Help & FAQ
Pitch to Staff estimates musical notes from recordings on your device, then shows a grand staff and piano-key views.
Quick start
- Choose an audio recording from your device.
- Keep the tab open while the browser decodes and analyzes the recording. GPU acceleration is on by default; you can switch it off beside the upload control.
- In Automatic mode, compare A, B, C and their available combinations. Generate Enhanced transcription and Also generate Vocal Ensemble start checked; turn either off if you want a lighter run. Vocal Ensemble requires Enhanced.
- After analysis, you may also select Enhanced, Hybrid (A plus supported acoustic discoveries), and Vocal Ensemble when those versions were generated successfully.
- To fine-tune A, B, or C, select it, enter Manual, and adjust the transferred detection settings before re-analyzing. The experimental acoustic versions have no single equivalent Manual parameter profile.
- Compare the recording with the staff and piano views. Use original-audio playback and score playback to check pitches and attacks.
- Under Download and print, export the selected score as MIDI, MusicXML, or JSON, or print the staff.
Score navigation and corrections
Choose a detected note in Start measure 1 at to begin the displayed staff there. Earlier notes are still included in the piano change view, playback and exports. Use Staff only and Piano only to switch views at the current approximate measure or note event. A curved tie across a barline represents one sustained pitch, not a repeated attack. Estimated score corrections can repair some octave shifts and nearby attacks but can also be wrong. Use their toggles and Reset/original option to compare with the uncorrected score. Automatic detection can sometimes split or re-trigger a sustained note incorrectly; review questionable passages.
Transcription methods
A, B and C are alternative automatic interpretations of the same neural model output. They can disagree on soft notes and repeated attacks, so compare them by listening. Enhanced uses additional acoustic evidence to estimate note intervals, rhythm and pitch alternatives. Hybrid retains version A as its anchor and attempts to add supported acoustic discoveries. Vocal Ensemble explores simultaneous pitches at shared attacks and is intended for layered singing and other polyphonic material. All three additional methods are experimental and may add wrong notes, especially harmonics or overtones. None has a guaranteed accuracy ranking.
When you switch score versions, the current staff and piano viewing position is preserved as closely as possible. Piano cards share the full pitch range of the currently selected score. Switching to a different version may change that range because the versions contain different notes.
During analysis and on long recordings
Analysis runs in your browser and may be slow on a phone, particularly when Enhanced and Vocal Ensemble are both enabled. A compact progress banner appears above the score. You can choose Cancel analysis, including from the mobile progress banner when the settings drawer is closed. Cancellation can be delayed while a model-processing step finishes. During re-analysis, the previous completed score stays visible and should be restored if the new analysis is cancelled or fails.
Enhanced, Hybrid, and Vocal Ensemble are no longer restricted to the first 60 seconds, but longer files are more demanding. Results beyond 60 seconds depend on your device completing the full inference. Files larger than 128 MiB are rejected before decoding as a memory precaution; a smaller compressed file can still decode into a large recording. Try a shorter excerpt or a desktop browser if processing runs out of memory. Keeping other demanding tabs closed may also help.
Frequently asked questions
Which audio formats work?
Common formats include MP3, WAV, M4A, and OGG when supported by your browser. Support varies across browsers and devices. Try WAV or MP3 if your file cannot be decoded.
Can I use the app without internet after my first visit?
Yes, on a supported HTTPS browser after the first online visit finishes preparing its offline files. Look for Offline ready · works without internet below the site heading before disconnecting. Then the local transcription model, A/B/C, Enhanced, Hybrid, Vocal Ensemble, Manual controls, staff and piano, playback, downloads, and internal help pages can work without a connection. The app caches website files and the neural runtime; it does not save your recordings or scores for later sessions. Keep an original audio file on your device if you plan to transcribe it offline. External contribution links and email require their respective services and a connection. Clearing site data, private-browsing restrictions, browser storage eviction, or an incomplete first-time download may require visiting online again.
Is my audio uploaded?
No. This build reads and processes selected recordings in your browser. Its transcription scripts and neural model load from this website. The app does not upload your audio file for analysis.
How accurate is transcription?
