Artificial Intelligence (AI) Training Dataset Market Outlook

According to our latest research, the global Artificial Intelligence (AI) Training Dataset market size reached USD 3.15 billion in 2024, reflecting robust industry momentum. The market is expanding at a notable CAGR of 20.8% and is forecasted to attain USD 20.92 billion by 2033. This impressive growth is primarily attributed to the surging demand for high-quality, annotated datasets to fuel machine learning and deep learning models across diverse industry verticals. The proliferation of AI-driven applications, coupled with rapid advancements in data labeling technologies, is further accelerating the adoption and expansion of the AI training dataset market globally.

One of the most significant growth factors propelling the AI training dataset market is the exponential rise in data-driven AI applications across industries such as healthcare, automotive, retail, and finance. As organizations increasingly rely on AI-powered solutions for automation, predictive analytics, and personalized customer experiences, the need for large, diverse, and accurately labeled datasets has become critical. Enhanced data annotation techniques, including manual, semi-automated, and fully automated methods, are enabling organizations to generate high-quality datasets at scale, which is essential for training sophisticated AI models. The integration of AI in edge devices, smart sensors, and IoT platforms is further amplifying the demand for specialized datasets tailored for unique use cases, thereby fueling market growth.

Another key driver is the ongoing innovation in machine learning and deep learning algorithms, which require vast and varied training data to achieve optimal performance. The increasing complexity of AI models, especially in areas such as computer vision, natural language processing, and autonomous systems, necessitates the availability of comprehensive datasets that accurately represent real-world scenarios. Companies are investing heavily in data collection, annotation, and curation services to ensure their AI solutions can generalize effectively and deliver reliable outcomes. Additionally, the rise of synthetic data generation and data augmentation techniques is helping address challenges related to data scarcity, privacy, and bias, further supporting the expansion of the AI training dataset market.

The market is also benefiting from the growing emphasis on ethical AI and regulatory compliance, particularly in data-sensitive sectors like healthcare, finance, and government. Organizations are prioritizing the use of high-quality, unbiased, and diverse datasets to mitigate algorithmic bias and ensure transparency in AI decision-making processes. This focus on responsible AI development is driving demand for curated datasets that adhere to strict quality and privacy standards. Moreover, the emergence of data marketplaces and collaborative data-sharing initiatives is making it easier for organizations to access and exchange valuable training data, fostering innovation and accelerating AI adoption across multiple domains.

From a regional perspective, North America currently dominates the AI training dataset market, accounting for the largest revenue share in 2024, driven by significant investments in AI research, a mature technology ecosystem, and the presence of leading AI companies and data annotation service providers. Europe and Asia Pacific are also witnessing rapid growth, with increasing government support for AI initiatives, expanding digital infrastructure, and a rising number of AI startups. While North America sets the pace in terms of technological innovation, Asia Pacific is expected to exhibit the highest CAGR during the forecast period, fueled by the digital transformation of emerging economies and the proliferation of AI applications across various industry sectors.

Data Type Analysis

The AI training dataset market is segmented by data type into Text, Image/Video, Audio, and Others, each playing a crucial role in powering different AI applications. Text da

• Version 4: Adding the data from "LLM-generated essay using PaLM from Google Gen-AI" kindly generated by Kingki19 / Muhammad Rizqi.
  File: train_essays_RDizzl3_seven_v2.csv
  Human texts: 14247 LLM texts: 3004
  
  See also: a new dataset of an additional 4900 LLM generated texts: LLM: Mistral-7B Instruct texts

  
  
  

• Version 3: "**The RDizzl3 Seven**"
  File: train_essays_RDizzl3_seven_v1.csv

• "Car-free cities"

• "Does the electoral college work?"

• "Exploring Venus"

• "The Face on Mars"

• "Facial action coding system"

• "A Cowboy Who Rode the Waves"

• "Driverless cars"

How this dataset was made: see the notebook "LLM: Make 7 prompt train dataset"

• Version 2: (train_essays_7_prompts_v2.csv) This dataset is composed of 13,712 human texts and 1638 AI-LLM generated texts originating from 7 of the PERSUADE 2.0 corpus prompts.

Namely:

• "Car-free cities"
• "Does the electoral college work?"
• "Exploring Venus"
• "The Face on Mars"
• "Facial action coding system"
• "Seeking multiple opinions"
• "Phones and driving"

This dataset is a derivative of the datasets

• LLM Generated Essays for the Detect AI Comp! by Radek Osmulski
• persuade corpus 2.0 provided by Nicholas Broad
• daigt data - llama 70b and falcon180b by Nicholas Broad
• Hello, Claude! 1000 essays from Anthropic... by Darragh

as well as the original competition training dataset

• Version 1:This dataset is composed of 13,712 human texts and 1165 AI-LLM generated texts originating from 7 of the PERSUADE 2.0 corpus prompts.

