I made a Lowrey organ part of a Suno bossa/dubstep song: how “Bosa Nova
Party Robot” became a MIDI + audio performance

September 27, 2026 | NITETIME.NET
Gopher edition of the MIDI process and song report.

I've been building a setup for NITETIME that takes a song generated with
Suno and gives my physical Lowrey organ its own part in the arrangement. The
latest example is called “Bosa Nova Party Robot.” It runs about 5 minutes
and 21 seconds, and the organ part moves between syncopated accompaniment,
short answers, heavier pulses, featured leads, and complete silence.

The basic idea is straightforward: play the recording through a mixer while
sending a separately prepared MIDI arrangement to the Lowrey. The instrument
produces its sound during playback, and I can work the balance and effects
at the mixer.

Getting that to work musically involves much more than connecting a MIDI
cable. The interesting part has been deciding what the organ should
contribute, how to keep it together with the recording, and when to let the
recording carry the song by itself.

Here is how this particular song was put together, and how the surrounding
system works.

What came out of Suno

I used Suno for the original ZIP of WAVs and MIDIs. For this song, the
source archive contained 11 instrument/vocal groups, each with WAV, MP3, and
MIDI exports:

- Lead vocals and backing vocals.
- Bass, drums, and percussion.
- Guitar and keyboard.
- Synth, strings, brass, and FX.

That gave the preparation process separate audio to inspect and separate
note data to work from. The MIDI exports were useful source material for
timing, pitch contours, bass movement, and harmonic evidence.

There was no separate full mix in this archive. The backing recording
therefore had to be reconstructed from the 11 aligned lossless WAV stems.
The initial reconstruction summed them at equal gain and applied a global
reduction of about 3.88 dB to control the resulting peak. A later playback
revision added 2 dB with a safety limiter. The earlier WAV was retained.

That distinction matters when describing the demo: the backing is a
reconstruction from the supplied stems, with a documented level revision. I
would not call it an untouched original stereo master.

The source MIDI also needed some housekeeping. Invalid key-signature
metadata was handled in separate analysis copies so the files could be read.
The original exports were retained. This repair was about making the
metadata usable while preserving the timing and notes for analysis.

The organ gets an arrangement of its own

The musical direction for the project is a recorded song featuring a strange
live organ arrangement. I want the Lowrey to have entrances, exits,
personality, and moments where it takes over.

An earlier approach in this project, called “Sentient Lower,” concentrated
on developing the lower-manual accompaniment while preserving established
upper and bass material. “THE ORGAN IS REAL” was the reference for that
approach, and that reference was approved on the physical organ.

The newer preparation workflow lets all the organ voices be rearranged
around the song. Each region has an assigned job:

- REST: release the organ and leave space.
- BED: add a sparse harmonic layer.
- GROOVE: establish rhythmic accompaniment.
- ANSWER: put short responses into gaps.
- BUILD: increase activity toward an entrance or feature.
- TAKEOVER: give the organ a prominent statement.
- PUNCH: use short, separated accents.

Those decisions are written into a timestamped map before the final MIDI is
rendered. The map for this song contains 22 regions, including the opening
and ending.

The region names are our arrangement labels. They describe what the organ is
intended to do at each point; they are not sections supplied in a formal
score.

How “Bosa Nova Party Robot” uses the Lowrey

The arrangement starts with eight seconds of organ silence. That gives the
recording's opening room to introduce itself. Sparse lower harmony follows,
and then the lower part moves into syncopated chord figures.

During the bossa-oriented passages, the lower part uses alternating beat and
offbeat attacks with short articulation. During builds and featured
sections, that changes to more insistent chord pulses. The intensity comes
from rhythm, articulation, and changing density as well as velocity.

The upper part has two main sources of material. Featured leads use selected
guitar, synth, and brass MIDI contours, adapted into the established upper
range and adjusted against the inferred harmony. The answers use fragments
of the preceding vocal contour, transformed into short responses in detected
quiet windows.

For those answers, the preparation script looks at the vocal audio for low-
energy gaps and combines that evidence with nearby vocal MIDI. It is a
concrete way to find possible response spaces. A quiet window is still an
inference about the audio, so it needs musical judgment when listening back.

The bass part selectively uses the supplied bass contour. Sparse sections
keep fewer attacks; bigger sections use more. The intent is to reinforce the
organ's role in that region without automatically following every bass note
in the source.

There is also a dedicated impact part for short upper chord hits, with lower
bass roots supporting selected accents.

The finished MIDI has a conductor track plus six musical tracks: vocal
answers, featured upper leads, lower bossa chords, lower dubstep pulses,
bass anchors, and brass/impact punches. Those six tracks are organizational
parts within two MIDI channels.

In this project, upper material goes to channel 1. Lower chords and bass
share channel 2. The arrangement uses MIDI note ranges of 60–83 for upper
material, 48–64 for lower chords, and 36–47 for the bass region. Those are
the working ranges used by this arrangement. The bass routing should not be
read as a claim that a separate physical pedal channel has been established.

Some moments to listen for

If you follow the arrangement map while listening, these are useful
landmarks:

Time              Intended organ role
0:00–0:08         Complete opening rest.
0:08–0:16         Sparse lower color.
0:16–0:32         First syncopated groove and selected answers.
0:32–0:40         First build.
0:40–0:56         First featured organ takeover.
1:20–1:36         Call-and-response region.
2:00–2:32         Extended central takeover.
3:04–3:20         Another organ feature.
3:20–3:28         Full organ dropout.
3:44–3:52         Short impact passage.
3:52–4:00         Another full dropout.
4:00–4:24         Final dance-floor takeover.
4:40–4:56         Percussion-focused feature.
4:56–5:12         Reduced lower texture.
5:12–5:20         Final build.
5:20–5:21.16      Organ release and recording tail.

