Skip to content

v0.7.3

Latest

Choose a tag to compare

@github-actions github-actions released this 16 Nov 19:35
· 129 commits to main since this release

Alpha software

transit is alpha: it changes every week, and some things are rough or missing. A release can change or reset your settings, the plug-in caches and the session format, so back up your sessions before upgrading. Tell us what breaks on the Discord or in the testing thread.

transit v0.7.3

This is a patch release. transit is alpha-quality software and the session format may still change, so back up anything you care about.

v0.7.2 (v0.7.2 notes) fixed recording, saving and plug-in hosting. This one continues with fixes found after it.

macOS: signed and notarized. This is the first release whose Mac build is signed with a Developer ID and notarized by Apple. A tester confirmed on Apple silicon that the disk image opens with only the standard first-launch prompt and that third-party plug-ins load. If macOS still blocks the app, the fallback is xattr -dr com.apple.quarantine /Applications/transit.app or right-click the app and choose Open. Windows builds are still unsigned.

Before you upgrade

  • ⚠️ Older builds and sessions with a pre-count zone: a session recorded with count-in notes from bar 1 has a clip before bar 1. A transit build from before this release cannot reach that part: it plays from bar 1, and moving the clip there puts it on bar 1 and shifts its notes later. Builds since v0.3.0 warn when they open such a session (it lists the pre_count need); keep it to this release or later.
  • ⚠️ Older builds and sessions with unplaced takes: a session saved by v0.7.3 that lists an unplaced take cannot be opened by v0.4.1 through v0.7.2, which report "not a transit session: invalid type: map, expected a string"; builds before v0.4.1 open it without the unplaced-take list. These entries arise when a take could not be placed: it was recorded wholly before bar 1, its track was deleted while recording, its playlist could not be copied, or the take was unreadable, damaged or too long to load. Keep such sessions to this release or later. Sessions with no unplaced takes still open in older builds.

Changed behavior

  • A stereo audio track whose input is a mono input records and hears that one channel on both sides; it used to take that channel and the next one. This includes a track made stereo from a mono track on a named mono path. (Details under Fixes.)
  • MIDI file export leaves muted clips out. (Details under Fixes.)
  • A MIDI take with notes played during the count-in now covers the whole count-in, so it overlaps an earlier clip on its track more often and merges into it or starts a take stack. Event > Time Operations > Move Song Start is gone. (Details under Highlights.)
  • When the audio device runs at a rate other than the session's, the output and the live input are now converted by a band-limited converter. It keeps more of the treble and folds back far less than the old linear conversion, within the limits listed under Fixes. Conversion also adds delay to live monitoring. In the tests through the fake device, the hardware measurement's extra reading was under a millisecond, where v0.7.2 added about one device buffer (see the fix for #724); that reading does not establish the exact monitoring delay. Measure hardware inserts and External Instruments again at the rate you use. A device at the session's rate is unchanged at speed 1. (Details under Fixes.)
  • Audio clips with a stretch other than 1 (Conform to Tempo, Elastic Properties on a track without an Elastic algorithm) now play and render through a band-limited reader, so playback, bounces, stems, freezes, Consolidate and clip renders of such clips differ from v0.7.2. The old read interpolated linearly between two samples: it dulled the treble and let sped-up clips fold high content back. Renders of existing sessions with stretched clips change once, to the correct sound. Neither command has a menu entry in this release (the per-clip stretch mode in #336 replaces them), so this affects sessions that already hold stretched clips. (Details under Fixes.)

Highlights

  • Count-in pickups are kept from the start of the session. Notes you play during the record count-in are kept where you played them, also when recording from bar 1 at the very start of a session. The take's clip then covers the whole count-in, so it starts before bar 1, in a pre-count zone whose bars count down (0, -1, ...); Return to Zero and Play from the start play the pickups, and a bounce or MIDI file export starts there. One rule everywhere: later in the session the clip also covers the whole count-in, where it used to start at the first pickup. Nothing is moved onto the beat. Controller moves during the count-in are kept too (hidden before the clip's start when no note was played), recording works from a cursor, a pre-roll or a cycle range in the zone, and a note played before the count-in started is said on the status line. Audio clips still start at bar 1: an audio take that starts in the zone keeps that audio in its file, before the clip's start, and the status line says so. The status line's "count-in notes before the session start were left out" is gone with the limit it reported (#441, #440).
  • Move Song Start is retired. Event > Time Operations > Move Song Start and the event.move_song_start command are gone: it moved only tick-based material, parting MIDI from audio, and the pre-count zone replaces its use for room before bar 1. The Time Operations window's reply no longer carries song_start (#441).

