Examples of edu.cmu.sphinx.frontend.window.RaisedCosineWindower
-
edu.cmu.sphinx.frontend.window.RaisedCosineWindower
Slices up a Data object into a number of overlapping windows (usually referred to as "frames" in the speech world). In order to minimize the signal discontinuities at the boundaries of each frame, we multiply each frame with a raised cosine windowing function. Moreover, the system uses overlapping windows to capture information that may occur at the window boundaries. These events would not be well represented if the windows were juxtaposed.
The number of resulting windows depends on the {@link #PROP_WINDOW_SIZE_MS window size} and the {@link #PROP_WINDOW_SHIFT_MS windowshift} (commonly known as frame shift in speech world). Figure 1 shows the relationship between the original datastream, the window size, the window shift, and the windows returned.
Figure 1: Relationship between original data, window size, window shift, and the windows returned.The raised cosine windowing function will be applied to each such window. Since the {@link #getData()} method returns awindow, and multiple windows are created for each Data object, this is a 1-to-many processor. Also note that the returned windows should have the same number of data points as the windowing function.
The applied windowing function, W(n), of length N (the window size), is given by the following formula:
W(n) = (1-a) - (a * cos((2 * Math.PI * n)/(N - 1)))
where a is commonly known as the "alpha" value. This variable can be set by the user using the property defined by {@link #PROP_ALPHA}. Please follow the links to the see the constant field values. Some values of alpha receive special names, since they are used so often. A value of 0.46 for the alpha results in a window named Hamming window. A value of 0.5 results in the Hanning window. And a value of 0 results in the Rectangular window. The default for this system is the Hamming window, with alpha 0.46 !). Figure 2 below shows the Hamming window function (a = 0.46), using our default window size of 25.625 ms and assuming a sample rate of 16kHz, thus yielding 410 samples per window.
Figure 2: The Hamming window function. @see Data
| |