8000 Releases · yeufonic/yeufonic · GitHub
[go: up one dir, main page]

Skip to content

Releases

v0.0.64

Choose a tag to compare

@dynamohum dynamohum released this 20 Nov 19:17

0.0.64: native Linux installer without Docker, Trash space for deleted takes and spaces, opt-in FIBO 2.0 across all installers, and fixes.

Windows: download Yeufonic-Setup-0.0.64.exe and run it over your existing install. Quit Yeufonic from its icon by the clock first (right-click, then Quit): closing the window leaves it running, and the installer waits for it to quit. Your library, settings, LoRAs, models and voices are kept.

Native Linux (no Docker): download yeufonic-0.0.64-linux-x86_64.run, mark it executable (chmod +x yeufonic-0.0.64-linux-x86_64.run), and run ./yeufonic-0.0.64-linux-x86_64.run. See linux/README.md for options and updates.

Docker: pull and rebuild (from any earlier version; a build with nothing changed takes moments):

git pull && docker compose up -d --build

If you run compose in the foreground, stop it with Ctrl+C and run docker compose up --build instead. Then reload the page.

Mac (Apple silicon): download Yeufonic-0.0.64.dmg, below, and drag it over the app in Applications. yeufonic.com/macos has the details.

What is new in 0.0.64

  • Native Linux installer without Docker (#77). Linux users with an NVIDIA GPU (x86_64, glibc 2.28+) can now download a single self-extracting installer (yeufonic-0.0.64-linux-x86_64.run) from the release page and run Yeufonic natively without Docker or sudo, with a choice of the ComfyUI or yue2.cpp engine, interactive setup and engine switching (yeufonic setup), detached background running (yeufonic start -d, stop, status, logs), and portable Linux binaries for FluidSynth and yue2.cpp. See linux/README.md.
  • Trash space for deleted takes and spaces (#78). Deleting a take (or deleting a space that holds takes) now moves its takes into a built-in Trash space at the bottom of the space list rather than deleting them immediately or dumping a deleted space's takes into Default. Takes in Trash stay out of playlists and search, show Play, Restore and Delete permanently, and can be restored individually or in bulk (using checkboxes or Select all). Deleting a take from Trash, pressing Empty Trash, or reclaiming Takes in Trash in Settings → Storage removes its audio, stems and artwork from disk permanently.
  • Opt-in FIBO 2.0 take-picture model across all installers. Bria's FIBO 2.0 with its turbo add-on (ComfyUI engine only, CC BY-NC 4.0 non-commercial use only) can now be chosen during setup instead of downloaded by hand: tick FIBO 2.0 (take pictures) in the Windows installer (setup.ps1 -Fibo), choose FIBO 2.0 in yeufonic setup on Linux (--picture-model fibo2), answer yes or run setup.sh --fibo on macOS (full-precision BF16), or run sh scripts/fetch-models.sh --fibo with Docker.
  • Hot keys button in the transport bar. Added a ? button beside M and V at the bottom left to open and close the Hot keys reference with a click as well as the ? key.
  • Storage lists the training checkpoints of every corpus's LoRA. It looked for them under the corpus's name, so a LoRA named otherwise (trained before the corpus was renamed, or under its trigger word) had its checkpoints missed, both in Storage and in the corpus's own checkpoint list. They are now found under the name the LoRA has.
  • Storage lists checkpoints whose corpus is gone, and greys out "Tick the ones that cost nothing" when none are free. A LoRA's training checkpoints stayed in the LoRA folder after its corpus was deleted, and nothing offered to remove them. Storage now lists them by LoRA (keeping the finished LoRA), and disables the free-items button when nothing listed is cost-free to remove.
  • The Windows app no longer fails to start now and then with "The app did not start within 3 minutes" (#76). When a connection to the app was dropped before it was accepted (the launcher's own check, while the app was still starting), Python's Windows event loop logged "Accept failed on a socket" and stopped listening, so the app ran on with nothing able to reach it. A dropped connection is now logged as a warning and the app keeps listening. The launcher's checks also no longer go through a system proxy.
  • Changing Harmony or Plan variety on a planned instrumental writes a new plan. Both steer only the planner, so with an instrumental open, changing them and pressing Render this score played the same plan again and gave a take that sounded almost the same. The main button now becomes Write a new plan when either differs from what the take's plan was written with: a new plan from the same structure, in a new take, the open one left as it was.
Earlier updates (0.0.63, 0.0.62)

