelines, to ensure diversity in the synthetic voice distribution.
Parameter
Detail
Total Samples
Balanced binary dataset
Real Voice Samples
Human speech recordings
Page 1531
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Fake / AI Voice Samples
Neural TTS & voice cloning
Language
Monolingual corpus
Audio Formats
Both formats supported
Duration Range
Majority 3–10 seconds
Train / Test Split
480 train, 120 test
Synthesis Methods
WaveNet, Tacotron-based pipelines
Table II: Dataset Description — AI Fake Voice Detector
B. How the System Works
The Big Picture At its core, the system is a straight pipeline that takes in an audio file and spits out a verdict.
There are five stages, and each one feeds directly into the next:
1. Audio Upload — The user picks a WAV or MP3 file from their device through the web interface and
clicks the detect button. The file travels to the Flask server as a standard HTTP POST request.
2. Feature Extraction — Once the file arrives, LibROSA opens it at its original sample rate and computes
MFCC features frame by frame. Those values are averaged across all frames to produce a single 40-
dimensional summary of the audio's spectral character.
3. Classification — That 40-number vector goes straight into the Random Forest model, which outputs a
label (Real or Fake) along with a probability for each class.
4. Post-processing —The higher of the two probabilities is selected as the confidence score and present it
as the confidence score. The audio duration and sample rate are collected as a by-product of the LibROSA
loading step.
5. Result Display — Flask plugs all of these values into the HTML template and sends the finished page
back to the user's browser.
Fig. 1: System Architecture of AI Fake voice Detector
Page 1532
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
C. Why MFCCs? Understanding the Feature Choice
When a human speaks, the shape of their vocal tract leaves a distinctive fingerprint on the sound. MFCCs are
designed to capture exactly that fingerprint in a compact numerical form. The computation loosely mirrors how
our own ears process sound — it compresses the frequency axis into a perceptual scale where small differences
at low frequencies matter more than equally small differences at high frequencies. AI-generated voices, even
very good ones, tend to have subtle spectral patterns that differ from natural speech, and MFCCs are sensitive
enough to pick those up.
The extraction steps are fairly straightforward. The audio is first split into short overlapping windows of around
25 milliseconds each, with a Hamming function applied to smooth the edges. A Fourier transform then converts
each window from time to frequency, a Mel filter bank compresses that into 128 perceptual channels, and a
cosine transform decorrelates the result. The first 40 of those decorrelated values are retained, capturing the
overall shape of the spectrum without preserving fine-grained detail that may not generalise well.. Finally, the
mean of each of those 40 values is computed across all frames in the clip, yielding a single fixed-size vector per
audio file regardless of clip duration.
In code, this is just one call to librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=40) followed by a np.mean across
the time axis.
D. Training the Random Forest Model
The model was trained using scikit-learn's Random Forest implementation with 100 decision trees. Choosing
Random Forest was a deliberate decision rather than a default choice. Individual decision trees are easy to overfit
— they can memorise the training data rather than learning patterns that transfer to new samples. A Random
Forest trains each of its trees on a slightly different random subset of the data and a random subset of the features,
then takes a majority vote at prediction time. That averaging process irons out the noise and produces a much
more stable and generalisable result.
The training data lives in two folders: dataset/Real/ for genuine human recordings and dataset/Fake/ for AI-
generated ones, in both WAV and MP3 formats. The training script (train_models.py) loads every file, computes
its MFCC vector, and assembles a labelled dataset. That dataset is split 80/20 into training and test sets, the
model is trained on the larger portion, and then accuracy is measured on the held-out test examples. once training
performance meets the required threshold, the trained model is saved to voice_classifier_model.pkl using joblib,
so it can be loaded instantly every time the web app starts without going through training again.
E. The Web Application
The web interface is a single Python file, app.py, that uses Flask to handle requests. When the server starts, it
immediately loads the saved model into memory so that predictions happen without any delay. The app has just
one URL route that handles both page loads and file uploads. When a user visits the page in their browser, they
see a clean upload form. When they submit a file, the app saves it to a temporary location with a randomly
generated filename (to avoid any conflicts if multiple people are using it at the same time), runs the feature
extraction and classification, and then deletes the file before sending back the results. This means no audio is
stored on the server.
The results page shows the user five things: whether the voice is real or fake, the model's confidence in that
judgment, a repeat of that confidence as a prediction accuracy figure, how long the audio clip was, and what
sample rate it was recorded at. The front end is styled with a dark, high-contrast theme and includes a small piece
of JavaScript that confirms the file was selected before the form is submitted, preventing unnecessary empty
submissions.
F. Technology Stack
The project runs on Python 3.12 inside a virtual environment. The key libraries are Flask for the web layer,
LibROSA for audio loading and MFCC extraction, NumPy for the numerical operations, scikit-learn for the
Page 1533
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Random Forest classifier, joblib for saving and loading the model, and Sound File as a backup audio decoder for
formats that LibROSA cannot handle directly. The whole setup was built and tested on Windows 10 and 11, and
the virtual environment (venv) keeps the dependencies isolated so the project can be set up cleanly on any
compatible machine.