This is an open source - publicly available dataset which can be found at https://shahariarrabby.github.io/ekush/ . We split the dataset into three sets - train, validation, and test. For our experiments, we created two other versions of the dataset. We have applied 10-fold cross validation on the train set and created ten folds. We also created ten bags of datasets using bootstrap aggregating method on the train and validation sets. Lastly, we created another dataset using pre-trained ResNet50 model as feature extractor. On the features extracted by ResNet50 we have applied PCA and created a tabilar dataset containing 80 features. pca_features.csv is the train set and pca_test_features.csv is the test set. Fold.tar.gz contains the ten folds of images described above. Those folds are also been compressed. Similarly, Bagging.tar.gz contains the ten compressed bags of images. The original train, validation, and test sets are in Train.tar.gz, Validation.tar.gz, and Test.tar.gz, respectively. The compression has been performed for speeding up the upload and download purpose and mostly for the sake of convenience. If anyone has any question about how the datasets are organized please feel free to ask me at shiblygnr@gmail.com .I will get back to you in earliest time possible.

In this Dataset contains both AI Generated Essay and Human Written Essay for Training Purpose This dataset challenge is to to develop a machine learning model that can accurately detect whether an essay was written by a student or an LLM. The competition dataset comprises a mix of student-written essays and essays generated by a variety of LLMs.

Dataset contains more than 28,000 essay written by student and AI generated.

Features : 1. text : Which contains essay text 2. generated : This is target label . 0 - Human Written Essay , 1 - AI Generated Essay

Software Model simulations were conducted using WRF version 3.8.1 (available at https://github.com/NCAR/WRFV3) and CMAQ version 5.2.1 (available at https://github.com/USEPA/CMAQ). The meteorological and concentration fields created using these models are too large to archive on ScienceHub, approximately 1 TB, and are archived on EPA’s high performance computing archival system (ASM) at /asm/MOD3APP/pcc/02.NOAH.v.CLM.v.PX/. Figures Figures 1 – 6 and Figure 8: Created using the NCAR Command Language (NCL) scripts (https://www.ncl.ucar.edu/get_started.shtml). NCLD code can be downloaded from the NCAR website (https://www.ncl.ucar.edu/Download/) at no cost. The data used for these figures are archived on EPA’s ASM system and are available upon request. Figures 7, 8b-c, 8e-f, 8h-i, and 9 were created using the AMET utility developed by U.S. EPA/ORD. AMET can be freely downloaded and used at https://github.com/USEPA/AMET. The modeled data paired in space and time provided in this archive can be used to recreate these figures. The data contained in the compressed zip files are organized in comma delimited files with descriptive headers or space delimited files that match tabular data in the manuscript. The data dictionary provides additional information about the files and their contents. This dataset is associated with the following publication: Campbell, P., J. Bash, and T. Spero. Updates to the Noah Land Surface Model in WRF‐CMAQ to Improve Simulated Meteorology, Air Quality, and Deposition. Journal of Advances in Modeling Earth Systems. John Wiley & Sons, Inc., Hoboken, NJ, USA, 11(1): 231-256, (2019).

Explore and download labeled image datasets for AI, ML, and computer vision. Find datasets for object detection, image classification, and image segmentation.

Around 500K essays are available in this dataset, both created by AI and written by Human.

I have gathered the data from multiple sources, added them together and removed the duplicates

1. Specifications Data size : 60,000 ID

Population distribution : the race distribution is Asians, Caucasians and black people, the gender distribution is male and female, the age distribution is from children to the elderly

Collecting environment : including indoor and outdoor scenes (such as supermarket, mall and residential area, etc.)

Data diversity : different ages, different time periods, different cameras, different human body orientations and postures, different ages collecting environment

Device : surveillance cameras, the image resolution is not less than 1,9201,080

Data format : the image data format is .jpg, the annotation file format is .json

Annotation content : human body rectangular bounding boxes, 15 human body attributes

Quality Requirements : A rectangular bounding box of human body is qualified when the deviation is not more than 3 pixels, and the qualified rate of the bounding boxes shall not be lower than 97%;Annotation accuracy of attributes is over 97%

1. About Nexdata Nexdata owns off-the-shelf PB-level Large Language Model(LLM) Data, 1 million hours of Audio Data and 800TB of Annotated Imagery Data.These ready-to-go Identity Data support instant delivery, quickly improve the accuracy of AI models. For more details, please visit us at https://www.nexdata.ai/datasets/computervision?source=Datarade

Snapshot img

US Deep Learning Market Size 2025-2029

The deep learning market size in US is forecast to increase by USD 5.02 billion at a CAGR of 30.1% between 2024 and 2029.

