Pitch Change by Resampling Tutorial

Roger Dannenberg

This tutorial shows how to change the pitch of a sound using resampling. In this example, a source sound is resampled to effectively play the sound faster and/or slower. This changes the perceived pitch. It also makes the sound play faster and/or slower, so the original duration is not preserved.

To control the playback speed, a control function is used. The function specifies the pitch shift in steps, so if the function is zero, there is no change. If the signal is 1, the signal is shifted up one half-step. If the signal is -2, the signal is shifted down a whole-step, etc.

The control signal can of course change over time. The control signal time corresponds to time in the resulting sound. Thus, if the control signal jumps from 0 to 5 at time 3, the resulting sound will jump up a musical fourth (5 half-steps) at time 3. This may or may not be time 3 in the source sound.

The code implements VARIABLE-RESAMPLE which takes two parameters:

The function works as follows: First, steps is converted to ratio, which is the speed change expressed as a ratio. E.g. when steps is 12 (an octave), ratio will be 2 (a 2:1 ratio) Second, ratio is integrated to get map. The map is a function from real time to the score. Note that the slope of the map at any time determines the amount of speedup. Finally, SND-COMPOSE is used to map the source sound according to the tempo map. Because SND-COMPOSE is defined to output the sample rate of the map, we coerce the sample rate of map to *sound-srate*.

This function could be improved by adding code to handle stereo inputs. Also, this function should really be implemented using snd-resamplev, which uses a high-quality resampling algorithm. Unfortunately, there a bug in snd-resamplev, so until the bug is fixed, snd-compose actually sounds better.

The resulting sound ends whenever either steps or snd come to an end. To control duration, you may want to make sure that the source sound is much longer than the control function. That way, even if you speed it up, it will still last long enough to exhaust the control function, and you know by design how long that is. You can also apply an envelope to the result to trim it to a known length.

So, here finally is the implementation: 

(defun variable-resample (steps snd)
  (let ((p1 (/ (log 2.0) 12))  ; p1 helps convert steps to a ratio
        ratio map)
    (setf ratio (s-exp (mult steps p1))) ; pitch ratio
    (setf map (integrate ratio)) ; map from real-time to sound time
    (snd-compose snd (force-srate *sound-srate* map))))
Here is an example: the PWL control function is initially zero (no transposition), but starting at time 1.0, it rises to 2.0 in 0.1s where it remains until time 3.0. The function ends at time 3.0. The sound to be modified in this case is a simple sinusoid with a duration of 8. In this case, the PWL function finishes at 3.0s, well before the OSC sound comes to an end. 
(play
       (variable-resample (pwl 1.0 0.0 1.1 2.0 3.0 2.0 3.0) (osc c4 8.0)))
Note that you can replace (osc c4 8.0) with any sound, including one read from a sound file using S-READ.