 |
Edinburgh Speech Tools
2.1-release
|
|
|
Edinburgh Speech Tools
2.1-release
|
|
Some examples of track manipulations.
The constructor functions can be used to create a track with zero frames and channels or a track with a specified number of frames and channels
tracks can be resized at any time:
by default, resizing preserves values in the track. This may involve copying some information, so if the existing values are not needed, a flag can be set which usually results in quicker resizing
If only the number of channels or the number of frames needs to be changed, this an be done with the following functions:
The preserve flag works in the same way with these functions
Values in the track can be accessed and set by frame number and channel number.
The following resizes a track to have 500 frames and 10 channels and fills every position with -5.
A well formed track will have a time value, specified in seconds, for every frame. The time array can be filled directly:
which fills the time array with values 0.01, 0.02, 0.03... 5.0. However, A shortcut function is provided for fixed frame spacing:
which performs the same operation as above. Frames do not have to be evenly spaced, in pitch synchronous processing the time array holds the time position of each pitch period. In such cases each position in the time array must obviously be set individually.
Some representations have undefined values during certain sections of the track, for example the F0 value during unvoiced speech.
The break/value array can be used to specify if a frame has an undefined value.
If a frame in this array is 1, that means the amplitude is defined at that point. If 0, the amplitude is undefined. By default, every frame has a value.
Breaks (undefined values) can be set by EST_Track::set_break(). The following sets every frame from 50 to 99 as a break:
frames can be turned back to values as follows:
It is up to individual functions to decide how to interpret breaks.
A frame's status can be checked as follows:
While channels can be accessed by their index, it is often useful to give them names and refer to them by those names.
The EST_Track::set_channel_name() function sets the name of a single channel:
An alternative is to use a predefined set of channel names stored in a map. A track map is simply a String List strings which describe a channel name configuration. The EST_Track::resize function can take this and resize the number of channels to the number of channels indicated in the map, and give each channel its name from the map. For example:
A convention is used for channels which comprise components of a multi-dimensional analysis such as cepstra. In such cases the channels are named TYPE_I. The last coefficient is always names TYPE_N regardless of the number of coefficients. This is very useful in extracting a set of related channels without needing to know the order of the analysis.
For example, a track map might look like:
This obviously gets unwieldy quite quickly, so the mapping mechanism provides a short hand for multi-dimensional data.
Here $ indicates the special status, "cep" the name of the coefficients, "-0" that the first is number 0 and "+8" that there are 8 more to follow.
Often functions perform their operations on only a single frame or channel, and the track class provides a general mechanism for doing this.
Single frames or channels can be accessed as EST_FVector : Given a track with 500 frames and 10 channels, the 50th frame can be accessed as:
now tmp_frame is 10 element vector, which is a window into tr: any changes to the contents of tmp_frame will change tr. tmp_frame cannot be resized. (This operation can be thought in standard C terms as tmp_frame being a pointer to the 5th frame of tr).
Likewise with channels:
Again, tmp_channel is 500 element vector, which is a window into tr: any changes to the contents of tmp_channel will change tr. tmp_channel cannot be resized.
Channels can also be extracted by name:
not all the channels need be put into the temporary frame. Imagine we have a track with a F0 channel,a energy channel and 10 cepstrum channels. The following makes a frame from the 50th frame, which only includes the cepstral information in channels 2 through 11
In addition to extracting single frames and channels, multiple frame and channel portions can be extracted in a similar way. In the following example, we make a sub-track sub, which points to the entire cepstrum portion of a track (channels 2 through 11)
Parameter sub behaves exactly like a normal track in every way, except that it cannot be resized. Its contents behave like a point into the designated portion of tr, so changing sub will change tr.
The first argument is the sub track. The second states the start frame and the total number of frames required. EST_ALL is a special constant that specifies that all the frames are required here. The next argument is the start channel number (remember channels are numbered from 0), and the last argument is the total number of channels required.
This facility is particularly useful for using standard signal processing functions efficiently. For example, the melcep in the signal processing library takes a waveform and produces a mel-scale cepstrum. It determines the order of the cepstral analysis by the number of channels in the track it is given, which has already been allocated to have the correct number of frames and channels.
The following will process the waveform sig, produce a 10th order mel cepstrum and place the output in sub. (For explanation of the other options see melcep
because we have made sub a window into tr, the melcep function writes its output into the correct location, i.e. channels 2-11 of tr. If it were no for the sub_track facility, either a separate track of the right size would be passed into melcep and then it would be copied into tr (wasteful), or else tr would be passed in and other arguments would have to specify which channels should be written to (messy).
Sub-tracks can also be set using channel names. The following example does exactly as above, but is referenced by the name of the first channel required and the number of channels to follow:
and this specifies the end by a string also:
sub_tracks can be any set of continuous frames and channels. For example if a word started at frame 43 and ended and frame 86, the following would set a sub track to that portion:
We can step through the frames of a Track using a standard iterator. The frames are returned as one-frame sub-tracks.
The EST_Track::channel, EST_Track::frame and EST_Track::sub_track functions are most commonly used to write into a track using a convenient sub-portion. Sometimes, however a simple copy is required whose contents can be written without affecting the original.
The EST_Track::copy_sub_track function does this
Individual frames and channels can be copied out into pre-allocated float * arrays as follows:
Individual frames and channels can be copied into the track from float * arrays as follows:
Auxiliary channels are used for storing frame information other than amplitude coefficients, for example voicing decisions and points of interest in the track.
Auxiliary channels always have the same number of frames as the amplitude channels. They are resized by assigning names to the channels that need to be created:
The following fills in these three channels with some values:
Tracks in various formats can be saved and loaded:
Save as a HTK file:
Save as a EST file:
Save as an ascii file:
The file type is automatically determined from the file's header during loading:
If no header is found, the function assumes the file is ascii data, with a fixed frame shift, arranged with rows representing frames and columns channels. In this case, the frame shift must be specified as an argument to this function:
1.8.11