The deep learning market is experiencing robust growth, driven by the increasing adoption of artificial intelligence (AI) in various industries for advanced solutioning. This trend is fueled by the availability of vast amounts of data, which is a key requirement for deep learning algorithms to function effectively. Industry-specific solutions are gaining traction, as businesses seek to leverage deep learning for specific use cases such as image and speech recognition, fraud detection, and predictive maintenance. Alongside, intuitive data visualization tools are simplifying complex neural network outputs, helping stakeholders understand and validate insights. 


However, challenges remain, including the need for powerful computing resources, data privacy concerns, and the high cost of implementing and maintaining deep learning systems. Despite these hurdles, the market's potential for innovation and disruption is immense, making it an exciting space for businesses to explore further. Semi-supervised learning, data labeling, and data cleaning facilitate efficient training of deep learning models. Cloud analytics is another significant trend, as companies seek to leverage cloud computing for cost savings and scalability. 

What will be the Size of the market During the Forecast Period?

Request Free Sample

Deep learning, a subset of machine learning, continues to shape industries by enabling advanced applications such as image and speech recognition, text generation, and pattern recognition. Reinforcement learning, a type of deep learning, gains traction, with deep reinforcement learning leading the charge. Anomaly detection, a crucial application of unsupervised learning, safeguards systems against security vulnerabilities. Ethical implications and fairness considerations are increasingly important in deep learning, with emphasis on explainable AI and model interpretability. Graph neural networks and attention mechanisms enhance data preprocessing for sequential data modeling and object detection. Time series forecasting and dataset creation further expand deep learning's reach, while privacy preservation and bias mitigation ensure responsible use.

In summary, deep learning's market dynamics reflect a constant pursuit of innovation, efficiency, and ethical considerations. The Deep Learning Market in the US is flourishing as organizations embrace intelligent systems powered by supervised learning and emerging self-supervised learning techniques. These methods refine predictive capabilities and reduce reliance on labeled data, boosting scalability. BFSI firms utilize AI image recognition for various applications, including personalizing customer communication, maintaining a competitive edge, and automating repetitive tasks to boost productivity. Sophisticated feature extraction algorithms now enable models to isolate patterns with high precision, particularly in applications such as image classification for healthcare, security, and retail.

How is this market segmented and which is the largest segment?

The market research report provides comprehensive data (region-wise segment analysis), with forecasts and estimates in 'USD million' for the period 2025-2029, as well as historical data from 2019-2023 for the following segments.

Application

  Image recognition
  Voice recognition
  Video surveillance and diagnostics
  Data mining


Type

  Software
  Services
  Hardware


End-user

  Security
  Automotive
  Healthcare
  Retail and commerce
  Others


Geography

  North America

    US

By Application Insights

The Image recognition segment is estimated to witness significant growth during the forecast period. In the realm of artificial intelligence (AI) and machine learning, image recognition, a subset of computer vision, is gaining significant traction. This technology utilizes neural networks, deep learning models, and various machine learning algorithms to decipher visual data from images and videos. Image recognition is instrumental in numerous applications, including visual search, product recommendations, and inventory management. Consumers can take photographs of products to discover similar items, enhancing the online shopping experience. In the automotive sector, image recognition is indispensable for advanced driver assistance systems (ADAS) and autonomous vehicles, enabling the identification of pedestrians, other vehicles, road signs, and lane markings.

