Sunofriend

Listen deeper. Create further.

Sunofriend turns separated music stems into editable MIDI and a balanced MIDI-derived WAV. The WAV is a clean musical interpretation that can make a song's rhythm, harmony and melody easier to hear; the MIDI can then be changed in GarageBand or another DAW.

It is local-first, open source and designed to compare several analytical and AI transcription methods rather than pretending one method always wins.

Sunofriend is an independent Unsigned Media Ltd project. Its name comes from the Hindi सुनो (suno), “listen.” It is not related to, affiliated with, endorsed by or sponsored by Suno Inc.

Easiest start: ask Codex

You do not need to clone the repository or understand Python first. Use Codex with access to a local folder on your Mac and follow the copyable prompts on the beginner website, or send them directly:

Use $skill-installer to install the Sunofriend skill from
https://github.com/N9-Developer-Empowerment/sunofriend/tree/main/skills/sunofriend.
Do not install the application yet. Tell me when the skill is available.

Then start a new message:

Use $sunofriend. I am new to music software. Help me choose between:
1. trying the built-in demo,
2. experimentally separating one finished song locally,
3. using stems I already have, or
4. getting stems I am allowed to process.
Explain one thing at a time. Inspect my Mac and show me an installation plan
before making system or network changes. Keep my audio local.

The skill should:

1. ask what you want to try;
2. inspect the Mac without changing it;
3. explain the exact setup plan in plain language;
4. ask once before preparing the source and again before installing the exact reviewed commit;
5. run the demo or process your stem folder; and
6. show the WAV, MIDI and ZIP in the order they are useful.

If the skill does not appear immediately, restart Codex and send the second message again. See AI-assisted first song for the complete newcomer journey.

No stems yet?

Choose the built-in demo. It creates a small copyright-safe synthetic song and runs the same automatic MIDI, WAV and ZIP workflow used for a real project. It is the quickest way to decide whether Sunofriend is useful before processing personal music.

Sunofriend also has an opt-in public experimental local separator for an authorised finished song on Apple-silicon macOS. Broad vocals plus complementary instrumental remains the default. The separately installed, immutable SCNet-large profile is now available as an explicit local vocals/drums/bass/grouped-other preview. Its objective offline canaries passed and its full-song outputs had no catastrophic defect reported; musical quality remains experimental. The working two-stem route produces a reconstruction check and local listening page. Read Experimental local stem separation before installing a separate pinned model/runtime or treating any estimate as useful.

Finished-mix separation currently has four distinct lanes:

The private synth and guitar models are not public profiles. Exact reconstruction demonstrates PCM accounting, not correct musical assignment.

Core-four preview available: after the separately approved SCNet setup, Sunofriend can estimate vocals.wav, drums.wav, bass.wav and grouped other.wav. Inspect the immutable profile first, then select the scope explicitly:

.venv/bin/sunofriend-separate profiles
.venv/bin/sunofriend-separate separate "/absolute/path/to/song.wav" \
  --scope core-four-stems-v1 \
  --out "/absolute/path/to/fresh-core-four-separation" \
  --rights-category owned

The command above is a read-only plan until --execute --confirm-rights is added. It never starts MIDI/Create automatically, and grouped other is not a single instrument.

For deeper separation, Sunofriend now has an opt-in Studio challenger for refining one exact SCNet grouped-other stem into either one guitar target or one keys target plus the transparent residual. The Apple-native htdemucs_6s MLX profile is installed separately; the keys lane is honestly labelled as a piano proxy. Its fraction-normalized local loader, synthetic canary and both full-song target mappings passed the objective offline gates. The completed fixed five-song review found no demonstrated useful guitar extraction and no successful piano extraction. The technically valid profile remains reproducible in Studio, but it is not promoted, selected or fed to MIDI. Plan a run with:

.venv/bin/sunofriend-separate refine-other \
  "/absolute/path/to/core-four-separation" \
  --target guitar \
  --out "/absolute/path/to/fresh-guitar-candidate"

Add --execute --confirm-rights only after reviewing the plan. The command never activates the target or starts MIDI automatically. The deterministic PCM24 proof, setup plan and bounded qualification evidence are documented in Refining grouped other in Studio.