Fixes

  • Rate conversion between the session and the audio device is band-limited. When the device runs at another rate than the session (it does not offer the session's rate, or another rate was saved for it in Audio Devices), the output and the live input (what you monitor, and an External Instrument's return, which a real-time render prints) went through linear interpolation, which dulled the treble (44.1 kHz on a 48 kHz device: -1.5 dB at 10 kHz, -6.3 dB at 20 kHz) and folded content above the device's limit back (96 kHz on a 48 kHz device at full level). Both now use the windowed-sinc reader built for the audio clip editor. The tests hold it to an error below -73 dB of the signal up to 0.42 of the lower rate (18.5 kHz when either rate is 44.1 kHz, 20.2 kHz at 48 kHz), and content that would land above 0.58 of the device's rate at least 73.5 dB down (78 dB when converting by a ratio of 1.5 or more). They also report, as observations rather than bounds: at every common pair of 44.1, 48, 88.2 and 96 kHz, tones within 0.0003 dB with everything else at least 89 dB down, and a 30 kHz tone converted down folding back at least 82.9 dB down. Above 0.42 of the rate is a transition band (observed: 20 kHz from 44.1 kHz on a 48 kHz device reads -0.38 dB). The converter looks ahead about 16 frames each way. Playback, the timing of MIDI sent out to other gear and where a take lands do not move. Conversion also adds delay to live monitoring. In the tests through the fake device, the hardware measurement's extra reading was under a millisecond; that reading does not establish the exact monitoring delay. The figures per rate pair and their limits are in the device-rate-converter-quality known issue. Notes played into a software instrument with the live input's look-ahead land on time in the tested cases (a 48 kHz session on 44.1 and 96 kHz devices). Measure hardware inserts and External Instruments again at the rate you use. A record count-in at such a rate keeps its full length, the take's start latch follows the first beat to the nearest output frame, what sounds across the start (a monitored input, a held note, a tail) stays continuous, and pickups land on the clicks they were played with (tested with whole bars on 44.1, 48 and 96 kHz devices). At equal rates, speed 1 and a whole-frame converter phase, finite samples pass through unchanged, as before (#601).
  • Live input no longer drifts later after a loop wrap or a count-in end, at any rate. In released builds, transit rendered a whole device buffer ahead whenever a block started in the middle of a callback. The live input (what you monitor, hardware insert and External Instrument returns, and what a real-time render prints from them) is read with each block, so after each loop wrap or count-in end inside a callback it moved later, with silence on the way, until it was about one buffer late. At a converted device rate this happened from the start, and a note played into a software instrument could land up to a buffer late. One example measured before the fix was 3.6 ms with a 128-frame buffer at 48 kHz on a 44.1 kHz device; no regression test pins that exact figure. v0.7.2 rendered the same way. Each callback now renders only what it plays. In the tests, the live input stays where it was through loop wraps at the session's rate, with no silence; at a converted rate it carries only the converter's delay described above, not a whole buffer; and buffers above 512 frames (600 and 1,024 tested) no longer start the live input with up to 512 frames of silence. Measure hardware inserts and External Instruments again at the rate you use (#724).
  • System Audio recordings keep their treble steady. The System Audio input (Windows) converts the captured output's rate to the input device's while it follows the drift between the two clocks, and it did that by linear interpolation at a position that never sits on a whole sample, so the treble swelled and dipped (at 48 kHz, up to 12 dB at 20 kHz and 5 dB at 15 kHz, about 2.4 times a second at a 50 ppm drift) and the result was recorded. It now uses the band-limited converter: a 15 kHz tone keeps its level within 0.001 dB in every 100 ms at drifts of 0, 50 and 1,000 ppm, a change of the drift trim never moves the read position (no click), and at the same rates with a drift trim of exactly zero every sample passes through untouched (on two real clocks the trim is rarely exactly zero, so the converter runs). The latency Audio Devices reports for System Audio is unchanged: the converter's look-ahead sits inside the ring's 30 ms margin (#603).
