Article details

Research area
Speech enhancement

Jahrestagung für Akustik der Deutsche Gesellschaft für Akustik


Ingo Schalk-Schupp, Mohamed Krini, Markus Buck, Andreas Wendemuth

Speech enhancement based on formant estimation


In this contribution, a computationally efficient method for enhancing speech quality and intelligibility by identifying and accentuating formants in speech signals is presented. The proposed method first identifies frequency regions containing speech formants and then accentuates those regions using an adaptive gain factor. The detection of formants can be performed using e.g. LPC techniques for estimating the vocal tract information of a speech sound and then searching for the LPC filter’s spectral peaks. A more efficient method is based in IIR smoothing of an amplitude spectrum along frequency using different smoothing constants. This method exploits the difference of a slowly varying and a fast varying spectral envelope for estimating the formants’ respective widths and positions. Once the formants are identified, different weighting rules can be utilized to boost the speech signal. A promising rule is the use of shaped windows concentrated on the central frequencies of formants and stretched to their respective widths. The gain values around the center of the shaped windows are dynamically adjusted depending on the presumed reliability of the formant estimation and combined afterwards with a regular noise suppression filter. Subjective evaluations indicated that the overall output quality can be enhanced – especially in high noise conditions.

Read/download now