The completed bounded candidate broadened the target from piano to the modern keyboard_synth family: electric piano, organ, synth pad and synth lead. Banquet was audited as a query-conditioned local research challenger alongside guitar, with one configuration and no post-feedback query hunt. Its historical no-effects planning commands were:

.venv/bin/python \
  scripts/plan-separation-other-refinement-query-runtime.py

.venv/bin/python \
  scripts/plan-separation-other-refinement-query-forward.py

.venv/bin/python \
  scripts/plan-separation-other-refinement-query-synthetic-report.py

.venv/bin/python \
  scripts/plan-separation-other-refinement-query-synthetic.py

.venv/bin/python \
  scripts/plan-separation-other-refinement-query-reference.py

It downloads, loads and executes nothing. The approved private checkpoint evidence established SHA-256 657295888781e62ef50593002720d2edb3858b9e5bbfabf0c54f715a0da4b9e2 and a network-denied static structure report. The source audit also found the required 341,546,630-byte OpenMIC PaSST checkpoint and rejected the upstream automatic download and unrestricted loaders. Its separately approved evidence-only download established SHA-256 dc229428753176e8be0373d25887116fc15b490af86f671cecf9ed76a0f287da without loading it. A separately approved evidence-only dependency step then resolved 28 exact CPython 3.12/macOS-arm64 wheels (99,354,620 bytes), recorded every SHA-256 and licence metadata under network-denied non-importing inspection, and committed the 2,800-byte lock with SHA-256 28249f5d6ab80d4b72a5f256ac435f3a2d150b1baa30d751754af44049c33b92. A later separately approved gate installed those exact 28 packages from the local cache into a fresh CPython 3.12.10/macOS-arm64 environment using --no-index --require-hashes. Eight relevant package modules imported under network denial with zero checkpoint opens, torch.load calls, network attempts or audio opens. A further explicitly approved gate then constructed the real 64-band Banquet adapter and both download-disabled PaSST variants and loaded the two exact local checkpoints with torch.load(weights_only=True, map_location="cpu"). All 1,069 Banquet and 159 OpenMIC PaSST keys, shapes and dtypes matched before strict loading, with no missing or unexpected keys, network attempt or audio open. The later approved single synthetic inference attempt then passed every objective gate under network denial: 2.85 seconds, 2,052,014,080-byte peak RSS, finite stereo float32 output at 44.1 kHz and 7.450580596923828e-09 maximum in-memory reconstruction error. It used only generated tensors and opened no audio. The second command prints the immutable forward contract bound to nine exact files from the pinned upstream revision and its setup-C configuration. The third prints the pure result validator contract: an objective failure must be retained and grants neither automatic retry nor activation. The fourth command is the immutable plan that bound the now-consumed one-attempt authority. The planning commands still construct no model, run no inference and read no audio. The retained report SHA-256 is bd5fa57716267488cfd9a0d1d69bc1627da6244d283fdaec5a5592234d51cec8. The very quiet synthetic target peak (4.118360084248707e-05) is explicitly not musical evidence. A later, separately approved rights-bound canary used the fifth command's frozen song-disjoint guitar, keyboard and synth query bank against three owner-authorised mixtures. Exactly nine CPU attempts completed under network denial in 64.01 seconds at 3,259,236,352-byte peak RSS. All 36 private PCM24 artifacts matched their hashes, every target plus residual reconstructed its reference with zero-LSB error, and the worker recorded zero network, forbidden-audio or unapproved-checkpoint attempts. The report SHA-256 is fd15e3ba9524a49ebd182f86fa5c50ea0f5b02e95cc776ee5943a09147206ea8. That is objective execution evidence, not musical validation. The completed bound review rated eight of nine targets not useful and one quiet keyboard target partly useful; several guitar/synth targets were heard as blank or had query hints unrelated to the mixture. The challenger is therefore technically valid but musically unsuccessful. It remains unregistered, unselected and unavailable to MIDI; the approved nine-attempt authority is consumed and poor feedback cannot trigger a query hunt. See Guitar and keyboard/synth query-challenger plan. The incremental code-structure work is tracked in Separation maintainability plan. Probability, delivery time and the alternative synth/guitar portfolio are tracked separately in the Fine-stem separation delivery forecast.