0.0.63 (2026-11-18)

  • Write a new structure for an instrumental that already has a plan
  • A new instrumental starts with no recording, and choosing Instrumental with a song open makes an instrumental of it

0.0.62 (2026-11-17)

  • A long score you write yourself is no longer cut off at six minutes
  • An instrumental keeps to the sections you give it, and a new one starts from the usual sections
  • The editor's columns can be resized, and a corpus song's style suggestion goes straight into a new song or your saved styles

The full history is in the changelog.

Windows installer SHA-256: 0322b28b79969e661df8cf747aada53cab8c28faf50009081292788c2a056dc4. The installer is unsigned, so Windows may show a SmartScreen warning.

Native Linux installer SHA-256: b2b407f5ce249573f4ef38c10dcc1859624246f1175080699ec85d904128de78.

Mac disk image SHA-256: 15c86a0dbe91bae76c585d848c3b608518ba9e8b8f3a8f4a74e78f8aed6a94b5. It is unsigned too, so macOS asks before opening it; gh attestation verify Yeufonic-0.0.64.dmg --repo yeufonic/yeufonic checks it is the one GitHub's Mac runners built.

yue2.cpp for Linux (8af01740686824ffd4a1839a457954f42e2a8ce7 with Yeufonic's patches)

Choose a tag to compare

yue2.cpp's server (https://github.com/ServeurpersoCom/yue2.cpp) at 8af01740686824ffd4a1839a457954f42e2a8ce7 with the patches in engine-cpp/patches, and stable-diffusion.cpp's server (https://github.com/leejet/stable-diffusion.cpp) at 0a067120b388bfefb9a87b96960cb548b97e9157, built for Linux x86_64 (glibc >= 2.28) with CUDA 12.8 runtime libraries bundled by linux/build-cpp.sh and .github/workflows/engine-cpp-linux.yml. Downloaded by the Linux installer when yue2.cpp is chosen; not a Yeufonic release. SHA-256 of the archive: 9d8759c31674abf653676162fe98998b27d5af884b5b17de69ff22ad1daa5772 (856514666 bytes)

FluidSynth 2.6.2 for Linux (x86_64 static)

Choose a tag to compare

FluidSynth 2.6.2 (https://github.com/FluidSynth/fluidsynth, LGPL-2.1) with libsndfile built in, statically linked for Linux x86_64 (offline WAV rendering only) by linux/build-fluidsynth.sh and .github/workflows/fluidsynth-linux.yml. Downloaded by the Linux installer so imported MIDI files can be heard; not a Yeufonic release. SHA-256 of the archive: e96a2954961671f0d51e174d8019d71a2148928b10616ea46cd6308314930552 (1048817 bytes)

Yeufonic 0.0.63

Choose a tag to compare

@dynamohum dynamohum released this 18 Nov 10:32

0.0.63: write a new structure for an instrumental that already has a plan, and two instrumental fixes.

Windows: download Yeufonic-Setup-0.0.63.exe and run it over your existing install. Quit Yeufonic from its icon by the clock first (right-click, then Quit): closing the window leaves it running, and the installer waits for it to quit. Your library, settings, LoRAs, models and voices are kept.

Docker: pull and rebuild (from any earlier version; a build with nothing changed takes moments):

git pull && docker compose up -d --build

If you run compose in the foreground, stop it with Ctrl+C and run docker compose up --build instead. Then reload the page.

Mac (Apple silicon): download Yeufonic-0.0.63.dmg, below, and drag it over the app in Applications. yeufonic.com/macos has the details.

