Acoustic event classification using multi-resolution HMM

Real-world acoustic events span a wide range of time and frequency resolutions, from short clicks to longer tonals. This is a challenge for the hidden Markov model (HMM), which uses a fixed segmentation and feature extraction, forcing a compromise between time and frequency resolution. The multiresolution HMM (MR-HMM) is an extension of the HMM that assumes not only an underlying (hidden) random state sequence, but also an underlying random segmentation, with segments spanning a wide range of sizes and processed using a variety of feature extraction methods. It is shown that the MR-HMM alone, as an acoustic event classifier, has performance comparable to state of the art discriminative classifiers on three open data sets. However, as a generative classifier, the MR-HMM models the underlying data generation process and can generate synthetic data, allowing weaknesses of individual class models to be discovered and corrected. To demonstrate this point, the MR-HMM is combined with auxiliary features that capture temporal information, resulting in significantly improved performance.