  • Stretched audio clips are read with a band-limited filter, in playback and in every render. A clip with a stretch other than 1 (Conform to Tempo, Elastic Properties on a track without an Elastic algorithm) used to be read by linear interpolation between two samples, in playback and in bounces, stems, freezes, Consolidate, Clip Effects Render and clip-gain renders, so the damage was written into files. Measured with -6 dBFS tones in a clip conformed from 133 to 130 bpm at 48 kHz, the largest error against the ideal tone is now -85.8 dB at 10 kHz (it was -13.7 dB) and -73.8 dB at 20 kHz (it was -2.6 dB); a clip sped up to twice its speed keeps what would play above half the sample rate 81.5 dB down, where the linear read let it through at nearly full level. A clip streamed from disk plays and renders what the same clip gives from memory, sample for sample, and an unstretched clip reads the file as before (a stretch that should be 1 but comes out a hair off, as after Conform to Tempo across a tempo change and back, is stored as exactly 1). Stated limits: content above 0.42 of the session rate (20.2 kHz at 48 kHz, 18.5 kHz at 44.1 kHz) is in the reader's transition band; a stretched clip costs processor time in proportion to how far it is sped up; a clip whose stretch is outside 1/32 to 32 (sped up or slowed down more than 32 times), which only a hand-edited session file can carry, is clamped to the nearest bound when the session opens or its tracks or clips are imported, with a message naming the track and clip (#604).
  • An audio take that ends before bar 1 now appears under Unplaced takes. Undo removes its listing and source together, and Redo restores both; the recorded file is always kept (#705).
  • Save As updates unplaced takes to name the saved audio, so their listings keep working after the autosave folder is removed.
  • A take whose track was deleted or whose playlist could not be copied is listed under Unplaced takes with its reason and file. Unreadable or damaged take files stay listed too. An empty take no longer adds an empty Record undo step.
  • After cycle recording before bar 1, the note gives the number of affected laps and their total audio kept before bar 1, including when only one lap is affected. A take kept entirely before bar 1 no longer says "Recorded nothing" (#706).
  • Bounce in Place and Commit no longer say the part before bar 1 stopped playing when the source track or that part's clips were already muted (#702).
  • Bounce in Place now says when Mute, Mute and deactivate, or Hide silences other clips before bar 1 on the source track, even when only clips after bar 1 were selected. Choose Leave to keep those clips sounding (#713).
  • Exports keep markers stored before the export start at the start and say how many were moved. This applies to project, MIDI file and consolidated exports; project exports could write negative marker times. A marker already on the start sample is not counted as moved, even when the bar line falls between samples (#701).
  • Linux, JACK system: the input device picked in Audio > Audio Devices is the one that opens. transit used to take the input from the output's device whatever was picked, and said the picked input was not available. If the output device has no inputs and none is picked, the first device that can record is used (#682).
  • A stereo audio track can take a mono input. On a device with one input channel its input menu used to list nothing; now a stereo track's input menu lists the stereo inputs, then the mono ones, and a mono input is recorded and heard on both sides. A stereo track with no input set reads "In 1" on such a device, which is what it records. Make Stereo on a mono track whose input has no stereo pair (the one input of a one-input device, or the last input of an odd count) keeps that input on both sides instead of refusing; a numbered input with a pair still becomes the pair ("In 3" becomes "In 3-4"), and a named input path is kept. Tracks wider than stereo and FX channels are unchanged. Changed behavior: a stereo track whose input is a mono input, including a track made stereo from a mono track on a named mono path, now records that one channel on both sides where it used to take that channel and the next (#681).
  • Audio Devices > I/O: a device whose saved I/O table had no inputs (saved while its input did not open) gets the default input paths ("In 1" and so on) when it opens with inputs. Paths you made stay, and a default you removed stays removed (#690).
  • Baking a looped MIDI clip no longer moves notes that sit before the loop start onto the clip's first tick: they stay where they are, as playback plays them (#441).
  • Dragging several clips toward the start now previews every clip stopping together, where the drop has landed since v0.7.1 (#441).
  • In a session with clips before bar 1, the consolidated and DAWproject exports start earlier and Bounce in Place prints from bar 1. Consolidated tracks' stems, MIDI files and tempo file all start at the bar line where the pre-count zone begins, so they line up with each other; bar 1 falls that many bars into each file, where they used to start at bar 1 and leave the zone out. Bounce in Place of clips that start before bar 1 prints from bar 1, with the part before it kept hidden in the printed file; with the source muted, deactivated or hidden (Mute is the default) that part is no longer heard, and the status line says so (choose Leave to keep it sounding); clips that end before bar 1 are refused. A DAWproject export starts at that same bar line: every time in the project moves later by those whole bars and the stems start there, so nothing sits before the project's start. The session length (the open log line and session.inspect's length_samples) counts the zone, and session.inspect reports start_samples and end_samples. Sessions without clips before bar 1 export and print exactly as before (#441).