Furthermore, image recognition plays a pivotal role in augmented reality (AR) and virtual reality (VR) applications, where it tracks physical objects and overlays digital content onto real-world scenarios. The model training process involves the backpropagation algorithm, which calculates

This dataset is part of the following publication at the TransAI 2023 conference: R. Wallsberger, R. Knauer, S. Matzka; "Explainable Artificial Intelligence in Mechanical Engineering: A Synthetic Dataset for Comprehensive Failure Mode Analysis" DOI: http://dx.doi.org/10.1109/TransAI60598.2023.00032

This is the original XAI Drilling dataset optimized for XAI purposes and it can be used to evaluate explanations of such algortihms. The dataset comprises 20,000 data points, i.e., drilling operations, stored as rows, 10 features, one binary main failure label, and 4 binary subgroup failure modes, stored in columns. The main failure rate is about 5.0 % for the whole dataset. The features that constitute this dataset are as follows:

• ID: Every data point in the dataset is uniquely identifiable, thanks to the ID feature. This ensures traceability and easy referencing, especially when analyzing specific drilling scenarios or anomalies.
• Cutting speed vc (m/min): The cutting speed is a pivotal parameter in drilling, influencing the efficiency and quality of the drilling process. It represents the speed at which the drill bit's cutting edge moves through the material.
• Spindle speed n (1/min): This feature captures the rotational speed of the spindle or drill bit, respectively.
• Feed f (mm/rev): Feed denotes the depth the drill bit penetrates into the material with each revolution. There is a balance between speed and precision, with higher feeds leading to faster drilling but potentially compromising hole quality.
• Feed rate vf (mm/min): The feed rate is a measure of how quickly the material is fed to the drill bit. It is a determinant of the overall drilling time and influences the heat generated during the process.
• Power Pc (kW): The power consumption during drilling can be indicative of the efficiency of the process and the wear state of the drill bit.
• Cooling (%): Effective cooling is paramount in drilling, preventing overheating and reducing wear. This ordinal feature captures the cooling level applied, with four distinct states representing no cooling (0%), partial cooling (25% and 50%), and high to full cooling (75% and 100%).
• Material: The type of material being drilled can significantly influence the drilling parameters and outcomes. This dataset encompasses three primary materials: C45K hot-rolled heat-treatable steel (EN 1.0503), cast iron GJL (EN GJL-250), and aluminum-silicon (AlSi) alloy (EN AC-42000), each presenting its unique challenges and considerations. The three materials are represented as “P (Steel)” for C45K, “K (Cast Iron)” for cast iron GJL and “N (Non-ferrous metal)” for AlSi alloy.
• Drill Bit Type: Different materials often require specialized drill bits. This feature categorizes the type of drill bit used, ensuring compatibility with the material and optimizing the drilling process. It consists of three categories, which are based on the DIN 1836: “N” for C45K, “H” for cast iron and “W” for AlSi alloy [5].

• Process time t (s): This feature captures the full duration of each drilling operation, providing insights into efficiency and potential bottlenecks.

• Main failure: This binary feature indicates if any significant failure on the drill bit occurred during the drilling process. A value of 1 flags a drilling process that encountered issues, which in this case is true when any of the subgroup failure modes are 1, while 0 indicates a successful drilling operation without any major failures.

Subgroup failures: - Build-up edge failure (215x): Represented as a binary feature, a build-up edge failure indicates the occurrence of material accumulation on the cutting edge of the drill bit due to a combination of low cutting speeds and insufficient cooling. A value of 1 signifies the presence of this failure mode, while 0 denotes its absence. - Compression chips failure (344x): This binary feature captures the formation of compressed chips during drilling, resulting from the factors high feed rate, inadequate cooling and using an incompatible drill bit. A value of 1 indicates the occurrence of at least two of the three factors above, while 0 suggests a smooth drilling operation without compression chips. - Flank wear failure (278x): A binary feature representing the wear of the drill bit's flank due to a combination of high feed rates and low cutting speeds. A value of 1 indicates significant flank wear, affecting the drilling operation's accuracy and efficiency, while 0 denotes a wear-free operation. - Wrong drill bit failure (300x): As a binary feature, it indicates the use of an inappropriate drill bit for the material being drilled. A value of 1 signifies a mismatch, leading to potential drilling issues, while 0 indicates the correct drill bit usage.

The dataset of the most popular text-to-image prompts.

  Dataset Details





  Dataset Description

Curated by: kazimir.ai Funded by [optional]: [More Information Needed] Shared by [optional]: https://kazimir.ai License: apache-2.0

  Dataset Sources [optional]

Repository: [More Information Needed] Paper [optional]: [More Information Needed] Demo [optional]: [More Information Needed]

  Uses

Free to use.

  Dataset Structure

CSV file… See the full description on the dataset page: https://huggingface.co/datasets/Kazimir-ai/text-to-image-prompts.

