Introduction
In the field of Music Information Retrieval (MIR), researchers frequently need to extract dozens of distinct audio features—such as rhythm, pitch, and timbre—from massive datasets. Traditionally, this process has been a significant computational bottleneck. Existing tools often forced a compromise between flexibility and efficiency. For instance, while the MIR Toolbox (Matlab) is user-friendly, it suffers from high memory overhead. Conversely, tools like Marsyas offer efficiency but limited parameter control, and VAMP plugins lack the ability to share intermediate computations during batch processing.
To address these challenges, YAAFE (Yet Another Audio Feature Extractor) was developed as an open-source framework designed to handle large-scale audio analysis with minimal memory usage and maximum speed.
The YAAFE Architecture
YAAFE’s efficiency stems from its unique two-stage dataflow architecture, which bridges high-level logic with low-level performance.
- The Parser (Python): When a user defines a "feature extraction plan," a Python-based parser analyzes the requested features to identify "computational redundancies." If multiple features require the same underlying mathematical step, such as a Fast Fourier Transform (FFT), the parser builds an optimized directed acyclic graph (DAG). This ensures that shared steps are calculated only once.
- The Engine (C++): The optimized graph is executed by a C++ engine that processes audio block-by-block. This approach allows YAAFE to analyze extremely long files without loading the entire audio signal into RAM, maintaining a low memory footprint.
Additionally, YAAFE ensures temporal alignment across all features. By centering analysis windows at identical time points regardless of varying window sizes, the framework produces perfectly synchronized data rows ready for machine learning applications.
Performance Benchmarks
To evaluate YAAFE, the authors compared it against industry standards like Marsyas and Sonic Annotator (a VAMP-based tool) using a dataset of 80 radio excerpts (40 hours of 32 kHz mono audio).
As shown in the table below, YAAFE significantly outperformed its peers in both processing speed and storage efficiency. For a standard feature set (including MFCC, Centroid, and Rolloff), YAAFE completed the extraction in just 6 minutes and 34 seconds—roughly eight times faster than Sonic Annotator and nearly four times faster than Marsyas.
Beyond speed, YAAFE utilizes the HDF5 binary format for output. This choice supports on-the-fly compression and avoids the overhead of text-based formats like CSV or ARFF, resulting in output files that are significantly smaller (1.22 GB for YAAFE vs. 2.7 GB for Marsyas).
Conclusion and Future Work
YAAFE successfully eliminates the computational redundancies inherent in traditional MIR workflows. While it currently excels at low-to-mid-level spectral features, future development aims to incorporate more complex high-level descriptors such as fundamental frequency (F0) and tempo estimation. Its proven scalability in large-scale evaluation campaigns like the Quaero project demonstrates its readiness for "big data" audio tasks.
Summary updated