Every change, in full
The README keeps one line per feature. This page has the rest: where each feature lives in the app, how it behaves, and what it does to a project you already have. Newest first.
Tag 4.0.0-beta.24.0.0-beta.2 LATEST RELEASE
The dubbing and mixing release. Subtitles come in as clips locked to their cues, tracks get a mixer, characters move into a library every project shares and each picks its own engine, and long projects stop lagging on every keystroke. A .tbaw from 4.0.0-beta.1 opens in it; read the box below first.
Before you open an existing project. Clips now have a pause between them, so a project that isn't hand-timed gets longer the first time it opens. Set Gap and Paragraph gap to 0 in Settings > Project for the old placement.
Text now splits into pieces of about 40 words, so clips whose pieces come out different are marked stale. A split_pattern saved in an old presets/*.json file is ignored.
A project that has takes, subprojects or imported audio needs 4.0.0-beta.2 or newer to open.
Each engine now keeps its own language. An Audio8 clip generated while the language box held Kokoro's code is marked stale once.
Engines
-
An Engine row in the Settings tab
The Settings tab starts with an Engine row. With a character selected it sets that character's engine, the same as Edit > Characters. With a clip selected it shows the clip's character's engine, greyed out ("Set on the character"). With nothing selected it is "Engine for new characters": Add in Edit > Characters and a new speaker from subtitle import get that engine (default Kokoro), not the engine of whichever character was selected. An Unknown speaker from a recording import still gets the first engine that clones voices.
When a character's engine changes, its voice stays if the new engine has a voice by that name; otherwise it becomes the new engine's default voice, or its first one, or none (Audio8 with no saved references). Variants stay. The switch can't be undone, and it's refused while a job runs.
Edits to a library character from the Settings tab now reach its library entry, as edits in Edit > Characters already did.
-
The Voices tab picks its own engine
An Engine box at the top of the Voices tab lists the engines with a voice editor (Kokoro's Mixing, Audio8's Voice Reference). Pick one to edit its voices without changing any character. Selecting a clip or character puts the tab back on that character's engine. A character on an engine with no voice editor (Dummy) leaves the tab where it was. Saving or deleting a voice updates the Settings tab's and Edit > Characters' voice lists.
-
Each engine keeps its own settings
Language, default voice, Parallel Threads and an engine's own Model settings (Audio8's temperature, top-p, top-k, max tokens) are stored per engine in
config_qt.json(engines). A project with Kokoro and Audio8 characters generates each with its own language and model settings, whatever is selected; before, both used the ones on screen. Speed, pitch, text splitting, output format and caching are still shared. The first launch moves the old single values across: the language to every engine that has it (Kokoro'sanever reaches Audio8), the voice to the default engine, the thread count to all. -
A project whose engine isn't installed
A character on an engine this install doesn't have keeps it. Its clips play from the audio they have and aren't marked stale while the files are there. They don't generate: Generate skips them and says "<engine> isn't installed: N clip(s) skipped", and the gutter shows no Generate button, with the reason as the tooltip. Edit > Characters and the Settings row show the engine as "(not installed)". Save and Save As keep the character's engine and the engine's files in the
.tbaw, so the project comes back whole where the engine is installed. (Save As used to drop entries it didn't recognise; it now copies them from the file the project came from.) Before, such a character silently switched to Kokoro and a Generate replaced its voice. -
Kokoro is optional, and engines can be plugins
Without the
kokoropackage the app still starts, with Audio8 and Dummy; a Kokoro project opens, plays and says Kokoro isn't installed. A separate package can add an engine through thekokorogui.enginesentry point. It appears in every engine picker, and one that fails to load is listed as not installed instead of stopping the app. CONTRIBUTING.md, "Adding an engine", has the steps.
Characters and projects
-
A character library shared by every project
Characters can live in a library (
characters/, one file per character) as well as in a project. A project's character linked to the library follows it: change its voice, FX or color in Edit > Characters and every project using it picks that up the next time it opens, or at once if it's open in another window. Each project keeps its own name for the character, and "Rename in library" copies it over.Characters made in a project stay local to it until you press "Promote to Library". Add > From library... brings library characters into a project. A project still carries its own copy of every character, so it opens and plays on a machine whose library doesn't have one. The dialog then says "Library (not found here)".
