Logarithmic Frequency Spectrum Shift Prediction for Speech Signal Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for obtaining differential pitch frequency information from speech signals are prone to errors due to background noise and require narrowing down the pitch frequency range, making it difficult to accurately extract this information without noise influence.
Innovation Solution
A feature extraction apparatus and method that calculates a frequency spectrum on a logarithmic scale, computes cross-correlation coefficients between frames, and predicts the shift amount of frequency spectra using these coefficients, reducing noise influence without restricting the pitch frequency range.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the range of the autocorrelation function of the predictive residuals is narrowed down to the vicinity of the accurate pitch frequency, then the measurement precision of differential pitch frequency information is improved, but the device complexity increases because the pitch frequency has to be calculated in advance and the range has to be suitably determined
Solution Approach 1:
The patent applies preliminary action by calculating the pitch frequency in advance and using it to determine the search range for the autocorrelation function. This preliminary pitch frequency calculation enables the subsequent differential pitch frequency extraction to be performed within a narrowed, optimized range, improving measurement precision while managing complexity through pre-computation
Solution Approach 2:
The patent segments the pitch frequency analysis by dividing it into two distinct stages: first calculating the pitch frequency, then using that result to define a specific range for the autocorrelation function. This segmentation allows each stage to be optimized independently, with the first stage providing information for the second stage
2Ease of operation
If the pitch frequency range is not narrowed down, then the ease of operation is improved, but the reliability of differential pitch frequency information deteriorates due to background noise influence
Solution Approach 1:
The patent introduces the pitch frequency calculation as an intermediary step that mediates between the broad pitch frequency range and the specific differential pitch frequency extraction. This intermediary provides a reference point that guides the autocorrelation function search, enabling reliable differential extraction without manually specifying the entire frequency range
Solution Approach 2:
By performing pitch frequency calculation in advance as a preliminary action, the system establishes a reference that automatically defines the appropriate search range for differential extraction, eliminating the need for manual range specification while ensuring reliability against background noise
Data Source
AI summary
A feature extraction apparatus includes a spectrum calculating unit that calculates, based on an input speech signal, a frequency spectrum having frequency components obtained at regular intervals on a logarithmic frequency scale for each of frames that are defined by regular time intervals, and thereby generates a time series of the frequency spectrum; a cross-correlation coefficients calculating unit that calculates, for each target frame of the frames, a cross-correlation coefficients between frequency spectra calculated for two different frames that are in vicinity of the target frame and a predetermined frame width apart from each other; and a shift amount predicting unit that predicts a shift amount of the frequency spectra on the logarithmic frequency scale with respect to the predetermined frame width by use of the cross-correlation coefficients.