“Cut the bass from the mix”
ProCraft: Instruction-Driven Music Production from Free-Form Text
Audio examples for the paper. ProCraft is a 0.5B codec language model that edits a 30-second mix from a freely phrased instruction: removing, extracting, accompanying, covering, re-instrumenting, denoising, and applying audio effects.
All systems run on the same test items with their released weights. Clips are loudness-normalised mono, 30 seconds.
All eight instrument families
The test sets of the paper target bass, drums, guitar and piano. These examples cover all eight families the model is trained on, from the Slakh test split: one item per task and family. Cover targets the four wide-range instruments and instrument transfer the seven pitched ones.
“Eliminate the drum elements from the audio.”
“Eliminate the guitar from the audio mix.”
“Remove piano from the mix”
“Eliminate the organ completely.”
“Eliminate brass from the audio mix.”
“Cut the reed frequency band.”
“Remove ensemble from the mix.”
“Remove all instruments except the bass”
“Extract drums to separate output channel.”
“Extract only the guitar track.”
“Isolate the piano track”
“Leave just the organ playing.”
“Cut all other instruments, leave only brass.”
“Isolate the reed and leave the rest.”
“Leave just the ensemble playing.”
“Design a bass line to match the track's rhythm and harmony.”
“Generate a fitting drums part for the song.”
“Generate a fitting guitar part for the track.”
“Generate a piano part that complements the track”
“Insert organ part matching track's harmony and rhythm”
“Design a brass part that complements the track's harmony.”
“Create a reed part that complements the track's harmony and rhythm.”
“Write a complementary ensemble part for the track.”
“replay the full song using only guitar parts”
“Arrange entire track as solo piano performance”
“Reimagine the full track as a solo organ performance.”
“Voice the full track using ensemble texture”
“Replace guitar with bass using same notes.”
“Swap the piano's sound with the guitar's timbre.”
“Transpose the ensemble's timbre to piano.”
“Replace guitar with organ, same notes, different tone”
“Turn the reed into a brass while keeping the same notes.”
“Swap bass for reed timbre, retain notes”
“Swap piano for ensemble, same notes, new tone.”
“Filter out the dripping water frequency”
“Eliminate the ambient breeze audio”
“Minimize ambient interference”
“Eliminate all non-music sounds.”
“Make it gritty and intimate.”
“Make it twangy and dusty.”
“Make it punchy and tight.”
“Add rich tube saturation and hall reverb to the piano.”
“Add bright saturation with tremolo and delay to the organ.”
“Add a deep low-pass with warm saturation to the brass.”
“Add a bright high-pass filter and a subtle saturation to the reed.”
“Add gritty tube drive with shimmering delay to the ensemble.”
Edits from free-form instructions
Six arrangement tasks and audio effects on both test domains, against the task-specific systems of the main results figure and, for audio effects, three text-to-effect systems. Systems marked ★ are deterministic: they separate or process the source instead of generating new audio.
“Remove guitar from stereo mix.”
“remove piano from the mix”
“Eliminate the guitar from the audio.”
“Eliminate the bass layer”
“Retain just the bass layer.”
“pull out just the drums”
“Boost drums, cut all else.”
“Isolate the bass track.”
“create a solo piano part to accompany this track”
“Layer a supportive guitar line matching the song’s tonality and rhythm.”
“Layer a fitting guitar part that blends with the existing music.”
“Add expressive piano layer to support track’s harmonic structure.”
“Reimagine the full track performed solely on piano”
“Reimagine the full track as a solo guitar performance”
“Transpose full track into standalone piano arrangement”
“Transpose full arrangement into solo guitar performance”
“replace the bass with a piano”
“Replace bass with guitar track”
“Turn the guitar into a piano sound”
“Replace bass with piano timbre, same notes”
“Remove the pouring water noise”
“Eliminate the wind noise from the track”
“Cut the thunder noise”
“Minimize the ambient layer”
“Make it deep, cinematic, resonant.”
“Make it crisp and grounded.”
“Make it crisp and hi-fi.”
“Make it earthy and bare.”
“Make it smoky and intimate.”
“Make it bright and sunlit.”
Edits combined with audio effects
Instructions that ask for an arrangement edit and an audio effect together, compared with the four free-text editors. The MoisesDB items are the eighteen items of the listening test.
“Remove the guitar. Make it ethereal, dreamy, hushed, reverberant, slow.”
“Remove the bass. Add pumping compression to the piano, airy high-pass filter to the drums, and warm saturation to the guitar.”
“Remove the piano. Make it echo-drenched and cavernous.”
“Remove the piano. Add low-passed warmth to the bass, compressed punch to the drums, and reverb-drenched haze to the guitar.”
“Remove the guitar. Make it earthy, weathered, hushed.”
“Extract the bass. Add punchy compression with a subtle saturation to the bass.”
“Extract the drums. Add a gritty bitcrush haze with tight compression and rhythmic flanger motion to the drums.”
“Extract the guitar. Add a warm tube saturation with a gentle high-pass to the guitar.”
“Extract the drums. Make it rhythmic and spacious.”
“Extract the bass. Make it raw, intimate, earthy.”
“Design a standalone bass line matching the track's rhythm and harmony. Apply a warm low-pass filter with gentle saturation and subtle tape noise.”
“Create a standalone piano part that complements the track. Add gritty bitcrush with analog noise.”
“Generate a fitting drums accompaniment for the song. Put on a low-pass filter haze, fuzzy saturation, slow modulation sweep.”
“Generate a standalone guitar accompaniment that fits the track. Put on dusty color noise with slow distortion.”
“Generate a complementary drum pattern for the track. Add compressed punch, tamed top-end, layered delay.”
“Reconfigure full track for solo piano performance. Put on tamed low-end, boosted mid-range, tight delay.”
“reimagine the full track as a solo guitar performance. Apply a lo-fi tape saturation with echo delay.”
“Transpose the full track into a solo guitar version preserving melody and harmony. Put on analog warmth with gentle noise and low-pass.”
“Arrange the full song as a solo guitar performance. Add warm bit-crush and tight delay.”
“Transpose the full song into solo piano format. Apply bitcrushed low-pass haze, boosted midrange grit, harsh low-pass filter.”
“Swap bass for piano, keep notes identical. Add gritty lo-fi bit crush with noise floor.”
“Change bass to guitar timbre, retain notes. Apply lo-fi saturation with slow flanger.”
“Swap the piano part for a guitar version. Put on a harsh high-pass cut with a tight low-pass.”
“Swap bass for piano, same notes, new tone. Add a soft bitcrush haze with dusty noise.”
“Transform guitar tones into piano timbre. Put on warm tube saturation with soft compression and gentle low-pass filtering.”
“Remove all unwanted audio elements. Put on vintage low-pass warmth, gritty analog noise, gentle compression.”
“Eliminate the non-musical rain audio. Apply subtle bitcrush haze, soft saturation, gentle noise floor.”
“Strip away the extra sounds from the track. Apply warm analog saturation with low-pass filtering and subtle tape noise.”
“Minimize ambient presence. Apply bitcrushed lo-fi warmth, short room ambience, tamed top-end.”
“Eliminate the background layer. Add tape hiss floor and bitcrushed grit.”
Failure cases
For each task on real audio, the test item on which ProCraft scores lowest by its task metric.
“Lower the bass frequency band.”
“separate out the piano”
“Craft a fitting guitar part for the song.”
“make a guitar cover of the whole song”
“reduce the background noise”