-
New starts from the library
File > New gets every library character instead of a copy of the last project's. With an empty library you get one "Default" character. The first launch of this version copies your
presets/*.jsonfiles into the library (the files stay), and characters in the project it opens that match a preset exactly are linked to it. -
Each character picks its engine
The Engine box in Edit > Characters replaces Options > Engine, so Kokoro and Audio8 voices can share a project. The Settings and Voices tabs follow the selected clip's character.
-
Voice variants for Audio8
A character can have named variants ("angry", "whisper"), each its own reference recording, set up in Edit > Characters. The transcript's Variant box picks one per clip.
-
Subprojects
A project can hold other projects. File > New Subproject turns the selected text into a subproject embedded in this one (a chapter, an episode segment, a scene), and File > Add Subproject... links an existing
.tbaw. The parent shows each as one read-only line and one block that plays the subproject's exported mix.Click the block to edit the subproject in the transcript and Settings tabs. Double-click to enter it on the timeline, with a breadcrumb back. Generate works through stale subprojects after the parent's own clips. New from text makes an EPUB one subproject per chapter. Characters come in three scopes: Global (the library), Project (shared by every subproject of a book) and Local.
Timeline and mixing
-
Pauses between clips
Clips placed one after another have 0.35 s of silence between them, and 0.9 s across a blank line. Both are in Settings > Project (Gap, Paragraph gap). A clip can override its own gap, and
[pause:1.5]in the transcript gives the next clip Auto-split creates a 1.5 s gap. The marker stays in the text and is never read aloud. -
Tracks appear when a character is used
A character gets a timeline track the first time you assign it text, and a track with no clips isn't drawn, so a big library doesn't open as dozens of empty lanes. An unused track keeps its mute, fader and pan for when it comes back.
-
Unified track layout
Settings > Project > Track layout > Unified puts every clip on a few lanes (3 by default) and moves to the next lane each time the speaker changes, so a conversation alternates lanes. Mute, solo, fader and pan then act per lane. Switching back puts each clip on its character's track. Either switch is one undo step.
-
Takes
Regenerating a clip keeps the old version. Right-click a clip > Take to switch back, or Delete take to drop one. A take recorded from text you've since edited says "old text". Old takes and their audio are saved in the project.
-
Ripple on regenerate
When a regenerated clip comes back longer or shorter, clips you placed by dragging that sit after it move by the same amount. Right-click a clip > Lock in time to keep it where it is, or turn ripple off in Settings > Project. Clips that overlap on one track get a red border.
-
Mixer
Each track has mute, solo, a fader, a pan slider and a volume automation lane. The
Abutton shows the lane: double-click adds a point, drag moves it, right-click deletes it, Alt-drag moves a segment. Clips have fade-in and fade-out handles on their top corners. Settings > Project > Auto-crossfade adds a 10 ms fade where clips overlap. Playback and export are stereo, and export follows mute, solo, pan, fades and automation. -
Markers and loops
Right-click the ruler to add a marker, and drag a flag to move it. Shift-drag on the ruler, or use "Loop to next marker", to loop a region. Export can write just the stretch between two markers.
-
Timecode
Settings > Project > Timecode shows the ruler and the transport clock as
HH:MM:SS:FFat a frame rate you pick, drop-frame included, from any start time. -
Convolution reverb
Audio FX > Spatial & Time > Convolution Reverb puts a clip in a recorded space. Pick an impulse response and set Mix (default 0.5). "Add..." copies a
.wavinto the impulse response library (presets/fx/ir/). Save bundles every impulse response the project uses, and a project's own copy wins over the library's. A missing one, or one longer than 30 seconds, plays dry and logs a warning.
Dubbing and subtitles
-
Subtitle import
File > Import Subtitles... reads SRT, WebVTT and ASS/SSA files. Each cue becomes its own paragraph at the end of the transcript and a clip locked in time at the cue's start, with the cue's text as the clip's source text and the cue's length as its target.
