Dataset Name: azspeech
Creator: Alas Development Center
Version: 1
Size: 1000+ hours, 400k+ voice files
Language: Azerbaijani
AzSpeech is a comprehensive voice dataset curated by the Alas Development Center, consisting of over 1000 hours of diverse voice recordings, totaling more than 400,000 individual voice files. This extensive collection has been meticulously compiled from various sources across the internet, ensuring a broad representation of linguistic nuances. The dataset aims to facilitate advancements in voice recognition technology, natural language processing, and machine learning research, offering a rich resource for developers, researchers, and organizations working in these fields.
Availability
Out of the extensive AzSpeech collection, 4k samples from the 400k available have been made accessible for review purposes. This initiative aims to provide a glimpse into the quality and diversity of the dataset, supporting the community's engagement with our voice data. Interested parties are encouraged to contact the Alas Development Center for access to the dataset and further collaboration.
Usage and Acknowledgments
Commercial Use:
Organizations interested in utilizing the AzSpeech dataset for commercial purposes are encouraged to get in touch with us. We offer access to the complete dataset on a paid basis. This approach enables organizations to explore the full extent of our dataset, tailored to meet the diverse needs of voice recognition technology, natural language processing, and machine learning applications.
Academic and Research Use:
Approximately 40% of the AzSpeech dataset (~400 hours) is designated for open-source use, aimed at supporting academic and research endeavors. Educational institutions wishing to access this portion of the dataset are required to form a partnership with the Alas Development Center. It is important to note that we will not be processing individual requests. Instead, our focus is on establishing collaborations with organizations that share our commitment to ethical data use. Organizations accessing the open-source data must fully comprehend and agree to our guidelines on data misuse prevention and adhere to our monitoring policy. This ensures the dataset's responsible use and aligns with our goals of advancing the field of voice technology research and development.
For educational institutions and research organizations interested in accessing the open-source portion of the AzSpeech dataset, please fill out the following form using your official company or institutional email. This process is designed to ensure that access is granted to legitimate academic and research entities committed to ethical and responsible use of the dataset.
Collection and Pre-processing
In the collection process for the AzSpeech dataset, all voice recordings have been sourced exclusively from public domains. Throughout this meticulous process, the Alas Development Center has adhered to international laws and regulations concerning data privacy, intellectual property rights, and ethical use of digital content. This adherence ensures that the dataset complies with global standards, safeguarding the interests of individuals and entities involved while fostering innovation and research in voice technology. Recognizing the importance of data quality for effective model training and research, we have undertaken a comprehensive preprocessing and denoising procedure to ensure the dataset provides ready data for users. This means the data is ready for immediate use in a range of applications, from fine-tuning text-to-speech and automatic speech recognition models to academic research.
Quality Assurance: Each voice file has undergone rigorous quality checks to ensure clarity and usability. This includes verifying the audio quality and ensuring the spoken content matches associated transcriptions.
Denoising: With advanced audio processing techniques, background noise has been significantly reduced in each recording. This denoising process enhances the purity of the voice data, making it more effective for training models that can distinguish nuanced vocal features.
Normalization: Audio files have been normalized to maintain consistent volume levels across the dataset. This standardization is crucial for avoiding bias towards louder or quieter recordings during model training.
Contact Information
For access to the AzSpeech dataset, partnership inquiries, or any other questions, please contact the Alas Development Center or or write to us on Linkedin.