What is new in 0.0.63

  • Write a new structure for an instrumental that already has a plan. An instrumental open in the editor lists the sections of its score, which can be moved, copied or removed but not rewritten. Write a new structure under the list turns it into the editable structure a new instrumental has, starting from the take's own sections and lengths: add, remove, rename and reorder them, then Write score plan writes a new plan into a new take, leaving the take that was open as it was. Back to this take's sections returns to the list.
  • A new instrumental starts with no recording. After an instrumental from a recording had been open, a new instrumental kept that recording and showed its sections instead of the structure to fill in.
  • Instrumental chosen with a song open makes an instrumental of it. With a song or a cover open in the editor, choosing Instrumental showed its score's sections and offered Render this score, which rendered the song again. The button is now Make an instrumental: a new take that plays the song's score with no vocal, the song left as it was, as the score panel's Make an instrumental does.
Earlier updates (0.0.62, 0.0.61)

0.0.62 (2026-11-17)

  • A long score you write yourself is no longer cut off at six minutes
  • An instrumental keeps to the sections you give it, and a new one starts from the usual sections
  • The editor's columns can be resized, and a corpus song's style suggestion goes straight into a new song or your saved styles

0.0.61 (2026-11-17)

  • The Mac version catches up: LoRA training, recorded voices, the yue2.cpp engine, style LoRAs and MIDI playback on a Mac; no longer experimental

The full history is in the changelog.

Installer SHA-256: b7f4ae920bf3dd8dd1a4aa8ce74365fd0d1cb29367da52ab123e0615b73fbcec. The installer is unsigned, so Windows may show a SmartScreen warning.

Mac disk image SHA-256: 923d11033048910ffc2b8441a32ff431042d0a845ca592f4badd6bd2d00caa0d. It is unsigned too, so macOS asks before opening it; gh attestation verify Yeufonic-0.0.63.dmg --repo yeufonic/yeufonic checks it is the one GitHub's Mac runners built.

Yeufonic 0.0.62

Choose a tag to compare

@dynamohum dynamohum released this 17 Nov 23:57

0.0.62: long songs and instrumentals: a long score you write yourself is no longer cut off at six minutes, an instrumental keeps to the sections you give it, the editor's columns can be resized, and a corpus song's style suggestion goes straight into a new song or your saved styles.

Windows: download Yeufonic-Setup-0.0.62.exe and run it over your existing install. Quit Yeufonic from its icon by the clock first (right-click, then Quit): closing the window leaves it running, and the installer waits for it to quit. Your library, settings, LoRAs, models and voices are kept.

Docker: pull and rebuild (from any earlier version; a build with nothing changed takes moments):

git pull && docker compose up -d --build

If you run compose in the foreground, stop it with Ctrl+C and run docker compose up --build instead. Then reload the page.

Mac (Apple silicon): download Yeufonic-0.0.62.dmg, below, and drag it over the app in Applications. yeufonic.com/macos has the details.

What is new in 0.0.62

  • An instrumental keeps to the sections you give it. With Sections or Timed sections, the planner took the list as a suggestion and wrote its own shape, often a fraction of a long structure (6 sections of 23, three minutes of ten). It now writes the sections listed, in that order and no others, as a song does with Follow this structure exactly. The planner still decides each section's length, so timed sections set the order and the names more than the seconds, and a very long structure can still end early where the planner reaches the most it can write in one plan.
  • A new instrumental starts from the usual sections. It kept the last instrumental's list, so after a long one every new instrumental started with dozens of sections. It now starts from intro, verse, chorus, verse, chorus, bridge, chorus, outro, and keeps only the way the structure is given (Let YuE2 decide, Sections or Timed sections).
  • A score written by hand is no longer cut off at six minutes. The length cap rises to fit a song's plan, and a cover's to fit its recording's score, but a score typed, pasted or edited longer kept the cap the take started with, and the render faded out there (six minutes unless Settings say otherwise). It now rises to fit any score when the take is rendered, unless you typed the cap yourself.
  • An emptied Length cap box no longer means a cap of six minutes. Clearing the box was read as a cap of 360 seconds typed by hand, whatever Settings said, so a long plan was turned away against it. An empty box is now no cap of anyone's: the take gets the one in Settings, which rises to fit a longer plan, and the box shows it again.
  • Lyrics too long for the length cap say so. Lyrics of many sections, with Follow this structure exactly on, can ask for a plan of a quarter of an hour, which is turned away when it runs far past the cap; the message called the plan unreadable and suggested Harmony Familiar. It now says how long the plan came out and which cap it was over, and to use fewer sections or raise the cap. A plan that runs on past short lyrics keeps the old advice.
  • A corpus song's style suggestion can be used straight away. Under the suggestion in a song's review, Use in a new song opens + Song with that style in the Style box, and Save as a style keeps it in the style list under a name you give it. Before, it had to be copied out by hand.
  • The editor's columns can be widened and narrowed. Drag the bar between two columns to give one more room and its neighbour less, down to a width each keeps its controls readable at; double-click the bar for the usual widths. The widths are kept in the browser. A wider first column also widens its style list, so more of each style shows.
Earlier updates (0.0.61)