  • MIDI file export writes each clip as it plays, and leaves muted clips out. Notes that start past a clip's end are left out, notes held over the end stop there, and a looped clip's repeats are written out. It used to write every note a clip held, past its end too, and a looped clip's content only once. A muted clip is no longer written, because a MIDI file cannot mark it muted and it would sound wherever the file is opened; a muted track is still written (#441).
  • A start trim of a MIDI clip hides notes instead of deleting them, and notes stored before an unlooped clip's start are now silent. Dragging a MIDI clip's start edge later, Trim Start to Insertion and Trim to Selection keep the notes and controller moves they pass, hidden: dragging the start edge back out brings them back exactly where they were played. A note that starts before the edge is silent as a whole, even when it is held past the edge, as it was when a trim deleted it. Existing sessions sound different in one case: a clip that already holds notes before its start without a loop (made by Clear Loop, or by dragging the loop handle back to the first cycle, on a looped clip whose start was trimmed) used to play those notes all at once on its first tick; they are now silent. Copies, pasted ranges, DAWproject exports and clip group files carry nothing the start edge hides; the session file keeps the hidden notes and lists midi_hidden_notes, so an older build that reads session markers warns before it plays them on the clip's first tick (#441).
  • A MIDI clip's start edge drags back as far as the session start, or further to the clip's earliest hidden note or controller move when that is earlier, and stops exactly there, unless another clip on the track is in the way, so a trim back reaches what a start trim hid; it can bring a pre-count zone back before bar 1. The trim drag's preview now also stops at a neighboring clip, where the drop does. The end edge and the loop handle of a clip that ends before bar 1 no longer stretch it to bar 1 (#441).
  • A whole-track print (Track > Bounce in Place on a track, Commit, track.bounce and track.commit) of a track with content before bar 1 says so when it silences that part: the print starts at bar 1, so with the source set to Mute, Mute and deactivate, or Hide, the part before bar 1 no longer plays, and the status line and the command's reply now say that Leave keeps it sounding (#700).
  • Linux, ALSA system: one audio dropout (an input overrun, or an output underrun while the audio thread waits) no longer ends the input for good with a CPU core left busy; the stream recovers after one gap and capture restarts, so monitoring and recording go on. Covered by tests on Linux; awaiting confirmation on hardware (#586).
  • Linux, ALSA system with a separate input device: pressing Record no longer fails at once with "recording start fell outside retained input history" when the start maps less than one input block (at most 100 ms) before the first frame captured after arming; the take starts at that frame and is marked approximate. A start on a separate input that no current input clock map can place (a late input callback, say) waits up to a second for one, then is placed by the last validated map and marked approximate, where it used to fail with "could not be mapped"; a device that records with its output is placed as before. Covered by tests; awaiting confirmation on Linux hardware (#587).
  • After recording with notes played during the count-in, the view scrolls back to show the take's start when the pickups landed left of it, including the first take that makes the pre-count zone. The view used to stay where it was, so the pickup was off screen and looked lost (#707).
  • macOS: the disk image's window background showed the inherited project's wordmark and artwork. The source and image now carry transit's wordmark and icon; not yet seen on a Mac (#727).
  • Linux: a plug-in editor embedded in its window is put back in place when another program moves it there, instead of staying offset or clipped until the window is next resized. (#697)