It is approximate. Dense arrangements, percussion, background noise, reverb, and overlapping instruments can cause wrong pitches, octaves, durations, chords, or timing. Always check the score before using exported files.
What recordings give the best results?
Start with clear isolated piano, guitar, or other pitched-instrument recordings. Reduce background noise and excessive reverb.
What do sensitivity, background filtering, and Recover soft attacks do?
Higher sensitivity can preserve quieter notes while introducing more false detections. Stronger filtering rejects more weak detections but may remove valid quiet notes. Recover soft attacks also uses changes in sustained pitch energy to find attacks the onset detector may miss; turn it off when noisy material produces too many extra notes.
What do BPM and time signature change?
They affect the notation layout and MIDI tempo/meter metadata, not the detected pitches or underlying performance note timings. In Find a notation tempo, tap the beat at least three times to get a BPM estimate, or after analyzing choose Suggest from score to see tentative suggestions based on detected attacks. Press a Use BPM button to apply a tempo; nothing changes automatically. A weak or irregular pulse may produce no score suggestion, and half-time/double-time options can both sound plausible. The meter must still be selected manually.
Which versions transfer to Manual mode?
A, B and C are decoder-derived versions. Choose one before switching to Manual to copy its exact detection parameters. Enhanced, Hybrid and Vocal Ensemble combine acoustic hypotheses and cannot be represented by one Manual threshold profile. Switch to A, B or C to adjust decoder settings.
Why does Vocal Ensemble show extra notes?
It searches for overlapping musical fundamentals, but a strong harmonic or overtone can be mistaken for another note. Compare against the original recording. More notes in a score do not automatically mean better accuracy.
Should I use GPU acceleration?
GPU/WebGL is enabled by default and can be faster. Some devices produce different notes on GPU. Turn it off to request CPU processing. If compatible GPU initialization fails, the app should fall back to CPU; report devices where this does not work.
Can score corrections make mistakes?
Yes. Octave and nearby-attack corrections are estimates. Review the recording, toggle corrections off or use Reset/original when you want to compare with the uncorrected detected score.
What do MIDI, MusicXML and JSON export?
All three reflect the currently selected and adjusted transcription rather than the original recording. MIDI carries notes and timing for a DAW or notation application. MusicXML carries editable staff notation for software such as MuseScore, although you should check the import and engraving in your editor. JSON contains structured note and diagnostic information. Printing uses your browser's print dialog. Check any result against the audio before sharing it.
What is MIDI?
MIDI stores pitches, note starts, note ends, and timing information. It is not the original audio recording. Open the downloaded .mid file in a compatible digital audio workstation or notation program.
Why is it slow or unresponsive on my phone?
Neural inference and the experimental acoustic versions use processing power and memory. Try turning off the optional Enhanced and Vocal Ensemble checkboxes for a lighter A/B/C run, closing other tabs, using a shorter recording, disabling GPU if GPU initialization fails, or trying a desktop browser. A longer recording might take significantly more time even when the file itself is small.
Can I stop a long analysis without losing my last score?
Use Cancel in the analysis controls or the progress banner on mobile. The app checks cancellation between processing steps, so the button may not stop the current model window immediately. A completed score from the prior analysis remains on screen while re-analysis runs and should be restored on cancellation.
Can I transcribe copyrighted music?
Use recordings you own, have permission to use, or are otherwise legally authorized to process. The tool does not grant rights to publish, distribute, or sell derived sheet music or MIDI.
Does contributing change my results?
No. Contributions are entirely voluntary and do not unlock features or change transcription quality.
Troubleshooting and bug reports
If audio cannot be loaded, verify it plays locally, try WAV or MP3, reload the page, or try a modern browser. If neural analysis fails, try disabling GPU acceleration. Keep the vendor/ and model/ folders intact when deploying.
When reporting a problem, provide your device, browser/OS, approximate recording duration, what happened, what you expected and steps to reproduce. Screenshots or error descriptions are helpful. Please do not send the original recording or note export unless you choose to share them. Email support@pitchtostaff.com or use the contact page. The address must be verified by sending and receiving a real test message before the support requirement is marked complete.