When the file names speakers (an ASS Name field, a VTT
<v Name>tag), a dialog asks once which character voices each: the narrator, a new character named after the speaker, or one you already have. Cancel imports nothing, and one undo removes the whole import, including the characters it made. Ripple never moves a cue. Dragging one moves it and it stays locked. -
Locked clips line up on their first word
Settings > Project > "Align locked clips to their first word" starts a locked clip a little early, by the silence before its first word, so the word lands on the clip's time instead of the breath before it. It's on by default in a project with a locked clip, and skipped for a clip with Trim Silence on.
-
Target durations and Fit to slot
Settings > Clip > Target (s) gives a clip a length to fill, and subtitle import sets it from each cue. The timeline draws that slot as a bracket, puts how full it is in the clip's label as a percentage, and tints the clip amber past 103% and red past 115%.
Right-click > Fit to slot changes the clip until it's within 3% of the target. Kokoro regenerates it at a new speed (0.7x to 1.6x, up to three tries). Audio8 has no speed control, so it time-stretches on playback (0.87x to 1.15x). A line still too long at that limit is marked Needs rewrite. Each fit is one undo step. "Fit all over slot" above the timeline, shown once any clip has a target, fits every clip that runs long.
While you type, a stale clip whose text reads longer than its target gets a wavy amber or red underline, judged from how fast each character has spoken in this project.
-
Original dialogue to dub against
File > Import Source Track... copies the original recording into the project. Settings > Clip > Original (s) holds the part of it a clip dubs (
start - end), and subtitle import fills it from each cue's times. The transport's Dub / Original / Both switch picks what you hear: the dub, the original under each clip, or both at -6 dB each, switchable mid-play. It stays on Dub, greyed out, until the project has a source track.Settings > Project shows the source track, a Source offset (where the cue times start in the file) and Remove. Export is still the dub alone.
-
Reference video
File > Load Video... (MP4, MOV, MKV, WebM, AVI, M4V) opens a Video tab beside Settings that plays the video muted and follows the transport's play, pause, stop and seek. Its Offset box sets which video time sits at the timeline's 0. Workspace > Simple hides the tab.
The project keeps the video's path, relative to the
.tbawwhen it can, and asks you to find the file if it has moved. "Bundle the reference video in the project file" in the Export dialog puts the video inside the.tbawinstead (off by default). Save only copies a video you picked with Load Video on this machine. A project someone sent you keeps the copy it came with. -
Source text
A clip can carry the original line it's dubbing. It shows in the transcript gutter's tooltip and next to a rough syllable count of both lines.
-
Review
Each clip has a status (to do, generated, approved, needs rewrite) and a note, set from the clip's right-click menu or Settings. The timeline's Show filter dims the rest. Export can write a cue sheet (
.csv) with times, character, source text, text, status and note. -
Word timing
Generated clips remember when each word is spoken. Kokoro reports its own timings in English. Audio8 and Kokoro's other languages get a Whisper pass after generating. During playback the transcript highlights the current word, Ctrl+click on a word jumps there, and export can write one subtitle per word.
Imported audio
-
Music beds with ducking
File > Import Audio... > "Music bed" adds a WAV, FLAC, OGG, MP3 or AIFF file on a Music track, locked in time at 0:00 or at the playhead, with its file name as a read-only line at the end of the transcript. Drag a bed's edges to trim it. Right-click for Loop (then drag the right edge to set how long it runs), Reset trim, Lock in time and Remove.
The
Dbutton on a track's header ducks it: the track drops while clips on tracks withoutDplay, by Settings > Project > Ducking (default -12 dB). The original dialogue neither ducks nor is ducked. Export sounds the same as playback. -
Edit a recording as text
File > Import Audio... > "Recording to edit as text" puts a recording's words in the transcript, each tied to where it was said. The transcript comes from Whisper (in the background, with progress) or from an SRT, VTT or ASS caption file, optionally with "Refine word timing with Whisper". A review step lists each clip with a play button, and a corrected word keeps its place in the recording. You pick one speaker for the whole recording from your voice-cloning characters, or "Unknown speaker". A caption file's speaker names go through the same mapping dialog as subtitles.