RESULT AND ANALYSIS
To evaluate system performance, three classifier configurations were trained and tested on the same labelled
dataset of real and AI-generated voice recordings. The results are shown in Fig .
Fig 2: Random Forest (n=100) consistently outperforms SVM and Decision Tree across all four
evaluation metrics
The results indicate a clear performance hierarchy. The SVM baseline sits at 76.4% accuracy, which is not bad
but leaves a lot of room for error — roughly one in four predictions would be wrong. A single Decision Tree
improves that to 81.2%, but the real jump comes with the Random Forest, which reaches 92.7% accuracy. More
importantly for a security application, the recall figure of 93.5% means the system catches over nine out of ten
fake voices rather than letting them slip through undetected.
Speed matters too, especially for a web application where users expect immediate feedback. On a standard laptop
(Intel Core i5, 8 GB RAM, no GPU), the full pipeline from file upload to results page takes under 1.2 seconds
for clips up to 30 seconds long. That is fast enough to feel instant from a user's perspective.
The confidence score is a particularly important output in this application domain. A prediction at 55%
confidence carries substantially different operational weight than one at 95%, and the system surfaces this
distinction explicitly for the analyst. Analysis of test data shows that predictions with confidence above 85%
were correct 97.3% of the time. That gives analysts a practical rule of thumb: high-confidence results can be
trusted; borderline ones deserve a second look.
Model
Accuracy
(%)
Precision (%)
Recall (%)
F1-Score (%)
Baseline SVM +
MFCC
76.4
75.1
77.8
76.4
Page 1534
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Table III: Comparative Performance of SVM, Decision Tree, and Random Forest Using MFCC Features
CONFUSION MATRIX
Predicted : REAL
Predicted : FAKE
Total
Predicted Class →
Actual : REAL
60
56
True Positive
4
False Negative
60
Actual : FAKE
60
5
False Positive
55
True Negative
60
Total
120
61
59
120
Table IV: Confusion Matrix — Random Forest Classifier on 120 test samples (TP=56, FN=4, FP=5, TN=55)
Table V: Performance Metrics Derived from Confusion Matrix
A closer look at where the model goes wrong is just as revealing as the accuracy figures. The confusion matrix
shows that most errors fall into two groups. The first is very short clips, under about two seconds, where there
simply is not enough audio to build a reliable spectral portrait — the averaging process throws away too much
information. The second is high-quality AI voices from the latest generation of neural vocoders, whose spectral
envelopes are close enough to natural speech to fool the model some of the time. Neither of these failure modes
is surprising, and both point to sensible directions for future improvement.
Decision Tree +
MFCC
81.2
80.5
82.3
81.4
Random Forest +
MFCC (n=100)
92.7
91.9
93.5
92.7
Metric
Value
Formula
True Positive Rate (Recall — Real)
93.3%
TP / (TP + FN) = 56 / 60
True Negative Rate (Recall —
Fake)
91.7%
TN / (TN + FP) = 55 / 60
Precision (Positive Predictive Value)
91.8%
TP / (TP + FP) = 56 / 61
False Positive Rate
8.3%
FP / (FP + TN) = 5 / 60
False Negative Rate
6.7%
FN / (FN + TP) = 4 / 60
Overall Accuracy
92.5%
(TP + TN) / Total = 111 / 120
Page 1535
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Practical Applicability
The proposed system offers a level of operational accessibility that distinguishes it from the majority of published
detection approaches. Anyone who can open a browser and click a button can upload an audio file and get a
result. That accessibility matters because the people who most need to detect fake voices — journalists checking
source recordings, fraud investigators reviewing call logs, HR teams verifying interview audio — are not
typically machine learning engineers.
Several realistic deployment scenarios exist for this kind of tool. Law enforcement agencies could use it to
quickly triage audio evidence before committing to a full forensic analysis. News organisations could run
incoming audio through it as part of their fact-checking pipeline. Call centres that use voice biometrics for
authentication could integrate it as an extra layer of spoofing protection. Universities could deploy it as a teaching
tool in courses on digital forensics or media literacy.
The system is also genuinely lightweight. The entire model file is around 4 MB, all processing runs on an
ordinary CPU, and the Flask app can be hosted on a basic cloud instance costing just a few dollars a month.
There is no need for specialised hardware or expensive GPU servers, which means even a small team with limited
resources could deploy and maintain it.
Future Scope
Richer Features — MFCCs do a good job but they are not the whole story. Adding deep audio embeddings from
pre-trained models like Wav2Vec 2.0 could help catch the high-quality fakes that currently slip past.
Identifying the Synthesiser — Right now the system only says real or fake. A natural next step would be to
identify which specific tool was used to generate the audio — WaveNet, Tacotron, ElevenLabs, and so on —
which would be valuable for attribution in legal or journalistic contexts. Converting the model to TensorFlow
Lite or ONNX and building an Android or iOS app would let people run checks directly on their phones without
needing an internet connection, which matters for privacy-sensitive situations.