0.0.61 (2026-11-17)

  • The Mac version catches up: LoRA training, recorded voices, the yue2.cpp engine, style LoRAs and MIDI playback on a Mac; no longer experimental
  • yue2.cpp's full-quality model can be chosen when installing, and LoRAs work with it

The full history is in the changelog.

Installer SHA-256: 27cff42e44c4641e23402432c84075ca111837defe0a2ee5ed7b9331dcec27bc. The installer is unsigned, so Windows may show a SmartScreen warning.

Mac disk image SHA-256: 28bd215a01e47496ffb5f4c689d806b966a4a25bc05c1b435da332c7bb7d1776. It is unsigned too, so macOS asks before opening it; gh attestation verify Yeufonic-0.0.62.dmg --repo yeufonic/yeufonic checks it is the one GitHub's Mac runners built.

Yeufonic 0.0.61

Choose a tag to compare

@dynamohum dynamohum released this 17 Nov 13:02

0.0.61: the Mac version catches up. LoRA training, recorded voices, the yue2.cpp engine, style LoRAs and MIDI playback now work on a Mac, lyrics are heard several times faster, and pictures no longer hold up other jobs. It is no longer marked experimental, though it is still much slower than an NVIDIA card. Also: yue2.cpp's full-quality model can be chosen when installing, and LoRAs work with it.

Windows: download Yeufonic-Setup-0.0.61.exe and run it over your existing install. Quit Yeufonic from its icon by the clock first (right-click, then Quit): closing the window leaves it running, and the installer waits for it to quit. Your library, settings, LoRAs, models and voices are kept. With yue2.cpp, the components page now offers its full-quality model beside the Q8 one.

Docker: pull, then rebuild the engine and bring it up again (the engine changed: its LoRA trainer is a newer commit, and the yue2.cpp image is new):

git pull && docker compose up -d --build

If you run compose in the foreground, stop it with Ctrl+C and run docker compose up --build again. Then reload the page.

Mac (Apple silicon): download Yeufonic-0.0.61.dmg, below, and drag it over the app in Applications; opening it runs the setup again, which now also asks about style LoRAs and recorded voices. yeufonic.com/macos says how to install it, how to choose the yue2.cpp engine, and how fast it is next to an NVIDIA card.