Then edit the text and the audio follows. Delete words and the recording closes up with a 5 ms crossfade. Cut, copy, paste or drag words and their audio moves with them, into another project too. Text you type inside a recording has no audio: it shows grey with a hollow circle in the gutter and splits the recording around it until you assign a character, which makes it an ordinary TTS clip. Recorded words have a faint underline and recording clips get a play button in the gutter. The recording is saved in the project under
audio/imported/.
Interface
-
A Settings window for program and project settings
Options > Settings… (Ctrl+,) opens a window laid out like OBS's: a page list on the left (General, Generation, Performance, Project, Timeline, Source track), the page on the right, and OK, Cancel and Apply at the bottom. Edits wait for Apply or OK, and Cancel drops them. Apply stays greyed out until a value differs from what the window opened with.
The Settings tab used to hold these too. It now keeps only what a voice sounds like: Engine, Language, Voice, Speed, the engine's Model fields, and Audio control. With a clip selected it still shows the clip's own fields (gap, take, status, note, source text, original range, target). What moved:
- General: Theme and Device, the same choices as their Options submenus.
- Generation: Target words per segment, the three Split at toggles, Output format.
- Performance: Enable segment cache, and per engine Parallel threads and Audio8's Cache reference encoding.
- Project: the subproject title, Gap and Paragraph gap, Auto-crossfade, Ripple on regenerate, Align locked clips.
- Timeline: Track layout, Ducking, Timecode.
- Source track: its offset and Remove. Remove now waits for Apply.
The Engine row stays at the top of the Settings tab so an engine and a voice are picked together. With a clip selected it's no longer greyed out: it switches the clip's character's engine, the same as selecting the character. With nothing selected it's still "Engine for new characters", and the Language and Voice under it now show that engine's defaults (what a new character starts with) instead of the first character's engine.
Nothing is stored differently: the app values stay in
config_qt.jsonand the project values in the project's settings, so existing projects and configs open as before. Each project value you change is still its own undo step.
Text and generation
-
The Lexicon applies to timeline clips
Before, only whole-document and JIT generation used it. Adding or removing a rule now marks the clips whose text it rewrites as stale, and no others.
-
Text splits at natural pauses, toward a word count
Settings > Target Words per Segment (default 40) replaces Split By. Each piece ends at the strongest boundary near the target: a paragraph, then a sentence end, then a pause (comma, semicolon, colon, dash, line break), each with its own on/off switch. A sentence is only cut when it runs past twice the target, and a word never is. Each piece is one file and one subtitle cue, so a long paragraph no longer leaves a clip stale forever.
-
Transcript details show segments, gaps and clip status
Options > Transcript details > Show details (off by default) adds overlays to the Transcript panel, each with its own toggle in the same menu. Segment boundaries: each segment of a clip alternates between the character's tint and a lighter one; a grey bar marks a cut at a pause and an amber bar a forced cut between two words, where the voice resets mid-phrase. A stale clip shows the segments Generate will make and updates as you type. Hover a segment for its number, word count, what it ends at, and its length and take (an estimate while stale). Clip info: a status dot and the clip's length in the gutter, and a line under the editor for the clip under the caret: character, engine and voice, segment, length, take, status and gap. Lexicon rewrites: a dotted line over text the Lexicon changes, with what it's spoken as in the tooltip. Gaps: the silence before each clip that starts a line, on its boundary rule. These are app settings in
config_qt.json; projects don't change. -
Whisper is the default auto-transcription engine
The Voices tab's Auto-Transcribe (for Audio8 References) runs Whisper large-v3-turbo locally through faster-whisper (MIT). It asks before the first-use download (about 1.6 GB) and offers another engine instead. Set
WHISPER_MODELin.envto pick a smaller model.python -m kokoro_gui.engine.asr ref.wav whisper --wordsprints timed words. If you picked Audio8 in the dock before, that choice stays until you change it. -
Inline tags are no longer spoken
A clip made from a
[Speaker]:or[Speaker:FX]:line, by Generate > Auto-split then generate or by typing the line and pressing Enter, covers its tag. Before, the engine read the tag aloud, and the splitter could make the tag a segment on its own (a name with spaces, such as[Old Man:Big Hall]:, always did). Now every clip's text drops its tags and[pause:x]markers before the Lexicon and the splitter run. The tag stays in the transcript and keeps the character's color, and subtitles and the cue sheet leave it out too. Projects don't change, but a clip whose text holds a tag is marked stale once and needs Generate. -
A tag's FX applies to its clip
The FX half of
[Alice:Radio]:used to be ignored on clips. Auto-split and the typed-line shortcut now set the new clip's FX preset toRadio, the same as picking it in the Transcript panel's FX box, in the same undo step as the clip. Change it in the Audio FX tab afterwards; editing the tag text doesn't change it. If no preset has that name, the clip is still made with its character's FX, and auto-split names the missing presets in a warning (the typed line says so in the status bar).
