Low Bit-Rate Speech Coding via MFCC Pseudo-Inverse Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in reconstructing speech from mel-frequency cepstral coefficients (MFCCs) lies in the information loss due to discarding the phase spectrum and mel-scale weighting functions, which affects the quality of reconstructed speech, especially at low bitrates, and existing methods struggle to maintain good quality while being compatible with distributed speech recognition systems.
Innovation Solution
The proposed solution involves computing a high-resolution mel-frequency cepstrum, using a Moore-Penrose pseudo-inverse to invert mel-scale weighting functions, and employing a least-squares estimation of the phase spectrum through the LSE-ISTFTM algorithm to reconstruct the speech waveform, allowing for efficient encoding and decoding at bitrates compatible with the ETSI Aurora DSR standard.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If mel-frequency cepstral coefficients are quantized at low bitrates, then transmission efficiency is improved, but speech reconstruction quality deteriorates due to information loss from discarding phase spectrum and mel-scale weighting functions
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing the pseudo-inverse of the mel-scale weighting functions matrix during system initialization. This pre-computed pseudo-inverse is then reused during the quantization and reconstruction phases, enabling efficient recovery of spectral information from quantized MFCCs without requiring complex real-time inversion operations, thus maintaining speech quality at low bitrates
Solution Approach 2:
The patent introduces an intermediary approach by using the pseudo-inverse of the mel-scale weighting functions as a mathematical mediator to reconstruct the power spectrum from quantized MFCCs. This intermediary transformation allows the system to recover spectral information that would otherwise be lost during quantization, bridging the gap between compressed representation and speech quality
2Device complexity
If phase spectrum is discarded during MFCC computation, then computational complexity is reduced, but reconstruction accuracy deteriorates
Solution Approach 1:
The patent applies inversion by reversing the typical MFCC computation process. Instead of computing MFCCs from the power spectrum (which discards phase information), the system inverts the transformation by using the pseudo-inverse of the mel-scale weighting functions to recover the power spectrum from quantized MFCCs, and then uses LSE-ISTFTM to estimate the phase spectrum and reconstruct the speech signal
3Manufacturing precision
If high-resolution mel-frequency cepstrum is used, then speech reconstruction quality is improved, but bit rate requirement increases
Solution Approach 1:
The patent applies parameter changes by quantizing the MFCCs using non-uniform quantization schemes that allocate more bits to coefficients with higher variance (typically the first few coefficients) and fewer bits to coefficients with lower variance. This parameter optimization allows high-resolution spectral representation to be achieved at lower overall bit rates by adapting the quantization precision to the actual information content of each coefficient
Data Source
AI summary
A method of (and concomitant computer software embodied on a non-transitory computer-readable medium for) generating speech comprising receiving a mel-frequency cepstrum employing a set of weighting functions, generating a pseudo-inverse of the set, reconstructing a speech waveform from the cepstrum and the pseudo-inverse, and outputting sound corresponding to the waveform. Also a corresponding method of (and concomitant computer software embodied on a non-transitory computer-readable medium for) encoding speech comprising receiving sounds comprising speech, computing mel-frequency cepstral coefficients from the sounds using a quantization method selected from the group consisting of non-uniform scalar quantization and vector quantization, and generating and storing codewords from the coefficients that permit recreation of the sounds.