The later bounded challenger prioritised synth first, then guitar and wind; acoustic piano was only an optional control and no longer stood in for modern keys. Its no-effects plan audited the public MVSep Mega 53 Stems release and an exact Apple-native BS-RoFormer source revision. The contract allowed only a native synth candidate plus a transparent residual, separated audible target presence from model usefulness, and permitted one configuration with no post-feedback tuning loop. The approved evidence-only gate locally verified the 1,368,919,887-byte checkpoint and 4,184-byte configuration against their exact published SHA-256 values, then inspected the checkpoint under network denial without loading it. A later evidence-only gate resolved and statically inspected 29 exact CPython 3.12/macOS 14+ arm64 wheels (127,527,173 bytes) without installing or importing them. The committed lock SHA-256 is 284d198c43e9074a4d645f005d937dd4e93b99e22aa21d942caaa1822b13d10b. An explicit follow-up installed that exact closure into a fresh isolated CPython 3.12.10 runtime and imported thirteen direct modules under OS network denial. There were zero connect or DNS calls, checkpoint/audio opens or torch.load calls; requests made one contained ::1 capability-bind probe. The exact 144,791-byte source archive was then materialised and all 64 files were verified against the sealed source inventory. The separately approved construction gate loaded the checkpoint exactly once with weights_only=True, matched 13,571 converted MLX parameter keys, shapes and dtypes, and retained the 53-stem model with zero network, audio or forward calls. That check also exposed and transparently adapted an upstream mismatch: the checkpoint requires transformer expansion 4, mask-head expansion 2 and float16 parameters, while the published adapter conflates the two expansion settings. The verified source was not mutated. Checkpoint use remains provisional local noncommercial evaluation and the profile is not public; one generated-tensor objective forward has now passed its separately approved gate. The frozen no-effects contract resolves the published 882,000-sample/441,000-step mismatch to an 881,664-sample chunk and 440,832-sample step, keeping both on the 512-sample STFT clock without padding, cropping or source mutation. It authorizes nothing by itself and permitted no automatic retry. The one call completed in 18.19 seconds at 15.42 GB peak MLX allocation with exact finite 53-role output and zero network/audio attempts. The later four-song Mega-53 synth canary completed in 37.44 seconds at 15.42 GB peak MLX allocation; the four-song BS-RoFormer-SW guitar canary completed in 52.79 seconds at 8.51 GB. Both ran under network denial, produced finite PCM24 target/residual artifacts and reconstructed within zero LSB. The bound human reviews found synth partly useful in 4/4 confirmed-present cases and guitar useful or partly useful in 4/4, with no catastrophic defects, bleed, artefacts or timing problems. Both therefore passed the frozen 60% private-Studio gate. The later fixed integration completed over the same eight reviewed windows: three sequential model loads made 16 bounded inference attempts and published 64 finite PCM24 artifacts with zero-LSB reconstruction and no network access. In the combined six-role review, synth was useful in 2/4 confirmed-present cases and partly_useful in 2/4; guitar was useful in 4/4. All eight outputs had no catastrophic defect. The exact no-effects outcome is private_six_role_integration_qualified. This is positive private Studio evidence, not a public six-stem product claim: three synth cases retained some content outside the synth estimate, the first downstream-MIDI review is methodology-limited, and model terms/resources still constrain release tier.

A no-effects integration plan bound the exact reports and reviews, reused the eight persisted primary estimates, and fixed a grouped-other-constrained three-way projection so synth, guitar and residual other could not be double-counted. The completed plan SHA-256 is 9507d1ef182a0060270033a770a823b758ba024e75cc42e768117e66893f1dec; the objective report SHA-256 is af0533233c4469c3914fbe2cf4eae1195de1ee545847c122a8a9f023f350d513. Record the completed review as a pure outcome without loading a model or reading audio:

python3 scripts/record-fine-stem-six-role-integration-outcome.py \
  SIX-ROLE-INTEGRATION-ROOT \
  --out FRESH/fine-stem-six-role-integration-outcome-v1

The completed review document SHA-256 is 407e4bf0ab686ceee2bcaa77473eca0a76b307b13b217e070cf5ae8a8cdb31ce; the pure outcome document SHA-256 is 85b63909743da20a0b68e9d2fc130d0120f99e88036653586f7507766cf5d6f9.