Speed
-
Typing and scrolling stay fast in long projects
The Transcript and Timeline panels no longer recheck every clip on every keystroke, scroll step or repaint. Each clip's stale flag is remembered and rechecked only when its text, character, overrides, segments or the generation settings change. The timeline updates only the clip blocks that changed, and it waits until you pause typing (60 ms) before updating. A waveform is decoded the first time its block scrolls into view. Measured on a 1,500-clip project (about 43,000 words): a keystroke went from 11 s to 17 ms, a scroll step in the Transcript panel from 1 s to 4 ms, a timeline refresh from 8 s to 23 ms, and redrawing the transcript highlights after a settings change from 1.7 s to 15 ms.
-
Options > Force refresh
Throws away everything the panels remember (stale flags, clip placement, post-processing settings, decoded waveforms, file checks) and rebuilds it from the project. Use it when a panel looks out of date, or after changing a project's files in another program while KokoroGUI kept focus.
-
Files deleted outside the app count when you come back to it
A segment file deleted by hand still marks its clip stale, but KokoroGUI now notices when its window becomes active again, after Generate, Open, Save and imports, and on Force refresh, instead of on every repaint. Nothing in the project file changes.
-
An FX change no longer freezes a long project
Changing a project-wide FX value, or the Settings tab's volume, pitch, normalize or trim, used to render every generated clip on the main thread when the transport reloaded, and the window stopped answering until it finished. Two background threads do that work now. A waveform on screen fills in when its render is ready, and the transport reloads once the renders are done, so the old sound keeps playing until then. On a 1,500-clip project with 750 generated clips, the longest freeze after switching reverb on went from 1.7 s to 0.14 s.
-
Undo copies what the edit changed
Assigning a character to a range, and the timeline's sub-range text replace, saved a deep copy of every clip and run so Ctrl+Z could put them back. They now save the runs around the edit and the clips those runs belong to. In a 1,500-clip project an assignment went from 50 ms to 6 ms and a replace from 47 ms to 5 ms. An edit that reaches more than 50 clips, or a project with imported recordings, still saves everything. Nothing in the project file changes.
Security
-
Private cache and voice folders
On Linux and macOS,
cache/,custom_voices/and the Audio8 reference folders are created owner-only (0700), and tightened to that if they already exist. Acache/that belongs to another user is left alone and a per-user one under~/.cache/kokorogui/is used instead. -
Impulse response names with a colon are refused on Windows
Windows read a convolution reverb name such as
a:b(from a project or an FX preset) as a drive-relative path and loaded the impulse response calledb. Such a name is now refused there, as it already was elsewhere, and the reverb plays without it.
Tag 4.0.0-beta.14.0.0-beta.1
The rebuild. 3.2.0 was a CustomTkinter form over one kokoro_engine.py. 4.0.0 is a PySide6 app over a transcript with clips, tracks and characters, pluggable engines with a second real one, and projects that live in one file.