What is new in 0.0.61

  • Mac: LoRA training works. Both halves of a style LoRA now train on a Mac's GPU, so the Train window is on there as everywhere else. ComfyUI loads the model straight onto a GPU that shares the computer's memory and leaves it there, and its weights stayed inference tensors that training cannot use; the trainer now copies them first, on a Mac only, so training on an NVIDIA card is as it was.
  • Mac: the yue2.cpp engine. setup.sh --engine yue2cpp sets the Mac up with yue2.cpp in place of ComfyUI, as the Windows installer can: yue2.cpp's server and stable-diffusion.cpp's, built by us for Macs on macOS 14 and later, with Ollama for Gemma beside them. Lyric drafts take seconds there instead of minutes, and pictures much less memory; songs render at about the same speed, and LoRAs are not trained.
  • Mac: recorded voices. The setup window asks once whether to set up recorded voices, as the Windows installer offers them (setup.sh --voices later), and the app then starts the voice service beside the engine. A voice trains and sings on the Mac's GPU: RVC, at the commit the other installs use, gets a small patch of ours that lets it, and the separator runs there too, with Metal's missing operations on the CPU.
  • Mac: hearing the words of a song is several times faster. Whisper now runs on a Mac's GPU (Apple silicon), with the same lines and timings as on the CPU, which it falls back to if the GPU fails.
  • Mac: the setup offers the style LoRAs. The setup window asks once whether to fetch the 39 style LoRAs by their authors that the Windows installer and Docker can fetch, with their descriptions, and remembers the answer (setup.sh --loras later). A LoRA file already in the folder is left as it is.
  • Mac: MIDI files can be heard without Homebrew. FluidSynth publishes no macOS build, so the setup now fetches one of ours, built unchanged from FluidSynth's source as a single program, where before a Mac needed Homebrew's.
  • Mac: take pictures no longer hold up the engine for minutes. With the ComfyUI engine on a Mac, Gemma runs on the CPU, so a picture's description took minutes and every job behind it, stems and lyric hearing among them, waited. There the default describer is now the take itself; Gemma still describes when it is chosen.
  • The full-quality model for yue2.cpp can be chosen when installing. With yue2.cpp, the installer's model size choice is now that engine's own: Low memory (Q8), ticked to begin with as before, or Full quality (BF16, 7.2 GB), which needs about twice the video memory. Each engine keeps its own choice, a size already downloaded is kept, and a file of your own named in settings.ini stays unless Full quality is chosen. On a Mac, setup.sh --cpp-model bf16.
  • yue2.cpp: LoRAs work with the full-quality (BF16) model. With the BF16 file chosen (YUE2CPP_MODEL), any LoRA that changes the sound, the instrumental and Realaudio ones among them, stopped the render with "llm2vae.bias is not staged by this model". yue2.cpp turns that file's small tensors into full precision as it loads them, and the LoRA merge then could not find them; a patch of ours lets it.
  • yue2.cpp: a picture no longer fails after one that stepped aside for a song. A take picture that gave way to song work could go on drawing on the picture server, which then refused the next picture's model with "model_busy", and that picture failed. Pictures beside the yue2.cpp engine are now drawn one at a time, and a picture waits for the server to finish one it was asked to cancel.
  • Windows: the taskbar shows Yeufonic's icon again. A test build of the installer gave its window the same taskbar identity as the real install's, and once it was removed Windows kept showing a generic icon for both. The window now has an identity of its own (and a test build another), so the button shows the Yeufonic logo.
Earlier updates (0.0.60)

0.0.60 (2026-11-15)

  • Take files that carry a take whole (#60), sampling values by number (#72), notes on takes (#71), and more

The full history is in the changelog.

Installer SHA-256: c2d4d296f03e6b93eecdb9e019b8c09084c003890acdca7c4f763f53303ecf68. The installer is unsigned, so Windows may show a SmartScreen warning.

Mac disk image SHA-256: 367d60d05871f9d1adb2be07717bc0c88865662c5ff862802a65af4dff4d5eee. It is unsigned too, so macOS asks before opening it; gh attestation verify Yeufonic-0.0.61.dmg --repo yeufonic/yeufonic checks it is the one GitHub's Mac runners built.

yue2.cpp for Windows (8af01740686824ffd4a1839a457954f42e2a8ce7 with Yeufonic's patches)

Choose a tag to compare

yue2.cpp's server (https://github.com/ServeurpersoCom/yue2.cpp) at 8af01740686824ffd4a1839a457954f42e2a8ce7, with the patches in engine-cpp/patches, built for Windows with CUDA 12.8 by .github/workflows/engine-cpp-windows.yml. Downloaded by the Windows installer when the yue2.cpp engine is chosen; not a Yeufonic release. SHA-256 of the zip: 945faa49f190dc6804912cd9e87ebf752158bd98bd6552148a20d1e62c2107ae

yue2.cpp for macOS (8af01740686824ffd4a1839a457954f42e2a8ce7 with Yeufonic's patches)

Choose a tag to compare

yue2.cpp's server (https://github.com/ServeurpersoCom/yue2.cpp) at 8af01740686824ffd4a1839a457954f42e2a8ce7 with the patches in engine-cpp/patches, and stable-diffusion.cpp's server (https://github.com/leejet/stable-diffusion.cpp), built for macOS 14 or later on Apple silicon with Metal by macos/build-cpp.sh and .github/workflows/engine-cpp-macos.yml. Downloaded by the Mac setup when the yue2.cpp engine is chosen; not a Yeufonic release. SHA-256 of the archive: 4478032b265982475897eb2023fa9553fc45ce63302ad582e0684377a2478333

FluidSynth 2.6.2 for macOS (Apple silicon)

