How to Prepare Audio for Accurate Music Transcription

Music transcription starts before a file is uploaded. A clear recording does not guarantee a perfect score, but a clipped, noisy, incomplete, or heavily layered recording forces any transcriber—human or automatic—to guess. A few minutes of input preparation can prevent hours of correcting false notes and broken rhythms.

The goal is not to make a commercial master. It is to make the target musical events easy to distinguish: clear attacks, stable pitch, enough context to find the beat, and as little competing energy as practical.

Choose the best available source

Start with the least processed, most isolated version you legally control. Preference usually runs in this order:

1. a direct track or stem for the target instrument;
2. a close live recording of one performer;
3. a rehearsal mix with the target prominent;
4. a mastered stereo mix; and
5. a phone recording made from a speaker.

Every step down adds ambiguity. A mastered mix combines instruments, effects, compression, and stereo placement. Playing that mix through a speaker and recording it again adds room reflections, device coloration, background noise, and a second conversion.

Do not convert a low-quality file to WAV and expect lost detail to return. Changing the container or sample format can make a file compatible with a tool, but it cannot reconstruct frequencies and transients removed by an earlier lossy encode.

Isolate the musical target

Automatic note detection works best when one intended source dominates. Spotify notes that Basic Pitch supports polyphonic instruments but works best on one instrument at a time. In a full arrangement, unrelated instruments can produce onsets and harmonics that resemble notes in the target line.

If you are recording a new performance:

• record the melody alone when possible;
• mute television, fans, notifications, and metronome speaker bleed;
• use headphones for accompaniment;
• place the microphone closer to the target than to the room;
• keep the performer in one position; and
• avoid touching the device or stand during the take.

If you already have a multitrack session, export the relevant stem. For a vocal melody, a vocal stem is more useful than the complete master. For a guitar solo, export that track without the drum bus and mastering chain when available.

Source-separation software can help when stems do not exist, but listen to the isolated output before using it. Separation may leave watery artifacts, missing attacks, or fragments of other instruments. Compare uncertain passages with the original mix.

Set recording level before the take

Clipping occurs when the input exceeds the system’s available level and peaks are flattened. Once clipped, different loud waveforms collapse to the same maximum value. That distortion introduces harmonics and destroys attack shape, both of which can confuse pitch and onset detection.

Audacity’s support guidance recommends monitoring the loudest expected passage before recording and lowering the input when the meter approaches the yellow or red region. Its manual gives roughly -6 dB as a practical target for a critical recording. The exact number is less important than leaving real headroom.

Run a test:

1. Set the final microphone position.
2. Play the loudest section at performance intensity.
3. Watch the input meter, not the playback volume.
4. Lower input gain if peaks approach 0 dBFS.
5. Listen through headphones for crackle or harsh distortion.
6. Record 15 seconds and inspect it before the full take.

Do not normalize or amplify a clipped recording and call it repaired. Lowering the level afterward only makes the distorted signal quieter.

Keep enough musical context

A ten-second highlight may contain the notes you want, but it can omit the evidence needed to interpret them. Beat tracking benefits from a stable lead-in. A pickup needs the following downbeat. A tied note needs its attack. A key may be clearer when the preceding cadence is present.

Include:

• a short lead-in before the first target note;
• the full pickup measure;
• at least several measures of stable pulse;
• the complete decay of the final note; and
• adjacent phrases when they clarify repeats or harmony.

Avoid cutting exactly on a transient. A hard edit through a waveform can create a click that looks like an onset. Use a zero crossing or a tiny fade when an editor provides that option, but do not fade away the actual first attack.

Long files are not always better. Uploading an hour-long rehearsal to transcribe one chorus increases processing and review scope. Make a deliberate excerpt with enough context, then name it clearly.

Avoid unnecessary processing

Aggressive cleanup can remove musical information. Strong noise reduction may produce metallic artifacts around sustained notes. A gate can cut quiet attacks or note tails. Heavy compression can raise background bleed between phrases. Stereo widening and reverb make sources less distinct.

Use the lightest processing that solves a known problem:

• high-pass only when low-frequency rumble is unrelated to the target;
• gentle noise reduction only after auditioning artifacts;
• no limiter unless preventing a new export from clipping;
• no reverb for a transcription reference;
• no pitch correction if the goal is to document the original performance; and
• no time stretching unless the transcriber explicitly expects it.

Slowing playback can help a human reviewer, but submit the original-speed file to a system unless its documentation says otherwise. Time stretching changes transients and may introduce spectral artifacts.

Pick a sensible file format

Lossless WAV or FLAC is a strong choice for a new recording or exported stem. High-quality M4A or MP3 can be acceptable when it is the best available source. Avoid repeated lossy re-encoding: editing an MP3 and exporting another low-rate MP3 compounds compression artifacts.

Confirm that:

• the file opens and plays from beginning to end;
• left and right channels are both present as expected;
• sample rate does not change mid-file;
• the export contains the intended take;
• there is no accidental silence-only channel; and
• the filename identifies song, section, instrument, and version.

A name such as harbor-lights-chorus-vocal-take2.wav is safer than audio-final-new.wav.

The correct output format for the score is a separate decision. See MusicXML vs MIDI vs PDF for the preservation trade-offs after transcription.

Make performance choices explicit

Some ambiguity is musical rather than acoustic. Tell the transcriber what to follow:

• lead vocal or accompaniment;
• melody only or all chord tones;
• concert pitch or a transposing instrument’s written pitch;
• standard tuning, drop tuning, or capo position;
• straight or swung feel;
• desired notation view; and
• whether expressive bends should become discrete notes.

A clean full mix can still support several valid transcriptions. A singer’s line, bass line, keyboard voicing, and guitar tab are different targets. Naming the target prevents the system from optimizing the wrong output.

For guitar, the same pitch can have multiple fretboard positions. The comparison of guitar tab and standard notation explains why an audio file may not uniquely determine fingering.

Run a pre-upload listening pass

Listen once without multitasking. Use headphones and speakers if both are available. Check:

• Does the target remain louder and clearer than competing sources?
• Are any loud attacks clipped?
• Do notes disappear under noise reduction or a gate?
• Is the pulse audible long enough to estimate tempo?
• Does the excerpt include the pickup and final decay?
• Are there unexpected count-ins, talkback, clicks, or notification sounds?
• Is the take in the key and tuning you expect?
• Do file duration and filename match your notes?

Then write down known trouble spots with timestamps. “The vocal is masked by cymbals at 0:42” is actionable. It tells a reviewer where confidence should be lower and where the original mix may be needed.

Review the result against the source

Good input reduces uncertainty; it does not eliminate review. Listen to the generated playback while following the notation. Check entrances, repeated notes, sustained notes, octave jumps, and measures with dense accompaniment.

Rhythm deserves a separate pass. The article on automatic transcription rhythm errors explains why a system can find plausible pitches but still require beat alignment and quantization cleanup.

VerseScore can turn a prepared recording into several readable views, but the best workflow remains source-aware: isolate what matters, preserve headroom, keep context, avoid destructive processing, and verify the output. Clear evidence produces fewer guesses—and a score that is much faster to trust.

Sources

• Spotify Basic Pitch input notes and limitations
• Audacity Support: setting recording levels
• Audacity Manual: recording