The downstream-MIDI no-effects plan is complete. It binds all eight exact confirmed-present synth/guitar artifacts and compares each target candidate with a sample-exact grouped-other control under identical BPM, key, tuning and transcription settings. It opens no audio, runs no model or transcriber, writes no MIDI and selects nothing. Reproduce that plan with:

.venv/bin/python scripts/plan-fine-stem-downstream-midi.py \
  SIX-ROLE-INTEGRATION-ROOT \
  FRESH/fine-stem-six-role-integration-outcome-v1/INTEGRATION-OUTCOME.json \
  --out FRESH/fine-stem-downstream-midi-plan-v1

The downstream-MIDI plan SHA-256 is 7afab38b0bd446e2de75b4c408b1e275e533298765f25d89280c055fbb63e1e4. Its separately approved private execution is complete. Under OS network denial, it verified all 24 input identities, wrote eight sample-exact grouped-other controls, made exactly 16 same-settings transcription attempts and published 16 private MIDI files plus 16 loudness-matched neutral previews. It made zero separator, checkpoint, network, selection or activation attempts. The report SHA-256 is 5f3ebf50c0097ca5a0169b63ed1eb4f2efc010d54b525321e5bfd3f621668b09. The eight-case A/B review is complete with review SHA-256 dc766790f97341521363f1705f90ab3dfa1456b1925b0e97e5e13d35e94c2103. Guitar candidate MIDI was preferred to its grouped-other control in 3/4 cases; synth was equivalent once and cannot_tell in 3/4. Recognisable-note and timing ratings were synth useful once and partly_useful three times, and guitar partly_useful three times and not_useful once. The completed review omitted the source reference, so these are methodology-limited findings rather than promotion evidence. The repaired local review now shows the exact source window beside A/B without changing the saved answers or requiring another checkbox. Poor results cannot disable the qualified six-role evidence.

The completed review is now reduced into a pure, hashed outcome. It records guitar as directional private evidence, records no isolated-stem synth advantage, and grants no model, audio, MIDI, source-selection or activation authority:

.venv/bin/python scripts/record-fine-stem-downstream-midi-outcome.py \
  PRIVATE-MIDI-CANARY-ROOT \
  --out FRESH/fine-stem-downstream-midi-outcome-v1

The outcome SHA-256 is d69863fb8ea59087ad6cfcd5669fee88db6afef6a7bd20809d42d88415c29a0c. A historical no-audio request then bound the exact four source-present synth cases and froze a three-arm attribution test: current synth estimate, provider synth/keyboard estimate and grouped-other control, all through the same transcriber with the source visible during review. It neither opens provider audio nor authorises the later 12-attempt run:

.venv/bin/python scripts/plan-fine-stem-synth-bottleneck.py \
  PRIVATE-MIDI-CANARY-ROOT \
  SIX-ROLE-INTEGRATION-ROOT \
  FRESH/fine-stem-downstream-midi-outcome-v1/MIDI-OUTCOME.json \
  --out FRESH/fine-stem-synth-bottleneck-request-v1

The request SHA-256 is d03f5ff549c5d778a6c07451c9f953be3fe29bc107e743ce5ae691e342e4419a. The missing Uni Ava pack was found locally and the deterministic first-pack policy is now frozen across all four songs. A separate qualification step bound 36 exact private input identities, wrote four source/provider PCM24 pairs and verified all four provider pack sums against the existing source windows. The sample correlations were 0.913–0.949, envelope correlations 0.946–0.980 and every best envelope lag was zero. The qualification report SHA-256 is 64b564798a6f338c66039e40981a5d562160c83451f2bc6d2a5411c56eba0eea. It loaded no model, ran no transcription and does not treat the provider Synth label as truth.

.venv/bin/python scripts/qualify-fine-stem-synth-provider-estimates.py \
  FRESH/fine-stem-synth-bottleneck-request-v1 \
  SIX-ROLE-INTEGRATION-ROOT \
  PRIVATE-PROVIDER-INPUTS.json \
  --out FRESH/fine-stem-synth-provider-qualification-v1

.venv/bin/python scripts/serve-fine-stem-synth-provider-review.py \
  FRESH/fine-stem-synth-provider-qualification-v1

The source-visible presence review recorded playback automatically, autosaved on the local server and had no listened checkbox. It completed all four cases before the separately hash-bound 12-attempt MIDI plan ran; poor musical feedback still cannot disable core four.