Choose a tag to compare

FluidSynth 2.6.2 (https://github.com/FluidSynth/fluidsynth, LGPL-2.1), built unchanged for macOS on Apple silicon with libsndfile built in, by macos/build-fluidsynth.sh and .github/workflows/fluidsynth-macos.yml. Downloaded by the Mac setup so MIDI files can be heard; not a Yeufonic release. SHA-256 of the archive: bb862c35de5d24e6c073e13aa7e4db53c7f7d4f39159f0f56b6905eefbfef466

Yeufonic 0.0.60

Choose a tag to compare

@dynamohum dynamohum released this 15 Nov 14:19

0.0.60: take files that carry a take whole, its recording and its audio included, so a cover you share or move arrives exactly as it was; and sampling values by number, notes on takes, a time signature to choose, saves for a DAW, an A–B loop, gapless Compare and a new look.

Windows: download Yeufonic-Setup-0.0.60.exe and run it over your existing install. Quit Yeufonic from its icon by the clock first (right-click, then Quit): closing the window leaves it running, and the installer waits for it to quit. Your library, settings, LoRAs, models and voices are kept.

Docker: pull, then rebuild the engine and bring it up again (the engine changed: the time signature, and lyrics heard in the language sung):

git pull && docker compose up -d --build

If you run compose in the foreground, stop it with Ctrl+C and run docker compose up --build again. Then reload the page.

Mac (experimental, Apple silicon): download Yeufonic-0.0.60.dmg, below; yeufonic.com/macos says how to install it and what does not work yet. Recorded voices are not on the Mac yet.

