Predicting fundamental frequency from mel-frequency cepstral coefficients to enable speech reconstruction

Area of Science:

• Speech processing
• Acoustic signal analysis
• Machine learning for audio

Background:

• Distributed speech recognition (DSR) systems often rely on mel-frequency cepstral coefficients (MFCCs).
• Traditional speech reconstruction methods require MFCCs along with fundamental frequency and voicing information.
• Reconstructing speech solely from MFCCs presents a significant challenge.

Purpose of the Study:

• To develop a method for reconstructing acoustic speech signals exclusively from MFCCs.
• To predict voicing classification and fundamental frequency directly from MFCCs.
• To evaluate the effectiveness of the proposed speech reconstruction method.

Main Methods:

• Proposed a novel speech reconstruction technique utilizing only MFCCs.
• Employed two maximum a posteriori (MAP) methods for predicting voicing and fundamental frequency from MFCCs.
• Utilized a Gaussian mixture model (GMM) for joint density modeling of MFCCs and fundamental frequency.
• Implemented a Hidden Markov Model (HMM)-based approach with state-dependent GMMs for localized density modeling.

Main Results:

• Achieved accurate voicing classification and fundamental frequency prediction compared to reference measurements.
• Speaker-independent experiments on male and female speech demonstrated the method's efficacy.
• Speech reconstruction using predicted parameters yielded speech quality comparable to using reference parameters.

Conclusions:

• Accurate speech reconstruction is feasible using only MFCCs by predicting essential acoustic features.
• The proposed MAP-based methods effectively estimate voicing and fundamental frequency from MFCCs.
• This approach simplifies DSR systems by eliminating the need for separate voicing and fundamental frequency transmission.