1. Specifications Data size : 200,000 ID

Population distribution : race distribution: Asians, Caucasians, black people; gender distribution: gender balance; age distribution: from child to the elderly, the young people and the middle aged are the majorities

Collection environment : indoor scenes, outdoor scenes

Collection diversity : various postures, expressions, light condition, scenes, time periods and distances

Collection device : iPhone, android phone, iPad

Collection time : daytime,night

Image Parameter : the video format is .mov or .mp4, the image format is .jpg

Accuracy : the accuracy of actions exceeds 97%

1. About Nexdata Nexdata owns off-the-shelf PB-level Large Language Model(LLM) Data, 1 million hours of Audio Data and 800TB of Annotated Imagery Data. These ready-to-go machine learning (ML) data supports instant delivery, quickly improve the accuracy of AI models. For more details, please visit us at https://www.nexdata.ai/datasets/computervision?source=Datarade

A large-scale hierarchical dataset of diverse student activities collected by Santa, a multi-platform self-study solution equipped with artificial intelligence tutoring system. EdNet contains 131,441,538 interactions from 784,309 students collected over more than 2 years, which is the largest among the ITS datasets released to the public so far.

Dataset Card for Analytics Vidhya Courses Dataset

This dataset contains details of online courses offered by Analytics Vidhya, including course titles, descriptions, number of lessons, and access links. The dataset is useful for building educational search and recommendation systems.

  Dataset Description

The Analytics Vidhya Courses Dataset consists of various online courses related to data science, machine learning, and AI, among other technical fields. The… See the full description on the dataset page: https://huggingface.co/datasets/kalki-ai/free-course.

This dataset consists of reviews collected from restaurants on a Korean delivery app platform running a review event. A total of 128,668 reviews were collected from 136 restaurants by crawling reviews using the Selenium library in Python. The dataset named as Korean Reviews.csv provides review data not translated to English, and the dataset named as English Reviews.csv provides review data translated to English. The 136 chosen restaurants run review events which demand customers to write reviews with 5 stars and photos. So the annotation of data was done by considering 1) whether the review gives five-star ratings, and 2) whether the review contains photo(s).

Data Cleaning Tools Market Outlook

As of 2023, the global market size for data cleaning tools is estimated at $2.5 billion, with projections indicating that it will reach approximately $7.1 billion by 2032, reflecting a robust CAGR of 12.1% during the forecast period. This growth is primarily driven by the increasing importance of data quality in business intelligence and analytics workflows across various industries.

The growth of the data cleaning tools market can be attributed to several critical factors. Firstly, the exponential increase in data generation across industries necessitates efficient tools to manage data quality. Poor data quality can result in significant financial losses, inefficient business processes, and faulty decision-making. Organizations recognize the value of clean, accurate data in driving business insights and operational efficiency, thereby propelling the adoption of data cleaning tools. Additionally, regulatory requirements and compliance standards also push companies to maintain high data quality standards, further driving market growth.

Another significant growth factor is the rising adoption of AI and machine learning technologies. These advanced technologies rely heavily on high-quality data to deliver accurate results. Data cleaning tools play a crucial role in preparing datasets for AI and machine learning models, ensuring that the data is free from errors, inconsistencies, and redundancies. This surge in the use of AI and machine learning across various sectors like healthcare, finance, and retail is driving the demand for efficient data cleaning solutions.

The proliferation of big data analytics is another critical factor contributing to market growth. Big data analytics enables organizations to uncover hidden patterns, correlations, and insights from large datasets. However, the effectiveness of big data analytics is contingent upon the quality of the data being analyzed. Data cleaning tools help in sanitizing large datasets, making them suitable for analysis and thus enhancing the accuracy and reliability of analytics outcomes. This trend is expected to continue, fueling the demand for data cleaning tools.

In terms of regional growth, North America holds a dominant position in the data cleaning tools market. The region's strong technological infrastructure, coupled with the presence of major market players and a high adoption rate of advanced data management solutions, contributes to its leadership. However, the Asia Pacific region is anticipated to witness the highest growth rate during the forecast period. The rapid digitization of businesses, increasing investments in IT infrastructure, and a growing focus on data-driven decision-making are key factors driving the market in this region.

As organizations strive to maintain high data quality standards, the role of an Email List Cleaning Service becomes increasingly vital. These services ensure that email databases are free from invalid addresses, duplicates, and outdated information, thereby enhancing the effectiveness of marketing campaigns and communications. By leveraging sophisticated algorithms and validation techniques, email list cleaning services help businesses improve their email deliverability rates and reduce the risk of being flagged as spam. This not only optimizes marketing efforts but also protects the reputation of the sender. As a result, the demand for such services is expected to grow alongside the broader data cleaning tools market, as companies recognize the importance of maintaining clean and accurate contact lists.