What is new in 0.0.60

  • Take files that carry a take whole (#60). Save → Take has a with its recording tick for a take made from a recording: the recording as uploaded goes in the zip under recording/, beside the recipe and any audio format chosen. Import take (and dropping on the library) takes such a zip, or a take file chosen with its recording. A take arrives as the kind it was saved as, every setting kept, so an imported cover renders as the original did: its recording joins the library (or the same one already there is used) and the cover is linked to it, and with its audio it arrives done. A cover imported without its recording renders its own score, and is linked to it once that recording is imported or chosen for it in the editor. This replaces importing a cover as a song with its recording's settings left behind.
  • The editor's header button is Advanced+, so it is not taken for the Advanced section in the editor's column.
  • Saving a take shows its progress in the top bar's job pill, as a render does: "Saving “Night drive” · Making the MP3 · 42%", or the stage for a mastered file (mastering in the page, then making the file) and for a DAW zip. The file is made in the background and fetched once it is done, instead of a download that said nothing until it arrived. While the engine is busy, the save is named under its job.
  • Advanced uses every number as typed, unless the engine cannot take it. A value outside a setting's usual range turns its box amber and is still used, for experimenting. One the engine would refuse (past its own limits, or a fraction where it takes whole numbers) is set to the nearest value it takes once the box is left, and the box says so; the server refuses such a value from anywhere else. Before, the Sampling and Follow the style/CFG boxes pulled any value outside their usual range back without a word, and the other settings quietly sent their default.
  • Make a new picture with a new prompt. Beside Make a new picture in the editor's Picture menu, under the whole picture and on a card with no picture: it asks for the art style of that one picture, starting from the one in Settings, and uses it in that picture's place only. Settings stay as they are, and the log names the style used.
  • Sampling values by number (#72). Advanced has a folded Sampling section at its foot with the planner's and the music stage's temperature, top_p, top_k and repetition_penalty. A blank box follows Plan variety or the Interpretation and shows its value; a value typed replaces that one alone. The values are saved with the take and its recipe file, Sing again and Try more keep them, and the card says (custom) beside the Plan or Interpretation they change. Variations has a tick to keep the music stage's values on every variation; left off, each is heard as its interpretation alone. Values away from the presets can make a plan come out unreadable or a render fail (a high planner temperature with a livelier Plan variety, for one): Reset to defaults puts them back, and the log names the values a failed plan or render ran with.
  • Follow the style is labelled Follow the style/CFG, the name it goes by elsewhere.
  • A take made from another no longer copies its note (Sing again and Try more copied it).
  • The editor's foot offers Render this score for a song whose plan is ready and not yet rendered, such as an imported take, beside Write score plan; the score's own row loses its copy meanwhile.
  • Hovering a take's lit pen shows its note in a card like the picture's, and a click on it edits the note.
  • Notes on takes (#71). The pen among a card's small buttons, and in the editor's top row, opens a box for a note of your own on the take: what it is for, the part to keep, w 5657 hat to cut. The pen is lit while the take has a note, and its tip shows it. Search takes finds takes by their notes, and a take's recipe file carries its note; a recipe from before notes imports as it did.
  • A new title in the editor stays with the new take. The title box renames the take open in the editor as it is typed in; when the editor is then left by making a new take (Write score plan, Create cover), the take it was made from gets its own name back. Leaving any other way (Escape, Close) keeps the rename.
  • Search the recordings when choosing one for a cover: a box at the top of the list finds recordings, uploads and corpus songs by every word typed, and Enter picks the first. It replaces the corpus songs' own filter.
  • The take panel's blocks fold. Click Style, Sound, Words or Score to fold it to its heading; the panel keeps it folded in this browser.
  • The take panel's score line names the time signature, and any the plan changes to on the way (3/4 → 4/4).
  • Key lock names the note. It offered major and minor keys, but the planner cannot be asked for a mode: a style naming one, or a score made to open with a minor key line, still came back in the plan's own mode. So the lock moves the plan to the note chosen and keeps its mode, as it always did, and now says so; a key lock set before to a minor key keeps working.
  • Choose the time signature. Advanced → Time signature writes the plan in 4/4, 2/4, 3/4 or 6/8, and the render follows it; Auto leaves it to the planner, which mostly writes in four. For songs and instrumentals from a prompt, and through MCP (time). A style that names the metre is not enough on its own, so the plan is made to open with the time signature and kept from changing it (with the engine rebuilt). A plan that still leaves it, or whose bars are not that long, is written once more; a second miss is kept, and its score says the time signature its bars actually make.
  • Save a take for a DAW. A new choice in Save makes one zip: the take as a 24-bit WAV (mastered, when the rack is on), its newest stems, the score as MIDI and as ABC, a chord chart with the tempo, key, time and each section's chords, and the recipe.
  • Compare switches with no gap. Every take is loaded in full and plays at once from the same instant, and a switch only changes which one is heard, so it lands on the same moment of the song. Before, each take was its own player, paused and started on every switch, which left an audible seam.
  • Loop a part with A–B on the player: pressed at the start of a part, such as a chorus, and again at its end, it plays that part round and round, shaded on the waveform, until pressed once more. Another take starts with no part set.
  • A first visit starts on Song, which needs only words and a style, rather than Cover, which needs a recording first; an empty library points to + Song, + Cover and + Instrumental. A page that was left on another kind still opens on it.
  • Each letter key means one thing. V opens and closes the visualizer whether or not anything is playing, and Shift+V is Variations; L is the logs and Shift+L the visualizer's lyrics. They used to change with what was playing or showing.
  • A new look, the default on new installs: the Studio theme with the Studio Slate style. Settings → UI style is new: Studio Slate gives each of the editor's three columns a colour of its own, and shows Corpora and Voices as cards with a filter box, each corpus with a bar of how many of its songs are included and each voice marked with where it came from and whether it is trained. Classic keeps the earlier look, and either goes with any theme. Studio (warm and dark) is now the default theme. An updated install takes the Studio Slate style and keeps its theme (Dark, or whatever was chosen) until another is picked in Settings.
  • Hear it plays again after a stop. Stopped and started with no setting changed, it stayed silent, because it took the voice it had already sung as playing; it now plays it again at once.
  • A mastered save keeps 24 bits. The rack renders in floating point, but it handed the server a 16-bit file, so a mastered FLAC or WAV lost its quietest detail; it now hands over 24-bit, and a mastered WAV is saved at 24 bits.
  • Lyrics from a recording are heard in the language sung. Whisper was told every song was in English, so words in another language came out wrong or turned into English; it now finds the language itself, from more than the opening seconds. On Docker the engine needs a rebuild for this (docker compose up --build).
  • A saved FLAC or MP3 carries its tempo and key from the take's score (BPM and INITIALKEY, TBPM and TKEY), where DAWs and DJ programs read them.
  • The DPX meter's top LEDs light. Full scale read as +16, so +18 and +24 never lit; full scale is now +24.
  • The unused code for ba...
Read more
0