  • Recording says when a take lost input, and where. When the take writer falls behind until the input queue is full, or the audio system (WASAPI) fills lost input with zeros, the take keeps its timing with silence there, as before; now a loss of 1 ms or more goes to the log when it begins and the status line counts it while recording, and Stop names how long and from how far into the take, for each cause (or that it started before the take began, in a count-in or pre-roll). Stop used to say only "the disk was too slow", a guess, and said nothing at all when a take file also failed; WASAPI's zeros were not reported per take. Invalid input samples (not a number or infinite) recorded as silence are counted too and named at Stop. Not counted: a JACK input port or an ASIO input channel without a buffer, which silences the whole stream rather than a stretch of it. The log also notes any stretch of exact digital silence of 100 ms or more between signal in a take, which names an interface that went silent while still running (#184).
  • ASIO: closing, reopening and stopping streams no longer risk a crash or replayed audio when a driver or a plug-in misbehaves. transit takes the driver's buffers out of use before the driver frees them, also when reopening fails. When a buffer switch arrives while transit is still busy with the previous one, that part of the buffer plays silence instead of repeating the audio from two buffers earlier, and a slow error report no longer makes buffer switches miss their turn. Stopping a stream while transit's ASIO thread is busy (for example while a driver's control panel is open) silences that stream at once and succeeds; the driver stops as soon as the thread is free. A start that finds an audio callback still running fails with a message saying so instead of hanging. If an audio callback (often a plug-in) still has not returned two seconds after its streams close, the driver is left loaded instead of being freed under it, and no ASIO driver opens again until transit restarts; the error says so (#524).
  • A hosted plug-in's own state is better protected. A change a plug-in reports itself (a patch or sample loaded in its own browser) marks the session modified, so Quit, New and Open ask to save; it used to go unnoticed until the next save. Every two minutes transit reads the plug-ins' current settings, also in a session with no unsaved changes, and autosaves when they changed, so a plug-in that never reports its changes is covered too; that read usually takes about a tenth of a second at most, plug-ins it did not reach are read at the next one, and it waits while the transport plays (it runs at Stop) or while a control is being moved. A plug-in whose settings change on every read while nothing else changes is left out of these timed reads for the rest of the session; it is still read before saving, Quit, New, Open and the changes below. Quit, New and Open also read every plug-in's settings before asking. Removing a plug-in, replacing an instrument (by another instrument or an External Instrument), deleting or ungrouping its track, changing the track's width, and an undo or redo (one step or a jump in Undo History) that takes a plug-in away now read its current settings first, so Undo or Redo brings it back as it was at that moment, not as of the last save; changing a track's width also no longer resets its plug-ins to the last saved settings. If the audio engine is still applying an earlier change, transit waits a moment for it; if that is not enough, the change is not made and the status line says to try again. A plug-in that does not answer in time (busy in a dialog, or stuck) keeps the settings it last stored: the change goes ahead with those, and the status line names the plug-in (after removing, replacing, deleting, ungrouping, an Undo History jump or a browser drop). A plug-in that reports a change all the time keeps its session reading modified (#574).

Known issues

  • macOS: when an older transit disk image is still mounted, the new one's window can show the old background, because the window finds its background by volume name. Eject the old one first (#727).
  • Bounce in Place of a take with count-in pickups prints from bar 1. Audio clips cannot yet start before bar 1, so audio played during the count-in stays in the take's file before the clip's start (#442, planned for v0.7.4).
  • Conform to Tempo and Elastic Properties have no menu entry in this release; the per-clip stretch mode replaces them (#336). Sessions that already hold stretched clips play and render them through the new reader.
  • At a converted device rate, starting Play or Record from a stop cuts audio that is already sounding (#725, planned for v0.7.4).
  • A session saved by this release that lists an unplaced take cannot be opened by v0.4.1 through v0.7.2 (see the warning at the top).
  • Converter limits and per-rate figures: device-rate-converter-quality and stretched-clip-reader-limits.

The full list is in the known issues.

Verifying a download

Every file of this release carries a build provenance attestation: a signed record that the file was built by this repository's release workflow from the tagged commit. With the GitHub CLI:

gh attestation verify <file> --repo transit-audio/transit

It says where and from what a file was built, not that the build is free of defects. The macOS build is also signed and notarized; the Windows installer and program are not code-signed yet. Details are in docs/releasing.md.