Which datasets are commonly used for benchmarking audio search algorithms?

Several datasets are widely used for benchmarking audio search algorithms, each serving different use cases and technical challenges. Common choices include AudioSet, Freesound, MUSAN, UrbanSound8K, LibriSpeech, and GTZAN Genre Collection. These datasets vary in size, audio types (e.g., music, speech, environmental sounds), and annotation quality, making them suitable for testing algorithms under diverse conditions. For example, AudioSet provides a large-scale collection of labeled YouTube clips, while LibriSpeech focuses on clean speech for voice-based search. Developers often select datasets based on their target application, such as music retrieval, voice query matching, or environmental sound detection.

AudioSet is a popular choice for general-purpose audio search due to its scale and diversity. It contains over 2 million 10-second YouTube clips labeled with 632 hierarchical classes, covering sounds like music instruments, animals, and human activities. This dataset is useful for testing algorithms that need to handle noisy or overlapping sounds in real-world recordings. Another key dataset, Freesound, offers user-uploaded audio snippets tagged with metadata, which helps evaluate systems that rely on community-driven labels. For speech-focused search, LibriSpeech provides 1,000 hours of read English speech from audiobooks, ideal for testing voice query accuracy in controlled environments. MUSAN adds synthetic noise and music to clean speech data, enabling robustness testing against background interference.

Specialized use cases often require tailored datasets. UrbanSound8K, for instance, contains 8,732 short urban environment clips (e.g., sirens, drilling) labeled into 10 classes, useful for training models to detect specific real-world sounds. The GTZAN Genre Collection, though smaller (1,000 30-second music tracks), remains a benchmark for music genre classification and retrieval. Developers also use synthetic datasets like DCASE (Detection and Classification of Acoustic Scenes and Events) challenges, which include multi-channel recordings and complex acoustic scenes. When evaluating audio search algorithms, metrics like mean average precision (mAP), recall rates, and query latency are measured against these datasets to assess performance trade-offs between accuracy, speed, and scalability. Choosing the right dataset depends on the specific problem, such as handling ambient noise, scaling to large catalogs, or supporting multilingual queries.