That source-visible three-arm MIDI comparison is now complete. Grouped other was preferred in all four synth cases, so Sunofriend retains grouped other as the MIDI control without discarding the separately qualified synth audio. The reviewed six-role excerpts can now be copied into an owner-only Studio package:

.venv/bin/python scripts/package-fine-stem-private-studio.py \
  SIX-ROLE-INTEGRATION-ROOT \
  FRESH/fine-stem-six-role-integration-outcome-v1/INTEGRATION-OUTCOME.json \
  PRIVATE-MIDI-CANARY-ROOT \
  FRESH/fine-stem-downstream-midi-outcome-v1/MIDI-OUTCOME.json \
  FRESH/fine-stem-synth-provider-midi-outcome-v1/PROVIDER-SYNTH-MIDI-OUTCOME.json \
  --out FRESH/fine-stem-private-studio-package-v1

The exact local build copied and re-hashed 72 PCM24 files across eight reviewed 15-second cases and wrote 16 explicit Studio catalogs. Each case has one six-role audio catalog and a separate grouped-other MIDI-control catalog; loading both as one source set would double-count synth and guitar. The package contains no MIDI, makes no source choice, loads no model and activates nothing. Its manifest SHA-256 is 4b5bce6daa00607a93257d33db3e1055ab651287458c21ccd637b304438053bd. This is a private evaluation slice, not a full-song or public six-stem claim.

The exact three-song full-song plan at SHA-256 869ac229d5c95c9c3d5eb2c9eb38da368056f6fe3c644de9830cc593313efb7d was approved and consumed once. It bound I am a Alien mashup for confirmed synth and guitar, Be Alone for synth and In the way for guitar. The guarded run ended in retained objective failure with no automatic retry. Its replacement failure retained complete SCNet, Mega-53 and guitar arrays but no guitar result receipt, guard counters or peak-memory measurement; both failed packages remain immutable.

A separately exact-hash-approved recovery reverified the retained PCM24, NPY and JSON identities under network denial and wrote the already-fixed 24 private PCM24 review artifacts. It loaded no checkpoint, constructed or loaded no model, ran no inference or canonicalisation and started no model worker. Its status is private_review_package_recovered_model_free_resource_gate_incomplete. This provides private full-song listening material, not full objective or resource qualification: the guitar and aggregate resource gates are incomplete and supported-ceiling compliance is unknown. The reproducible urllib3 IPv6 loopback socket.bind import path is consistent with the failed effects guard, but the missing historical worker receipt means it is not proven as that failure's cause. The future worker import order is repaired without weakening bind denial.

The bound review is complete with status human_listening_complete_no_selection. Playback was recorded for all 24 full-song items and all four confirmed-present windows. All three cases were overall useful and reported no catastrophic defect. Vocals, drums, bass and residual other were useful in 3/3 cases; synth and guitar were each useful in 2/2 confirmed-present cases. No scored role reported bleed, artefacts or timing/join problems. Synth and guitar each reported some missing content in 2/2 cases; core roles reported none. There were no cannot_tell or not_tested ratings. This is positive private musical evidence, not objective/resource qualification or a public six-role product.

The immutable private outcome is now recorded with status private_full_song_six_role_listening_evidence_recorded_resource_gate_incomplete and canonical document SHA-256 fa5d1d24627dce4cb1e27175055f1e3d5a3a70683b98e2376d92ee125bc2163c. It preserves the positive listening aggregate and every incomplete objective/resource/product boundary above.

The original planning command and its approval are historical, consumed and non-reusable. Do not recreate or rerun that plan or solicit its old approval. Any future model run needs a new bounded plan and explicit authority. Validate the exact recovered report without loading a model:

.venv/bin/python scripts/recover-fine-stem-full-song-six-role.py \
  --plan PRIVATE-PLAN-ROOT/FULL-SONG-SIX-ROLE-PLAN.json \
  --validate-report \
  PRIVATE-RECOVERY-ROOT/TECHNICAL/FULL-SONG-SIX-ROLE-RECOVERY-REPORT.json

The recovered package grants no public activation, source choice, MIDI, hosting, redistribution or upload. Poor listening feedback remains recordable without disabling core four. See Private full-song six-role canary plan.

Inspect the already-completed recovered full-song review locally with:

.venv/bin/python scripts/serve-fine-stem-full-song-six-role-review.py \
  PRIVATE-RECOVERY-ROOT \
  --plan PRIVATE-PLAN-ROOT/FULL-SONG-SIX-ROLE-PLAN.json

See Synth-first fine-stem challenger.

