Skip to content

v0.23.1 — new ears for the lyrics, and Space works again

Choose a tag to compare

@github-actions github-actions released this 20 Sep 21:09
· 191 commits to main since this release
3c71669

The lyrics engine has been replaced end to end: it hears sung words far more accurately and invents nothing into silence. A machine with the old engine is offered the new one once, at launch — nothing downloads until you say so. Separated songs stream instead of loading whole, cutting open time and memory sharply, on the computer and the phone. And on the computer, Space plays and pauses again after a button was clicked, instead of pressing that button again.

Движок распознавания слов заменён полностью: он гораздо точнее слышит спетые слова и ничего не выдумывает в тишине. Компьютер со старым движком получит одно предложение перейти на новый при запуске — ничего не скачивается без согласия. Разделённые песни передаются потоково, а не загружаются целиком — это заметно ускоряет открытие и экономит память на компьютере и на телефоне. А на компьютере пробел снова играет и ставит на паузу после нажатия кнопки, а не нажимает её ещё раз.

Mac and Windows:

  • The lyrics engine is new, and it only listens now — no more online
    subtitle lookalikes.
    Whisper is gone; every transcription and every
    Check & align now runs on Qwen3-ASR, with a purpose-built aligner timing
    the words it heard. Across the songs it was measured on: it hears more of
    what is actually sung (word error 0.167 against 0.278 for songs with
    nothing to find online), it invents nothing — 55 phrases whisper wrote
    into pauses and instrumental breaks are gone, all of them — and it lands
    each word closer to where it is sung (median timing error 0.08s against
    0.19s). On a machine whose graphics card could accelerate neither
    engine it was measured faster than whisper too, not just more accurate.

  • The switch is offered once, and the old model cleans up after
    itself.
    The recogniser and its aligner are one 3.5 GB download. If your
    copy of SingZ already has a whisper model on disk, the next launch opens
    the model manager once with the new one highlighted and says why —
    nothing downloads until you press Get, and if you close it, it does not
    ask again. Once the new engine is fully installed, the old whisper model
    is removed on its own.

  • Re-running Check & align on a song you already checked is much
    faster.
    What the engine heard in a song's vocals is now kept beside it,
    so asking again — after editing a line, say — goes straight to timing the
    words instead of listening from scratch. On the songs it was measured on,
    a repeat came back three to six times faster.

  • Space plays and pauses the song again, after clicking anything.
    Click Mute, a fader, or almost any other button, then hit Space: instead
    of playing or pausing, it pressed that button a second time — muting and
    unmuting, retapping the tempo. A recent change widened a rule meant for
    one dialog into every button on the song screen. Space is the
    song's again; a field you are typing into, a focused dialog and a
    fader's own arrow keys keep theirs.

  • Cancelling a lyrics job gives you your words back, and a pick from
    Change… can no longer land in the wrong song.
    Press Cancel on a
    transcription or a Check & align and the lyrics you had went blank until
    you reopened the song — cancelling was being read as a verdict on the
    words, when it only means "stop, and keep what I had." It keeps them now.
    Picking a different set of words from Change… had the same gap from the
    other side: if you switched songs before the pick answered, it could be
    drawn into whatever song was open by then instead of the one you asked
    it for. Both are fixed now: cancelling keeps what you had, and a pick
    lands only in the song you made it for, however long it takes to answer.

  • Opening a project from outside your library folder can no longer make
    SingZ sync itself in an endless loop.
    A project living outside the
    library — copied in from elsewhere, or opened straight from another
    drive — was correctly left out of every sync, and that correct skip was
    then read as new work, which queued the next sync, which skipped it
    again, forever: a real Drive listing every few seconds for as long as
    SingZ stayed open. It syncs everything it should and quietly leaves that
    one alone now.

Mac, Windows, iPhone and Android:

  • Separated songs open faster and hold far less in memory. Every split
    since 0.23.0 keeps its lead and backing vocal lanes uncompressed, and
    that format was being decoded in full rather than streamed — on the
    phones, a five-minute separated song held over 800 MB in memory where an
    ordinary song holds under 20; on the computer, opening one took several
    times as long as an all-compressed song. Those lanes stream now, the same
    way the rest of the song already did.