The dropouts are part of the composition. After a prominent organ passage,
stopping it gives the next entrance a different weight. The MIDI checks
include the designated rest windows, because a held note crossing into a
rest can spoil the intended contrast even if there are no new attacks there.

The harmony and tempo work

No score was supplied with this song. Harmony was inferred from the exported
MIDI, using tonal note evidence from keyboard, synth, guitar, strings, and
brass, with bass evidence helping choose the root. The analysis produced 179
harmony windows.

That is useful working harmony for the arrangement, with the usual limits of
inference. It is not an authoritative transcription of the composition.

The supplied tempo map contains 719 tempo events, ranging from about 131.64
to 145.91 BPM. The arrangement retains that map. Its timestamped musical
decisions are converted back into MIDI ticks using the map, so those
decisions can coexist with the changing tempo.

The recording lasts 321.16 seconds. The supplied MIDI ends at roughly 320.38
seconds, so the prepared MIDI includes the remaining approximately
0.780-second audio tail without stretching the recording. The organ is
already resting during the final part of the song.

The player that ties it together

The local application is called nitetime-midi. It runs on Linux and uses
Python/FastAPI for the application, Mido/RtMidi for MIDI, sounddevice and
NumPy for audio handling, and a plain JavaScript browser interface.

A saved link associates a MIDI file with its WAV or MP3. Selecting that MIDI
starts the linked recording too. Queue playback, repeat, and random playback
use the saved pair as well.

The signal flow is:

Prepared recording → computer audio output → Yamaha MG-XU USB return
Prepared organ MIDI → USB MIDI interface → Lowrey → mixer audio input
                                             Mixer → capture / broadcast

The browser is the control surface. The audio runs on the computer connected
to the hardware, including when I control the page from another device.

This gives me separate control of the recording and organ at the mixer.
Riding faders and effects is part of how I use the system. The final balance
depends on that physical mix as well as the written MIDI.

Before paired playback starts, FFmpeg prepares the recording as 48 kHz
stereo PCM on temporary disk. The application memory-maps the current
recording, and the MIDI scheduler follows the audio output's DAC clock and
reported latency. That gives the two sides a common timing reference during
playback.

There is an adjustable start offset for a pair when needed. This song's
saved offset is zero. The application also supports an optional fixed MIDI
clock for external rhythm equipment, but that option is not enabled for this
pair. The changing tempo map inside the MIDI is a separate part of the
song's timing.

What the checks actually tell me

For this version, the saved MIDI validation report records 2,953 note
attacks: 367 on channel 1 and 2,586 on channel 2. Velocities range from 44
to 79. The recorded peak simultaneous note counts are three on channel 1 and
five on channel 2.

The report records no hanging notes and no overlapping instances of the same
pitch on the same channel. It also checks the intended rest intervals and
the duration. These are useful checks before sending the arrangement to
hardware.

Synchronization deserves a more qualified description. The analysis compares
supplied MIDI attacks with nearby detected audio onsets at several points
across the stems. The keyboard checkpoints have median nearest-onset
differences of roughly 44–172 ms, and the guitar checkpoints are roughly
45–78 ms. Bass checkpoints are roughly 154–245 ms.

Some other checkpoints are much worse, including multi-second outliers in
brass and vocal material. Matching the nearest detected onset does not prove
that the two events represent the same musical attack. The saved report
marks synchronization verified, but those outliers mean I would not describe
the analysis as proof of perfect alignment. It provides supporting timing
evidence, and listening remains necessary.

The player logs show this arrangement being started several times on
September 27, including 8:32, 8:38, and 8:41 in the morning. The preparation
documents still carry their earlier “physical audition not performed”
status. Playback history establishes that it was started; it does not
capture my listening impressions or certify the registration and mix.

The practical controls matter too

The player has Stop and Panic handling for the hardware, along with limits
on outgoing note velocity, MIDI volume, and expression. The current caps are
80. These are MIDI value limits; the physical instrument and mixer still
determine the acoustic level.

The application also has a remote organ panel with a learn process for
supported button messages, and an OBS overlay showing a real waveform from
the Yamaha audio input. The scope is driven by captured audio samples. It
gives me a view of the actual input signal during a session.

Those features make the project more useful as a performance and broadcast
setup. Preparing a song, controlling its playback, hearing the instrument
through the mixer, and seeing the captured signal all belong to the same
working environment.

Why I'm interested in this combination

The part I want to explore is what happens after a song has been generated.
Separate stems and MIDI provide material that can be selected, transformed,
and arranged for a physical instrument. The Lowrey then brings its own sound
and the mixer brings another layer of performance decisions.

For “Bosa Nova Party Robot,” the important choices include the first eight
seconds of silence, the syncopated lower part, the short vocal answers, the
heavier featured passages, and the full dropouts before the next entrance.
Each one changes the organ's relationship to the recording.

I'd be interested to hear from people doing similar things with home organs,
hardware synths, or other MIDI instruments. How are you turning generated or
exported song material into parts that give the hardware its own musical
role? And if you listen to this one, where would you give the organ more
space, and where would you let it get wilder?