What Pitch Bend Detection Adds to Audio to MIDI (and When It Actually Matters)

The short answer: pitch bend detection keeps a continuous glide as a real, smooth pitch-bend curve in the MIDI instead of rounding every frame to the nearest equal-tempered note. That is the difference between a guitar bend sounding like one elastic note and a bend sounding like a staircase of five stiff semitones. It matters for instruments that actually bend — electric guitar, fretless bass, violin, slide guitar, most sung notes with vibrato — and it barely matters for piano, drums, or tightly quantized electronic tracks, which were already close to the grid. The catch is that a standard MIDI pitch bend is a single value per channel, so it only behaves when one voice owns the channel; a polyphonic chord where two notes bend in different directions cannot be represented that way.

What a pitch bend actually is

A note in equal temperament sits on a fixed ladder: A4 is exactly 440 Hz, B4 is 493.88 Hz, and everything between them is "not a note." But real instruments do not stay on the rungs. A guitarist pushes a string and the pitch slides up a quarter of a step and back; a violinist's finger never quite stops; a singer's vibrato wobbles a few cents around the centre. That continuous wobble is pitch bend — a deviation from the nearest written pitch that is smaller and smoother than a full semitone.

When a converter ignores it, the only thing the writer can do is pick the rung each frame is closest to. A note that slid a third of a step upward becomes a short higher note stuck on the end; a vibrato becomes a row of separate notes trembling around the centre. That is not wrong as data, but it is not the sound you played.

How a converter turns a glide into MIDI

A frame-based detector already solves half the problem: at every step it estimates a frequency in hertz, not a note name. To keep the bend, the writer does two extra things instead of rounding:

Pick a centre note. It chooses the equal-tempered pitch the track is closest to overall — say, the note C.
Store the deviation, not the rounded value. For each frame it works out how far the measured frequency sits above or below that C in cents, then writes that offset as a MIDI pitch-bend message rather than snapping the frame to C or to a neighbour.

The result is one note-on, one note-off, and a stream of bend messages in between that trace the curve. The rhythm still lands on frame boundaries, but the pitch inside the note is smooth. A few converters take this route — Spotify's Basic Pitch, for example, tracks pitch continuously and writes the offset from the nearest note as bend events rather than locking each frame to a key — while many browser and plugin converters still emit only note-on and note-off and leave the curve as a staircase.

The numbers that decide whether it is smooth

Two separate resolutions are in play, and only one of them is tight enough to worry about.

1. The MIDI bend resolution. A pitch-bend message is 14-bit: a value from 0 to 16383 with 8192 in the middle meaning "no bend." The default sensitivity a synth assumes is ±2 semitones, and a semitone is 100 cents, so the full travel is ±200 cents across 8192 steps. That is 200 ÷ 8192 ≈ 0.024 cents per step. The smallest pitch change a human can hear is around 5 cents, so the MIDI side has roughly 200× more resolution than you could ever notice. The bend channel is never the bottleneck.

2. The detector's frame rate. This is the real limit. If the analyser steps through the audio at, say, 86 frames per second, the bend curve is sampled 86 times a second. A typical vibrato runs at 5–7 Hz with a depth of 30–60 cents; sampling a 6 Hz wobble at 86 Hz clears Nyquist by more than 7×, so the wobble survives intact. A fast trill or a sub-100 ms flick of the wrist, though, is shorter than the frame spacing and will be smoothed or missed — not because MIDI is coarse, but because the detector only looks 86× a second.

Where the bend resolution actually lives. The MIDI value is far finer than human hearing; the frame grid is what rounds the curve.
Stage	Resolution	Versus human JND (~5 cents)
MIDI pitch bend (±2 st, 14-bit)	0.024 cents / step	~200× finer
Detector at 86 fps on a 6 Hz vibrato	~14 samples / wobble	smooth, but finite
Detector at 86 fps on a 40 ms flick	~3 samples	blurred or lost

When it matters

Pitch bend detection earns its keep whenever the performance lives between the notes:

Electric guitar bends and slides. A whole-tone bend is the textbook case — without bend tracking it becomes three or four stacked semitones instead of one elastic note. This is why a bend-heavy take and a bend-aware converter are a poor match for a note-only writer; see what a bend does to your notes on guitar.
Vocals and bowed strings. Sung notes sit slightly sharp or flat and carry vibrato constantly. Bend tracking preserves the human wobble; rounding flattens singers into a parody of themselves.
Fretless and slide instruments. There is no grid at all — every note is a position on a continuum — so without bend data the MIDI is a permanent approximation.

