Speech signal processing used to the sub-frame, frame shift, the concept of windowing, filtering, noise reduction, synthesis, the basic concept of each reproduced as follows:
First, the sub-frame
Voice data and video data is different, this is no concept of frames, but in order to transport and storage, we collected data is a section of the audio. In order to be able to batch processing procedures will be performed according to the specified segment length (number of samples or period of time), our structured programming data structures, this is the sub-frame.
Second, the frame shift
Because of our common signal processing methods require signal is continuous, it must be said that the signal start to finish without interruption by a disconnect. However, we sampled or sub-frames of data are turned off, so to retain the overlapped portion of data from frame to frame, in order to satisfy the continuity requirement, which is partially overlapping data frame shift.
Third, windowing
Frame shift introduced when we said, we claim a method of processing a signal are signal is continuous conditions, but when the sub-frame processing intermediate disconnected, in order to satisfy the conditions we will portioned by a piece of data with the frame data length, this data is the data within the window function throughout the cycle changes from the minimum to the maximum, and minimum.
Fourth, the filter
We know that we are dealing with voice is a kind of acoustic sound waves is a matter waves. Filtered literally understood to filter some of the waves of different frequencies. According to Fourier transform, we know that any wave can be decomposed into several superimposed sine and cosine waves, and from the viewpoint of probability theory, i.e., the weighting filter. Filtering effect is to give a different signal components of different weights. The simplest loss pass filter, the low-frequency signal is directly to the weights 0, and 1 weight part of the high frequency. For more complex filtering, such as Wiener filtering, based on statistical knowledge signal will have to design weights.
When the higher frequency component of the signal allowed by the filter, such a filter is called a high-pass filter.
While allowing lower frequency component of the signal by the filter, such a filter is called a low-pass filter.
When only a signal component in a certain frequency range by a filter, such a filter called a band pass filter.
When the signal component in a certain frequency range allowed by the filter, such a filter is called a band stop filter.
Fifth, noise reduction
From the statistical signal processing, noise reduction may be as a kind of filter. Purpose of noise reduction is to highlight the influence of noise suppressed signal itself. From this point, the signal to noise is a high weight to the noise and a low weight. Wiener filter is a typical noise reduction filter.
Sixth, synthesis
General speech processing, the first sub-frame, and then divided into individual sub-band in the frequency domain processing, post-processing the time-domain transformed into, synthesized speech. Seen from the description, the speech synthesis and the reverse process is framed, we ensure that the signal after the data conversion process to return to the original state. After each frame is converted into the respective sub-band synthesis are superimposed as the time series data of one frame.