Find the message, then the fix.
Most failed conversions are caused by one of six things: an unsupported or silent audio track, a section that is too short, a browser that cannot unpack the decoder, not enough memory for the file, a page that was interrupted, or audio with no clear pitched notes. The converter names the stage it reached, so you can match the text below to the cause instead of guessing.
Back to the converterThe six failures, in order of frequency
| What you see | What it means | What to do |
|---|---|---|
| “This file type is not supported.” | The extension is not on the accepted list. The check runs on the file name before any audio is read. | Export as MP3, WAV, M4A, FLAC, OGG, or one of the listed video formats. Renaming a file does not change its contents. |
| “No readable audio track was found.” | The decoder opened the container but could not decode an audio stream. Common with video files that only carry a picture, and with unusual codecs inside a familiar extension. | Check the file plays with sound in a normal player. Video containers are read directly, so extract the soundtrack as WAV only when the codec inside the container is the part that fails. The video page lists the containers that are accepted. |
| “The selected section is too short or empty.” | After trimming, less than a tenth of a second of audio remained. | Move the start earlier or the end later. Leave the end field blank to use the rest of the file. |
| “No clear pitched notes were found.” | The model ran, but nothing crossed its confidence threshold. Speech, drums, silence, very quiet recordings and heavily compressed material behave this way. | Choose a section with a sustained melody, raise the level in your editor, and avoid sections that are only rhythm or talking. |
| “This browser cannot unpack the local decoder.” | The page needs the DecompressionStream API to expand the bundled decoder. | Use a current Chrome, Edge, Firefox or Safari. Older browsers and some in-app browsers do not provide it. |
| “The media decoder could not be downloaded.” | The first conversion fetches about 10 MB of decoder data. A dropped connection or a blocking extension stops it. | Check the connection, allow requests from midilace.com, and retry. The download is cached after the first success. |
The bar stops moving on a long file
The converter shows the stage it is in — decoder, decode, model, notes — instead of a time estimate. On a long recording the model stage can dominate, because every second of audio becomes many frames of pitch evidence to score. Two things help immediately:
- Convert a section, not the whole file. Use First 30 seconds or type a start and end time. The exported MIDI then starts at the beginning of your section, which is what you want when you are testing whether the tool suits a recording.
- Press Cancel and try a shorter section. Cancel stops the running worker instead of waiting for it. Nothing is left half-written: the page returns to the input state and you can convert again.
Memory is the other limit. There is no fixed file-size cap in the page, but the browser must hold the file, the decoded audio and the model at once. A phone with little free memory can fail on a file that a desktop handles without complaint. If a large file fails silently on a phone, move to a desktop browser before assuming the recording is unusable.
On a phone the wait is real but not a failure. We ran an 8.7-second vocal clip on the live site in a phone-sized viewport with the processor slowed to a quarter speed, the way a mid-range handset behaves: the conversion finished in 135 seconds, the note preview and the MIDI download both worked, the layout did not overflow sideways, and the page logged no errors. Expect a phone to take two to three times as long as a laptop for the same clip, and use a short section rather than a whole song.