For the separate grouped-other refinement route, bind an exported listening JSON to its exact result without selecting a source or starting MIDI:

.venv/bin/sunofriend-separate review-other \
  "/absolute/path/to/refinement-result" \
  "/absolute/path/to/sunofriend-other-refinement-listening.json" \
  --out "/absolute/path/to/fresh-private-feedback.json"

For development across several songs, the fixed private corpus runner freezes five authorised songs, separate guitar and piano-proxy windows, one model configuration and exactly ten reviews. It places the Sunofriend estimates beside local Moises/Suno comparison excerpts without treating provider stems as ground truth or using feedback to tune between cases:

.venv/bin/python scripts/create-other-refinement-corpus-review.py --plan

The prepared package is served only on localhost so every audio control works reliably. Its ten-review bundle can be recorded as hash-bound evidence without selecting a source, promoting a model or starting MIDI. See Refining grouped other in Studio.

Not sure what a stem is, why a drums stem contains several drums, or where to get authorised examples? Read Stems: what they are and where to get them.

For your own music, valid starting points include:

• separate tracks exported from a DAW;
• stems from a song generator or project you are allowed to process;
• separated tracks exported from Moises or another stem separator; or
• the built-in Sunofriend demo.

Plans and export features change, so ask the agent to check the provider's current official help before you subscribe. Sunofriend does not download songs or grant permission to process music. Its default alpha is a broad two-stem estimate, and even the core-four preview is not lost multitrack recovery or a replacement for every specialist separation service.

Already have separate, synchronized parts as MP3, FLAC, M4A, AIFF, Ogg or WAV? The local source-import-folder command can check and prepare 2–64 files as one canonical WAV stem project before Simple or Studio runs. It does not separate a finished mix, repair alignment or prove the musical downbeat. One broad drums part can now produce review-required mixed-kit MIDI; it does not create separate kick, snare or cymbal audio files. The narrower source-import command still prepares one standalone asset. Both import routes now retain automatic, unreviewed key/BPM evidence. Inspect one file without importing it with sunofriend musical-metadata SOURCE; when preparing separated stems, pass the original finished mix with source-import-folder --metadata-source ORIGINAL_MIX. Explicit and filename metadata remain authoritative, and tuning observations are never promoted automatically. See source preparation.

What you receive

The automatic result contains:

AUTOMATIC-SONG/
├── START-HERE.txt
├── MIDI/individual parts and combined-gm-interpretation.mid
├── SOUNDS/named starter-sound MIDI, short previews and GarageBand guide
├── AUDIO/balanced-midi-song-interpretation.wav
├── TECHNICAL/automatic-musical-metadata.json and mix evidence
├── garageband-mix-recipe.md
└── sunofriend-automatic-midi-and-wav.zip

Listen to the WAV first. Then follow SOUNDS/INSTRUMENTS-START-HERE.md and import either the combined MIDI or the separate sound-aware MIDI files into GarageBand. Each one requests the same named General MIDI starter sound heard in the interpretation. The exact automatic-primary MIDI stays unchanged under MIDI/, and every starter sound remains an editable, unreviewed choice.

The WAV contains rendered MIDI, not the original stems. It is a creative interpretation of notes, rhythm and structure, not an exact reconstruction or a human-approved release master.

Simple or Studio?

Both modes use the same conversion engine. Studio preserves Sunofriend's main strength: different methods can be best for different instruments or phrases.

Hear an example

Listen to “Out of Place” on SoundCloud. It is shared as an example of Sunofriend interpolation: a new MIDI-derived musical interpretation, not an exact copy of source audio.

For a documented four-tool workflow, hear “The Aisle at Lidl” and explore its compact worked MIDI pack.

More detail

Privacy, rights and limits

• Current processing, review notes and feedback stay on the local Mac.
• Use only music you own or are authorised to process.
• GarageBand or another instrument supplies the final sound; MIDI does not contain a stem's exact texture, voice or effects.
• Mixed or noisy stems and melody extraction remain difficult. Automatic results are useful starting points, not ground truth.
• The current release is an alpha local macOS application. A hosted service for people without suitable hardware is a future direction.

Contribute

Reports from beginners, other Macs, other DAWs, separation results, operating systems and AI tools are especially useful. Use the beginner first-song report, the DAW / AI compatibility report, or read CONTRIBUTING.md.

Contact and support

For a private or general enquiry, email [email protected]. Please do not email stems, vocals, unreleased music, MIDI, project files or private review notes.

Use the public issue forms for reproducible first-song and compatibility reports. For a security concern, follow SECURITY.md and do not open a public issue. The full routes and privacy notice are at sunofriend.com/contact.

Sunofriend is licensed under Apache-2.0.