ANA-Mix: A Synthetic Corpus of Mandarin Speech in Airport Noise Conditions

This paper presents the Airport Noise-AISHELL Mix (ANA-Mix), a rich and realistic dataset tailored for advancing speech recognition and interactive systems in complex airport acoustic conditions. The noisy speech dataset is constructed by combining the publicly available AISHELL-3 Mandarin speech dataset with the environmental noise data actually collected at airports. The AISHELL-3 dataset provides a rich variety of high-quality sentence recordings, while the airport noise data captures a variety of typical airport noise scenarios, including crowd conversations, luggage rolling, and boarding announcements. A data mixing method is used to superimpose clean speech and randomly selected airport noise in waveforms to create 200,000 sets of noisy speech samples, including approximately 100,000 sets of single-person noisy speech and another 100,000 sets of multi-person (2~4 speakers) speech. This voice construction results are close to the actual deployment environment. The dataset constructed in this study can be used for a variety of tasks such as speech recognition, voiceprint recognition, and speech enhancement, demonstrating its potential value in improving the performance of voice interaction systems.