Correct results that still need editing
A conversion that finishes is an estimate, not a transcription. Three patterns are normal and are fixed in your MIDI editor rather than by converting again:
| Pattern | Why it happens | Edit |
|---|---|---|
| A sustained note appears as two or three blocks | Falling energy at the tail drops below the threshold, so the note is closed and reopened. | Select the blocks and merge them, or extend the first one to the intended end. |
| Extra notes an octave above a real note | The model also hears the second harmonic of a strong fundamental. | Delete the octave duplicates. They are usually shorter and quieter than the true note. |
| Notes arrive a few milliseconds early or late | Onsets are detected from the attack, and a soft attack is harder to place than a hard one. | Quantise to the grid if the performance is rhythmic, or nudge the whole passage. |
| Chords collapse into one or two notes | Simultaneous pitches share harmonics and mask each other. | Record or convert chord voicings one at a time when the part matters. |
Too many notes, or too few? Change the sensitivity before you edit
The control above the Convert button sets how confident the model must be before an event becomes a note. It is the first thing to try when the note list looks wrong, because it changes what the model accepts instead of deleting notes afterwards. Measured on our published 12-second guitar excerpt:
| Setting | Notes exported | Real notes found | Spurious notes |
|---|---|---|---|
| More notes | 113 | 42 of 45 | 71 |
| Balanced (default) | 68 | 41 of 45 | 27 |
| Fewer, cleaner | 55 | 39 of 45 | 16 |
If the result is cluttered with notes you cannot hear in the original, switch to Fewer, cleaner and convert again. If quiet, legato or sustained notes went missing, switch to More notes. The measurements behind these numbers explain why deleting duplicate notes by hand is often the wrong first move.
The right setting depends on the material. On a solo voice the clutter is mostly vibrato, so the quieter setting is often the better default: the same 8.7-second take gave 70 events with 27 real notes found at More notes, 47 events with 25 found at Balanced, and 36 events with 23 found at Fewer, cleaner. On a dense piano arpeggio the problem is the opposite of missing notes — the export was roughly 1.7 times too busy — so moving towards Fewer, cleaner is the first thing to try there as well. The measured cases are on the voice and piano pages.
Then fix the rest by hand, on the roll
After converting, the estimated notes appear in a piano roll. Click a note there to remove it; a note that is two short blocks is easier to delete than to merge by hand. The MIDI download follows your edits, and an Undo removal button restores the last note. Notes with an orange dashed border sit on the same pitch with a short gap — they are often sustain tails the model split into two blocks, and they are the first candidates to delete. This is the only correction the model cannot make for you: whether a vibrato wobble is one note or two is a judgement call, and you can hear the difference while the algorithm cannot. We tried to automate it — a minimum-note-length filter at 80 ms removed four spurious notes from the voice take but also deleted two real ones — so letting you decide in the roll is safer than a rule.
Our published test files show these patterns with the input and the exported MIDI side by side, so you can see what an unedited result looks like before you trust one. See the test file library.
When "no clear pitched notes" is the correct answer
Drums are the clearest case of a capability limit rather than a failure. The converter estimates note pitches, and a kick, snare or hi-hat is broadband noise with no stable pitch — there is nothing for a pitch estimate to hold. That is why the message above, "No clear pitched notes were found," names drums alongside speech and silence as material that behaves this way.
Because the limit is the task, the sensitivity control does not rescue a drum take: it changes how confident the model must be before an event becomes a note, not what kind of content it can detect, and noise-shaped energy never becomes a pitch. What the tool measures in a full-band recording is the pitched material — a bass line, a keyboard comp, a held tone — and that is what to convert when the drums are the reason you opened the page.
Transcribing drums is a different problem with a different output: onset times and instrument classes instead of note events. Some commercial transcription services list drums among the instruments they support; if the drum part is the point, that is the category of tool to look at, and the MIDI it produces is a percussion map rather than a note list.
Questions that come up during a failed conversion
Does converting the MP3 to WAV first help?
Sometimes, for a specific reason: it removes a decoding step and any doubt about the container. It cannot restore notes that were never clear in the mix, and it does not improve a noisy recording. If the MP3 decodes and the model returns notes, the extra conversion usually changes nothing.
Why does the first conversion take so much longer than the second?
The first run downloads the decoder and the note model, then unpacks the decoder in memory. Later conversions in the same tab reuse them. Reloading the page clears that cache.
Is my file uploaded when a conversion fails?
No. The page has no upload endpoint; decoding and note detection both run inside the tab. A failure is a local failure, and the recording never leaves the device.
Can I retry the same section without choosing the file again?
Yes. Change the start or end time and press Convert. The selected file stays loaded, and the previous result is cleared when the section changes.