Python music sequencer timing problems
research at johnohagan.com
Thu Dec 11 02:21:24 CET 2008
On Wed, 10 Dec 2008, badmuthahubbard wrote:
> I've been trying to get the timing right for a music sequencer using
> Tkinter. First I just loaded the Csound API module and ran a Csound
> engine in its own performance thread. The score timing was good,
> being controlled internally by Csound, but any time I moved the mouse
> I got audio dropouts.
> It was suggested I run the audio engine as a separate process, with
> elevated/realtime priority and use sockets to tell it what to play,
> and that way, too, people could set up servers for the audio on
> different CPUs. But I've found that the method I came up with for
> timing the beats/notes is too slow (using threading.Timer on a
> function that calls itself over and over), and the whole thing played
> too slowly (and still gave me noise when moving the mouse). I've been
> using subprocesses, but I'm now wondering if sockets would or could
> make a difference.
> The overall goal is this: when the user wants to audition a piece,
> create an audio engine process with elevated/realtime priority. This
> engine also has all the synthesis and sound processing rules for the
> various instruments, due to the way Csound is structured. Set up a
> scheduler- possibly in another process, or just another thread- and
> fill it with all the notes from the score and their times. Also, the
> user should be able to see a time-cursor moving across the piece so
> they can see where they are in the score. As this last bit is GUI,
> the scheduler should be able to send callbacks back to the GUI as well
> as notes to the audio engine. But neither the scheduler nor the audio
> engine should wait for Tkinter's updating of the location of the time-
> cursor. Naturally, all notes will have higher priorities in the
> scheduler than all GUI updates, but they won't necessarily always be
> at the same time.
> So, I have a few ideas about how to proceed, but I want to know if
> I'll need to learn more general things first:
> Create both the scheduler and the audio engine as separate processes
> and communicate with them through sockets. When all events are
> entered in the scheduler, open a server socket in the main GUI process
> and listen for callbacks to move the cursor (is it possible to do this
> using Tkinter's mainloop, so the mouse can be moved, albeit
> sluggishly, at the same time the cursor is moving continuously?); the
> audio engine runs at as high priority as possible, and the scheduler
> runs somewhere between that and the priority of the main GUI, which
> should even perhaps be temporarily lowered below default for good
> Create the audio engine as an elevated priority process, and the
> scheduler as a separate thread in the main process. The scheduler
> sends notes to the audio engine and callbacks within its own process
> to move the GUI cursor. Optionally, every tiny update of the cursor
> could be a separate thread that dies an instant later.
> Closer to my original idea, but I'm hoping to avoid this. All note
> scheduling and tempo control is done by Csound as audio engine, and a
> Csound channel is set aside for callbacks to update the cursor
> position. Maybe this would be smoothest, as timing is built into
> Csound already, but the Csound score will be full of thousands of
> pseudo-notes that only exist for those callbacks. Down the road I'd
> like to have notes sound whenever they are added or moved on the
> score, not just when playing the piece, as well as the option of
> adjusting the level, pan, etc. of running instruments.
I've recently been fooling with something involving timing and synchronising
multiple note-on/note-off events, and also tried threading and subprocesses
without much success - out of the box, the high-tempo precision wasn't there.
But I had more luck with this approach, if it's not way too simple for your
purpose (I'm not using a gui, for example); in psuedocode:
from time import time, sleep
start = time()
for event in music:
duration=len(event) #Really, the length of the event
timer = time()
remaining = start + duration - timer
if remaining < 0.001:
sleep(remaining / 2)
start += duration
IOW, just check the time, wait half the remaining note duration, check the
time again, etc, till you've reached your desired precision level (in this
case 0.001 sec). The halving of each succesive sleep() means that there is
only a handful of calls to time() per note, 5-10 depending on the tempo. (Of
course it could be any fraction, I just pulled that out of a hat; it would
probably be better too if the fraction decremented as the deadline
Even with this naive algorithm, I'm getting accuracy of <0.001 sec
consistently, up to stupidly high tempos like 3000 bpm (as measured by
inserting a "print remaining" in the loop) without using a separate timing
thread or "sync pulse", and with only a few calls.
Any processes started this way stay in sync, even if individual events are
delayed, because the "start" variable is incremented by the correct (not
actual) duration of the note. (But I guess that's just scheduling,
right? :) )
Obviously there is a limit imposed by arithmetical precision, but I've left
things running all day without seeing a discrepancy.
In terms of audio process communication, I'm using sox in a subprocess to play
samples (and its cheesey synth sound!), and fluidsynth via a socket for midi.
The latter is far superior and seems to handle whatever is thrown at it, and
with the jack driver with realtime priority, the dropouts you mentioned
caused by mouse movements and so on disappear.
Of course, this doesn't prove that using a socket is better; I don't think sox
is jack-aware, and if the subprocesses could have different priorities from
the parent process, that may help with the difficulties we're having with
them. Anyone out there know about that?
I've been putting off getting into Csound, but by all accounts it's _the_
audio language. Good luck with your project, it sounds interesting.
More information about the Python-list