When it does not matter

Save your attention for the cases above; the rest of your library will not thank you for bend tracking:

Piano and other struck keys. Each note is already a fixed pitch struck once. There is no curve to keep.
Drums and percussion. Rhythm, not pitch, is the content; a bend message on a kick is meaningless.
Quantized electronic tracks. If the source was drawn on a grid in a DAW, it is already note-shaped and pitch-stable.
Heavily detuned or noisy material. A bend curve traced through noise is just a noisy curve; the cleaner win there is a shorter note setting, not bend tracking.

The monophonic trap

This is the part most people miss. In standard MIDI, pitch bend is a channel message: one value, applied to every note currently sounding on that channel. There is no per-note bend. So:

If a converter puts a single melodic line on its own channel (the normal case for a lead or a vocal), bend tracking works perfectly — the whole channel bends together, which is exactly what the instrument did.
If two notes on the same channel bend in different directions at the same time, standard MIDI cannot show it. One bend value has to win, and the other note is wrong.
In a polyphonic chord where different voices independently slide, the format itself is the limit — no detector, however good, can fix it without a non-standard per-note extension that most players ignore.

That is why bend detection and monophonic or single-voice conversion go together, and why it falls apart in dense polyphony. The trade-off is laid out in what each detection mode misses.

What a bend looks like when your converter does NOT support it

If the tool only writes note-on and note-off, every bend becomes a staircase. Concretely, a one-second whole-tone guitar bend (200 cents) reported at 86 fps with a note change every semitone (100 cents) comes out as roughly two to four short notes stacked end to end, each a fixed pitch, with silent or overlapping gaps between them. A vibrato becomes six to nine fragments around the centre note. The melody is still recognisable; the expression is gone, replaced by a tremolo.

The practical tell: open the MIDI in a piano-roll editor. A true bend shows as a sloped line on the pitch axis; a missing bend shows as a column of separate blocks. If you see blocks where you played a glide, your converter rounded it.

How to tell if your result kept the bend

Convert a short bend-heavy clip (a guitar bend, a sung "ahh" with vibrato).
Open the exported MIDI in any DAW or piano-roll view and look at the pitch trace, not the note count.
A kept bend is one note with a sloped or wavy pitch line. A lost bend is several rigid blocks.
If your writer supports it, confirm a pitch bend event exists in the file rather than assuming — some players draw a smooth line even when the underlying data is stepped.

One more caveat: a bend written correctly can still sound flat if the instrument on the other end ignores the message. Many samplers and drum machines need pitch bend enabled, and some ignore it entirely. The bend has to survive both the writer and the player.

Questions people actually ask

Does audio to MIDI keep vibrato?

Only if the converter writes pitch bend. With bend tracking, vibrato survives as a wavy pitch line inside one note. Without it, the same vibrato becomes a row of separate notes trembling around the centre — audible as a machine-gun tremolo, not as human warmth.

Is a bend the same thing as portamento?

Related but different. A bend is a continuous offset from a fixed note; portamento is the glide between two separate notes. Both can be stored as bend messages, but a converter has to decide which frames belong to one note and which are a transition — and that decision is exactly where bend detection either shines or smears.

Why does my converted solo sound stiff even though the notes are right?

Almost always missing bend and velocity expression. The note names can be perfect and the performance still feels robotic because every micro-glide was rounded to the grid. Bend tracking is the half of the sound most note-only converters throw away.

Can polyphonic conversion keep bends on every voice?

Not with standard MIDI. Because bend is one value per channel, a chord where several voices slide independently cannot be represented faithfully — at best each voice gets its own channel and they bend consistently, at worst one bend value has to cover notes that wanted different ones.

Will the bend survive into the audio when I play the MIDI back?

Only if your playback instrument responds to pitch-bend messages. A synth with bend enabled will glide exactly as recorded; a sampler with bend turned off, or a drum machine, will play the fixed centre note and ignore the curve. Check the receiving instrument, not just the file.

Convert a file now — runs in your browser, nothing is uploaded.

What pitch bend detection adds to audio to MIDI (and when it actually matters)