Noise estimation in voice signals using short-term cepstral analysis

Area of Science:

• Speech processing
• Acoustics
• Signal analysis

Background:

• Accurate noise level estimation is crucial for voice analysis.
• Traditional methods may struggle with complex noise components in voiced speech.
• Cepstral analysis offers a robust approach to spectral feature extraction.

Purpose of the Study:

• To develop and validate a cepstral-based method for estimating the noise baseline in voiced speech.
• To introduce a new harmonics-to-noise ratio (HNR) estimation technique utilizing the noise baseline.
• To assess the accuracy of the proposed HNR estimation on synthetic voice signals.

Main Methods:

• Cepstral analysis of voiced speech, including aspiration noise.
• Fourier transform of the liftered cepstrum to obtain a noise baseline estimate.
• Comparison of the cepstral method to moving average filtering of the logarithmic spectrum.
• Development of a new HNR estimation technique based on the noise baseline.

Main Results:

• The cepstral-based noise baseline estimation is comparable to moving average filtering.
• The noise baseline is influenced by harmonic resolution and glottal source spectral tilt.
• The proposed HNR estimation technique yields accurate results on synthetic voice data.

Conclusions:

• Cepstral-based noise baseline estimation provides a reliable method for voiced speech analysis.
• The novel HNR estimation technique demonstrates high accuracy.
• This approach offers a valuable tool for quantitative voice assessment.