I wanted to hum a melody into FL Studio. Hearing the pitch was the easy part.

A MIDI note is simple: a number from 0 to 127, a start and an end. A voice is not. It slides into notes, wobbles with vibrato, drifts flat over a long vowel, breathes in the middle of a phrase, and is rarely in tune unless you have trained for years. Turn that stream straight into note names and one held note becomes six: C4, C4 a little sharp, C4 a little flat, B3 for a single frame, C4 again.

The brief I wrote for myself was short: a producer should be able to hum an idea and get usable MIDI, without knowing which notes they are singing. Behind it sat firm constraints. Everything runs locally, with no cloud service. One voice at a time. Windows and FL Studio first, but with the music logic kept apart from FL Studio, so other DAWs stay possible. And a rule that shaped the whole project: do not fake support where the host doesn't provide it.

The brief asked for a real comparison of pitch methods, weighing accuracy, CPU, latency and complexity, with no cloud dependency allowed in version 1.

YIN works by asking how well a sound matches a delayed copy of itself. It computes that difference for every possible delay, normalises it so quiet and loud notes compare fairly, then takes the first delay that dips under a threshold. The best delay, refined with a parabola through its neighbours, is one period of the note, so frequency is the sample rate divided by it. The depth of the dip also says how periodic the sound is, and PitchPen turns that into a confidence for every moment.

The numbers are chosen for voices. The window is about 32 ms: 1,536 samples at 48 kHz, scaled for 44.1, 88.2 and 96 kHz so it always covers the same time. It moves forward a quarter of a window at a time, which gives roughly 125 measurements a second. A longer window sees more of a low note and is steadier, but adds delay. The result is a total of about 24 ms (half a window plus one step), which is short enough that the pitch curve feels attached to your voice. Searching only between 60 Hz and 2 kHz keeps the sub-period false dips and rumble out, while still covering deep voices and whistling.

Two small decisions matter more than they look. First, when no delay dips under the threshold, the deepest local minimum is still accepted if the sound is periodic enough, so unsteady singing yields a lower confidence instead of a gap. Second, each measurement is stamped at the centre of its window, not the start. Without that, every note would land 16 ms before it was sung.

I was honest with myself about one deviation. The scope planned a more careful offline pYIN pass after capture. What shipped is YIN plus a pitch-hold step, because the live analysis thread already produces the frames, and reusing them means any setting change can rebuild the melody instantly. The cost is that the tracker itself can't be re-run with different parameters after capture. A proper pYIN pass remains on the list.

On generated voices, with harmonics and vibrato and a weak fundamental, the pitch tracker returned the right note every time. My own first real take was different: the notes landed high, around C6, and flipped between neighbouring semitones, so what came out of the Piano Roll didn't sound like what I had hummed. I'm not a trained singer, which made it the right test.

That one take changed the defaults. Loose tolerance and Auto Octave are now on from the start. Auto Octave lifts the whole melody by whole octaves so its middle sits around F#4, which is where instruments are comfortable. The capture summary also reports the range it found ("Captured 9 notes, E4 to C#5"), so a melody in the wrong register is obvious at a glance. The point of a tool for non-singers is that the forgiving settings are the ones you get without asking.

Audio code has one hard rule: the thread that fills the sound card's buffer must never wait. PitchPen's audio thread only filters (DC removal, a 70 Hz high-pass, a noise gate), measures level, mixes everything down to mono, and pushes samples into a fixed-size lock-free queue. It allocates nothing, locks nothing and touches no files. A separate high-priority analysis thread reads the queue, runs the pitch tracker and keeps a short history for the live curve. The interface thread only reads results.

While you capture, the analysis thread also keeps every pitch frame. When you press Stop, everything after pitch tracking, which is segmentation, key detection, quantizing and velocities, runs on those frames and is quick enough to feel instant. That's why changing a preset, the tolerance or the grid after capture feels instant, and why the raw audio never needs to be stored.

One trap came from the framework. JUCE's stand-alone wrapper starts with the microphone input muted to prevent feedback, a default I found in its source. Left alone, PitchPen would have shown a dead meter on first launch. PitchPen unmutes it, and stays safe by outputting silence unless monitoring is switched on.

A pitch tool needs known answers. The test suite generates its own sound: sine waves checked at 44.1, 48, 88.2 and 96 kHz, a noisy tone that must score lower than a clean one, silence that must produce nothing, and a two-note phrase with a pause between them that must come out as two notes with the right pitches. Note cleanup is tested with hand-made pitch frames, including the C4-with-a-glitch case. In total there are 52 tests and 231 assertions, all passing. Outside the suite, a probe fed voice-like tones with strong harmonics and a weak fundamental, male and female ranges, and every frame came back as the right note.

Two extra harnesses sit outside the suite. One loads the real VST3 the way a DAW does: stereo layout, the editor open, audio running, and checks it stays alive. It earned its keep when, early on, the plugin window flashed and closed inside FL Studio. I feared a crash, but the harness showed the editor was fine. The cause was dropping an effect plugin onto a Piano Roll, which isn't something a plugin can be dropped on. The other feeds a generated voice through the real engine in real time and checks that A4, E4 and G4 come out with the right timing.

A third tool made the product page's screenshots, and found a bug by doing it. It runs the real editor, feeds it a generated hummed phrase with vibrato, out-of-tune notes and one deliberate slip, and saves the window. In the first set, Raw and Clean looked identical. My "raw" view was showing a melody that had already been through the pitch hold, so the Before/After view showed nothing. It now shows a naive conversion where every pitch change is a note, and a test checks that it contains more notes than the cleaned melody.

PitchPen 1.0 does the whole loop on my machine: hum into the microphone, see the pitch curve, get a cleaned and editable melody with a detected key, fix it with the mouse, and send it into an FL Studio 2026 Piano Roll as ordinary notes. A generated voice gets the right notes with the right timing, even through vibrato, an out-of-tune take and a single-frame slip.

What it isn't: it follows one voice and can't do chords. It captures first and edits after, so you can't yet play an instrument live with your voice. It is classic signal processing, not a neural network, so noisy rooms and whispers can beat it, and the scope's more careful pYIN pass isn't in. It has no pitch bend or expression. It is Windows only, the one-click Piano Roll route is FL Studio only, and I have tested it on one machine with one DAW. The installer isn't code-signed. Next on the list: a live mode, pitch bend for gliding notes, the pYIN pass, bridges for other DAWs, and signing.

It was never meant to be sold. It exists because I wanted to get a tune out of my head and into a project without first learning which keys to press.