Component Analysis

The data cleaning tools market can be segmented by component into software and services. The software segment encompasses various tools and platforms designed for data cleaning, while the services segment includes consultancy, implementation, and maintenance services provided by vendors.

The software segment holds the largest market share and is expected to continue leading during the forecast period. This dominance can be attributed to the increasing adoption of automated data cleaning solutions that offer high efficiency and accuracy. These software solutions are equipped with advanced algorithms and functionalities that can handle large volumes of data, identify errors, and correct them without manual intervention. The rising adoption of cloud-based data cleaning software further bolsters this segment, as it offers scalability and ease of

The quality of AI-generated images has rapidly increased, leading to concerns of authenticity and trustworthiness.

CIFAKE is a dataset that contains 60,000 synthetically-generated images and 60,000 real images (collected from CIFAR-10). Can computer vision techniques be used to detect when an image is real or has been generated by AI?

Dataset details The dataset contains two classes - REAL and FAKE. For REAL, we collected the images from Krizhevsky & Hinton's CIFAR-10 dataset For the FAKE images, we generated the equivalent of CIFAR-10 with Stable Diffusion version 1.4 There are 100,000 images for training (50k per class) and 20,000 for testing (10k per class)

References If you use this dataset, you must cite the following sources

Krizhevsky, A., & Hinton, G. (2009). Learning multiple layers of features from tiny images.

Bird, J.J., Lotfi, A. (2023). CIFAKE: Image Classification and Explainable Identification of AI-Generated Synthetic Images. arXiv preprint arXiv:2303.14126.

Real images are from Krizhevsky & Hinton (2009), fake images are from Bird & Lotfi (2023). The Bird & Lotfi study is a preprint currently available on ArXiv and this description will be updated when the paper is published.

License This dataset is published under the same MIT license as CIFAR-10:

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

DEEP-VOICE: Real-time Detection of AI-Generated Speech for DeepFake Voice Conversion This dataset contains examples of real human speech, and DeepFake versions of those speeches by using Retrieval-based Voice Conversion.

Can machine learning be used to detect when speech is AI-generated?

Introduction There are growing implications surrounding generative AI in the speech domain that enable voice cloning and real-time voice conversion from one individual to another. This technology poses a significant ethical threat and could lead to breaches of privacy and misrepresentation, thus there is an urgent need for real-time detection of AI-generated speech for DeepFake Voice Conversion.

To address the above emerging issues, we are introducing the DEEP-VOICE dataset. DEEP-VOICE is comprised of real human speech from eight well-known figures and their speech converted to one another using Retrieval-based Voice Conversion.

For each speech, the accompaniment ("background noise") was removed before conversion using RVC. The original accompaniment is then added back to the DeepFake speech:

(Above: Overview of the Retrieval-based Voice Conversion process to generate DeepFake speech with Ryan Gosling's speech converted to Margot Robbie. Conversion is run on the extracted vocals before being layered on the original background ambience.)

Dataset There are two forms to the dataset that are made available.

First, the raw audio can be found in the "AUDIO" directory. They are arranged within "REAL" and "FAKE" class directories. The audio filenames note which speakers provided the real speech, and which voices they were converted to. For example "Obama-to-Biden" denotes that Barack Obama's speech has been converted to Joe Biden's voice.

Second, the extracted features can be found in the "DATASET-balanced.csv" file. This is the data that was used in the below study. The dataset has each feature extracted from one-second windows of audio and are balanced through random sampling.

Note: All experimental data is found within the "KAGGLE" directory. The "DEMONSTRATION" directory is used for playing cropped and compressed demos in notebooks due to Kaggle's limitations on file size.

A potential use of a successful system could be used for the following:

(Above: Usage of the real-time system. The end user is notified when the machine learning model has processed the speech audio (e.g. a phone or conference call) and predicted that audio chunks contain AI-generated speech.)

Kaggle The dataset is available on the Kaggle data science platform.

The Kaggle page can be found by clicking here: Dataset on Kaggle

Attribution This dataset was produced from the study "Real-time Detection of AI-Generated Speech for DeepFake Voice Conversion"

The preprint can be found on ArXiv by clicking here: Real-time Detection of AI-Generated Speech for DeepFake Voice Conversion

License This dataset is provided under the MIT License:

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

A dataset of 70,000 fashion images with labels for 10 classes. The dataset was created by researchers at Zalando Research and is used for research in machine learning and computer vision tasks such as image classification.

call