• Posts
  • RSS
  • ◂◂RSS
  • Contact

  • Whistle-based Synthesis

    October 14th, 2019
    whistling, music  [html]
    I'm reasonably happy with my Bass Whistle, where I can whistle and have it come out as a decent sounding bass. I've been using it when playing mandolin in a duo, and it fits well there. When playing piano or with a piano player, however, there's already bass so something that falls in a different place in the overall sound would be better. That could be melody, though I can't whistle fast enough for everything, but probably something simpler: harmonies, riffs, horn lines.

    When I take my current software, optimized for bass, and tell it to synthesize notes a few octaves up, it sounds terrible:

    • Raw whistled input:
      (mp3)
    • Bass version (needs headphones or good speakers):
      (mp3)
    • Treble version:
      (mp3)

    I'm using simple additive synthesis with the first four harmonics, which means adding together four sine waves. I think what's going on is that higher notes need more complexity to sound good? Playing around with distortion and fading the harmonics at different rates it sounds a bit more interesting:

    • Adding distortion:
      (mp3)
    • Adding fade:
      (mp3)
    • Adding both:
      (mp3)

    I'm still not very happy with it, though. It sounds artificial and silly. There are good synthesizers, the product of decades of work on turning "play this note at this time" into good sounding audio, so perhaps I could use my pitch detection to drive a standard synthesizer?

    I made some stand-alone open source software that pipes the pitch detection through to MIDI. This was kind of tricky: MIDI doesn't have a way to say "play this frequency". Instead you just have "play this note" and "bend the note by this much". How to interpret pitch bend is up to the synthesizer, but generally the range is ±2 half steps. So we need some math:

    in: wavelength
    in: sample_rate
    in: current_note
    
    # Convert from "this wave is 23.2 samples long"
    # to "the frequency is 1896.6 HZ".
    frequency = sample_rate / wavelength
    
    # MIDI is equal tempered, with each octave divided
    # into twelve logarithmically equal pieces.  Take
    # A440 as a reference point, so represent our
    # 1896.6 HZ as "25.29 half steps above A440":
    distance_from_a440 = 12 * log2(frequency / 440)
    
    # A440 is A4, or midi note 69, so this is 94.29.
    fractional_note = 69 + distance_from_a440
    
    # MIDI uses a note + bend encoding.  Stay on the
    # same note if possible to avoid spurious attacks.
    if (current_note and
        current_note - 2 < fractional_note
                         < current_note + 2)
      integer_note = current_note
    else
      integer_note = round(fractional_note)
    
    # To compute the pitch bend, we first find the
    # fractional part of the note, in this case 0.29:
    fractional_bend = fractional_note - integer_note
    
    # The bend will always be between -2 and +2, a
    # whole tone up or down.  MIDI uses 14 bits to
    # represent the range between -2 and +2, so -2 is 0
    # and +2 is 2^14.  The midpoint is 2^13, 8192:
    integer_bend = round((1 + fractional_bend / 2)
                          * 8192 - 1)
    
    # The bend is 14bits which gets split into two 7-bit
    # values.  We can do this with masking and shifting.
    bend_least_significant = integer_bend & 0b1111111
    
    bend_most_significant =
        (integer_bend & 0b11111110000000) >> 7
    
    out: integer_note
    out: bend_least_significant
    out: bend_most_significant
    

    Initially I screwed this up, and thought pitch bend was conventionally ±1 semitone, and didn't end up catching the bug until I wrote up this post.

    I have this working reasonably well, except that when I bend more than a whole note I get spurious attacks. Say I slide from A to C: the slide from A to Bb to B can all be done with pitch bend, but then once I go above B the system needs to turn off bent A and start working with a new note. I would love to suppress the attack for that note, but I don't know any way to communicate that in MIDI. I don't know what people with existing continuous pitch electronic instruments do?

    A second problem I've run into is that what sounds like a steady whistled pitch actually has a lot of tiny variations. Consider this input: (mp3)

    This sounds reasonably steady to me, but it isn't really. Here are some successive zero crossings:

    wavelength (samples) frequency (hz) midi note
    39.02 1130.07 49.3300
    39.26 1123.41 49.2277
    38.66 1140.68 49.4918
    39.25 1123.62 49.2309
    38.90 1133.71 49.3857
    38.85 1135.21 49.4087

    My synth doesn't mind, and just passes the variability through to the listener, where it's not a problem. I track where in the wave we are, and slowly adjust the rate we move through the wave to match the desired frequency:

    • Bass output:
      (mp3)
    • Pure sine treble output:
      (mp3)
    When I pass that variability into some regular synths, however, even when I don't cross note boundaries I get a wavery output. I think this may be another artifact of using a synth that isn't designed for continuous pitch input? Or possibly the problem is that real MIDI pitch wheels don't just suddenly jump from +23% to +49% over a 40ms period, and so they haven't needed to design for it?

    I can fix this some on my end by averaging the most recent pitches to smooth out the variability, but then it stops feeling so responsive (and quick slides don't work, and if I start a note slightly sour it takes longer to fix it). I think the answer is probably "find a better synth" but I'm not sure how to figure out what to use.

    Still, I like this a lot, and I think there's something here. If you have a mac and want to play with this, the code is on github.

    Comment via: facebook, lesswrong

    Recent posts on blogs I like:

    More on the Deutschlandtakt

    The Deutschlandtakt plans are out now. They cover investment through 2040, but even beforehand, there’s a plan for something like a national integrated timetable by 2030, with trains connecting the major cities every 30 minutes rather than hourly. But the…

    via Pedestrian Observations July 1, 2020

    How do cars fare in crash tests they're not specifically optimized for?

    Any time you have a benchmark that gets taken seriously, some people will start gaming the benchmark. Some famous examples in computing are the CPU benchmark specfp and video game benchmarks. With specfp, Sun managed to increase its score on 179.art (a su…

    via Posts on Dan Luu June 30, 2020

    Quick note on the name of this blog

    When I was 21 a friend introduced me to a volume of poems by the 14th-century Persian poet Hafiz, translated by Daniel Ladinsky. I loved them, and eventually named this blog for one of my favorite ones. At some point I read more and found that Ladinsky’s …

    via The whole sky June 21, 2020

    more     (via openring)


  • Posts
  • RSS
  • ◂◂RSS
  • Contact