computeIIF-Matlab - a Matlab implementation for the computation of IIFs

Recently, invariant integration features (IIFs) have been proposed as an alternative to the standard feature extraction methods for automatic speech recognition.

The IIFs are based on the idea of constructing invariant features by integrating nonlinearly transformed input signals over a finite transformation group. In the context of feature extraction for ASR, monomials up to a certain order are a good choice as nonlinear functions. Experiments showed that the IIFs perform superior compared to MFCCs and PLPs, especially in mismatching training-testing conditions.

This is a Matlab implementation for the computation of IIFs, which was used within our experiments. The implementation computes IIFs together with log-energy and delta features. In addition, reduction and decorrelation matrices are contained within the download.

References

2011

  • Müller, F. and Mertins, A.: Contextual invariant-integration features for improved speaker-independent speech recognition. Speech Communication, vol. 53, no. 6, pp. 830 - 841, 2011
    BibTeX DOI Website PDF

2010

  • Müller, F. and Mertins, A.: Invariant Integration Features Combined with Speaker-Adaptation Methods. Proc. Int. Conf. Spoken Language Processing (Interspeech 2010-ICSLP), Makuhari, Japan, pp. 2622--2625, Sept. 2010
    BibTeX PDF

2009

  • Müller, F. and Mertins, A.: Invariant-integration method for robust feature extraction in speaker-independent speech recognition. Proc. Interspeech 2009, Brighton, Sept. 2009
    BibTeX PDF