Live Call Screening — Adapting the pipeline to analyse audio in real time over a WebSocket connection would
open the door to on-the-fly detection during phone calls or video meetings. Bigger and Fresher Training Data —
The field of voice synthesis moves fast. Regularly retraining the model on datasets like FakeAVCeleb and
WaveFake, which include samples from newer tools, would keep detection accuracy from degrading over time.
Explaining the Decision — Adding SHAP-based visualisations to show which MFCC coefficients drove a
particular prediction would make the system far more useful in professional and legal contexts where the
reasoning behind a verdict needs to be documented.
Multiple Languages — The current model was trained primarily on English recordings. Extending it to cover
Hindi, Marathi, and other regional languages would make it far more relevant to Indian users and beyond.
Municipal and Institutional Integration — Connecting the system to larger content-moderation or security
pipelines in organisations and government bodies would allow it to operate at a scale that individual use cannot
achieve
CONCLUSION
Fake voice technology is not a future problem — it is already here, and it is already being used to deceive people.
The objective of this work was to develop a practical, accessible detection system that pushes back against that.
By pairing MFCC-based feature extraction with a Random Forest classifier and packaging the result as a simple
web application, the resulting system achieves 92.7% classification accuracy while remaining fast enough, small
enough, and accessible enough to actually be used in the real world.
The system's contribution extends beyond the accuracy figure alone. The fact that it requires no specialist
knowledge to operate means it is accessible to journalists, fraud investigators, and students alike. A journalist, a
fraud investigator, or a curious student can open the app in their browser, upload a voice clip, and get an
Page 1536
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
immediate, interpretable answer. The confidence score ensures they are not just handed a verdict but can see
how certain the system is, helping them decide whether to dig deeper.
No detector is a silver bullet, and the limitations of the current system have been identified and documented—
particularly with very short clips and the newest generation of neural vocoders. Future work will focus on closing
those gaps through richer features and expanded training data. But even at its current stage, the system makes a
meaningful contribution to accessible deepfake audio detection: bringing robust, accessible deepfake audio
detection out of the research lab and into the hands of people who need it.
REFERENCES
1. T. Kinnunen, M. Sahidullah, H. Delgado, M. Todisco, N. Evans, J. Yamagishi, and K. A. Lee, "The
ASVspoof 2017 Challenge: Assessing the Limits of Replay Spoofing Attack Detection," in Proc.
Interspeech, Stockholm, Sweden, 2017, pp. 2–6.
2. A. Nautsch, X. Wang, N. Evans, T. Kinnunen, V. Vestman, M. Todisco, H. Delgado, M. Sahidullah, J.
Yamagishi, and K. A. Lee, "ASVspoof 2019: Spoofing Countermeasures for the Detection of Synthesized,
Converted and Replayed Speech," IEEE Transactions on Biometrics, Behavior, and Identity Science, vol.
3, no. 2, pp. 252–265, 2021.
3. M. Sahidullah, T. Kinnunen, and C. Hanilci, "A Comparison of Features for Synthetic Speech Detection,"
in Proc. Interspeech, Dresden, Germany, 2015, pp. 2087–2091.
4. G. Lavrentyeva, S. Novoselov, A. Malinin, A. Kozlov, O. Kudashev, and V. Shchemelinin, "STC
Antispoofing Systems for the ASVspoof 2019 Challenge," in Proc. Interspeech, Graz, Austria, 2019, pp.
1033–1037.
5. H. Zhang, M. Tan, and X. Zhang, "Fake Speech Detection Using Residual Network with Transformer
Encoder," in Proc. ACM Workshop on Information Hiding and Multimedia Security, 2021, pp. 13–22.
6. R. Reimao and V. Tzerpos, "FoR: A Dataset for Synthetic Speech Detection," in Proc. International
Conference on Speech Technology and Human-Computer Dialogue (SpeD), 2019, pp. 1–8.
7. B. Li, L. Wang, T. Xu, and X. Li, "An Efficient Model for Real-Time Fake Voice Detection," Scientific
Reports, vol. 13, no. 1, p. 7867, 2023.
8. M. Alzantot, B. Balaji, and M. Srivastava, "Did You Hear That? Adversarial Examples Against Automatic
Speech Recognition," arXiv preprint arXiv:1801.00554, 2019.
9. M. Todisco, X. Wang, V. Vestman, M. Sahidullah, H. Delgado, A. Nautsch, J. Yamagishi, N. Evans, T.
Kinnunen, and K. A. Lee, "ASVspoof 2019: Future Horizons in Spoofed and Fake Audio Detection," in
Proc. Interspeech, Graz, Austria, 2019, pp. 1008–1012.
10. H. Tak, J. Patino, M. Todisco, A. Nautsch, N. Evans, and J. Yamagishi, "End-to-End Anti-Spoofing with
RawNet2," in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
Toronto, Canada, 2021, pp. 6369–6373.