Examples of TDHS

• de.maramuse.soundcomp.effect.TDHS
  For better decision of which way to go, we will implement a few planned modes of operation: Mode 0: Don't care about artifacts. Just do the transformation and let the buffer over/underflows happen. This is the cheapest way and its base algorithm will be used in any following mode. The buffer handling here will be: Determine the resampling factor F (f_out/f_in). For a semitone upwards, this will e.g. be 2^(1/12). Reject any resampling factors outside (0.1 .. 2). Split this number F into 2 factors, one (U, fractional) for upsampling and one (D, natural) for downsampling, so that F=D/U and D = 10. In our example, D=10, U=10*2^(-1/12) For each D samples that we will finally emit, we must insert U samples into the pipeline. Since U is not natural, we must do some error distribution there: some times we have to insert INT(U) samples, sometimes INT(U+1). We will aggregate the frational error to decide which case is relevant for each input sample. This upsampled data stream is LP-filtered with a SVF with an edge frequency of SAMPLERATE/2/D (actually the filter runs on a data stream D times as fast, so this simulates filtering on SAMPLERATE/2 since the filter does not know about the faked higher sample rate). Then, from the resulting upsampled and filtered data stream, take every D-th sample as final output. For F<1, this leads to buffer underruns, since we fill the pipe with less data than we attempt to get from it. For F>1, the opposite will happen: the pipe content will grow. Since we use a FIFO ring buffer, both cases will equally lead to "jumps" in the input stream by the size of the ring buffer. Mode 0 ignores these glitches, they will be addressed in following modes. Mode 1: split the input signal into two paths. one is double band pass filtered via state variable filters at an edge frequency of 20 and 50 hz, and then used for edge detection (upward zero crossings) to get a period duration D (in samples) hint. This should be largely cheaper than the usual autocorrelation mechanisms, and in most cases at least as precise (unless we have periodic signals, where autocorrelation is brilliant. But periodic signals will be more an exception here). the other path is upsampled by a large scale factor S (with S/F being natural and s<=100) then, depending on whether to shift up or down, either S*D sized fragments are cut out with windowing, or they are duplicated with windowing, from a buffer size of S*D/abs(F-1). After this, the resulting signal is low pass filtered by half the original sampling frequency, then every S/F-th sample is taken out to form the output sample stream. The upsampling may be a by non-natural factor. It is simply done by repeating the input value the necessary number of times. In case the factor is non-natural, it is approximated by spreading the error between repeating the ceil and floor values alternatively. This is really no problem since the resulting error is mitigated by the anti alias filtering that comes afterwards. It is essential that the downsampling is carried out by a natural factor since no error spreading is possible with post processing, because the downsampling is the last step. We will not allow any base frequency below audible range (more precisely, not even below SAMPLINGRATE/100). Otherwise, the buffer lengths would get way too long, and the delays would get really annoying. Therefore, the edge detection will not look at larger portions than 100 samples. Should the edge detection fail, the largest allowed buffer size is applied. Note that fast changes in the frequency ratio will almost surely produce artefacts since the pointer arithmetics in the up/downsampling buffers are not prepared for this parameter to change. Mode 2: Do a regular covariance detection. Limit this covariance to a relatively short range, but search in a longer range. So we can find shorter similar curve fragments within a signal of longer period without having to do the expensive covariance detection over thousands of samples. A range of 30 samples of best covariance should suffice, but search this within a range of at least 2000 samples (SAMPLERATE/20Hz). This method is more expensive, but I expect it to better suppress artifacts. As mode 1, it does not necessarily require a periodic input. My favourite candidate. Mode 3: @author jssr67

Examples of de.maramuse.soundcomp.effect.TDHS