The app
-
Qt frontend, the only frontend
python main.pyorrun.batlaunches a PySide6 window of dockable panels: the transcript, the Settings, Audio FX, Lexicon and Voices tabs, the timeline and the transport, under File / Edit / Options / Workspace menus. Every panel can be dragged, stacked into tabs or floated. Workspace > Advanced / Simple / Reset layout are saved layouts, and Simple hides the timeline so the transcript gets the full height. The CustomTkinter app (gui.py) is gone, and PySide6 is a regular dependency. -
Welcome screen
Launch opens the last project, then puts a dialog over it: recent projects (the open one first, Resume as the default button), New project, New from text file, Open other, right-click to drop a row, Clear list, and a details pane with the file's path, modified time, character and clip counts, audio length and engines. Untick "Show at startup" for a silent resume. File > Welcome... brings it up any time.
-
Undo and redo
Edit > Undo/Redo over a plain-Python undo stack. Typing undoes like a normal text editor. Character and FX assignments and timeline moves have their own history, and Ctrl+Z reverts whichever happened most recently.
-
Dark and light themes
Options > Theme, dark by default. One palette module feeds the custom-painted widgets, the Qt palette and a stylesheet for every control. The UI font is Segoe UI / Inter / Noto Sans at 10pt, the transcript one point larger. Timeline clips have rounded corners, a waveform in the clip's own darker shade and a black or white label picked by contrast. The default character palette is eight hues at one lightness and doubles as the color picker's presets.
Transcript and timeline
-
A project, not a text box
The old "generate this text" input is a project: text with clips, tracks and characters layered on top. The transcript is the source of truth and the audio is generated from it, tracked per clip by a hash of what produced it. Generate works on every stale clip in one pass instead of the whole transcript every time. A project with no clips yet still generates the whole transcript at once. Auto-split turns a
[Speaker:FX]-tagged transcript (optionally per paragraph) into clips and generates them in one action. -
Transcript with character highlighting and a live gutter
Each run is tinted by its character, so you see speaker boundaries without reading the inline
[Speaker:FX]:syntax, which turns into a real assignment the moment you finish a tagged line. Character and FX combos above the editor show the caret's clip and reassign the selection or the whole clip. The gutter labels each character or FX change and shows a play button beside each stale clip. Stale text is dash-underlined, and thin rules show where clips end and where Auto-split would cut. Copy and paste carry the character assignment along. -
A multi-track timeline on a real seconds axis
One lane per character. Clips sit end to end in text order at their real duration once generated, and at an estimate (dashed outline, no waveform) before. A ruler with a playhead, a fixed track-header column, Ctrl+wheel zoom. Dragging a clip pins it to a time or moves it to another character's track, and dragging it before an earlier clip also moves its text there. Shift+drag inside a clip carves out a sub-range and replaces it with fresh TTS under any character. Each clip has its own FX button.
-
Playback
Play, pause, stop, click the ruler to seek, a playhead across all lanes, and the transcript highlights and scrolls to the clip being played. Space toggles playback anywhere but the text editor, and Ctrl+Space toggles it everywhere. One
sounddevicestream mixes the timeline in its callback, so the position is sample accurate. Loop toggle included. -
Export
File > Export mixes every clip down to one file (wav, mp3, flac or ogg) at its timeline position, optionally with a
.srtand per-clip files (<base>_001_Narrator.wav). It warns when clips are stale and offers to generate first. The dialog also holds the project's two bundle options: whether to bundle generated audio, and the audio format for new segments.
Settings and FX
-
Settings and Audio FX follow the selection
Voice, speed, volume, pitch, normalize and trim live in a Settings tab that reads and writes whatever's selected: the project's defaults, one clip's overrides, or a character. Editing a character affects every clip using it unless that clip has its own override. Audio FX works the same way, with a prompt before changing a character several clips share. The timeline's FX button selects the clip and raises the tab.
-
Audio FX are non-destructive
Clips are generated as raw model output. The FX chain, volume, pitch, normalize and trim are applied when the transport, the export or the timeline waveform reads them. Move a slider, pick an FX, toggle "Apply": you hear it on the next play, and the clip stays generated. A clip with its own FX override counts as FX-on even if its character's FX is off.
-
Characters replace bare presets
Existing
presets/*.jsonfiles become characters on first load (one per file, or a single "Default" seeded from your last settings if you had none), each with its own highlight color. Edit > Characters... edits name, color, voice and FX. The preset files themselves are untouched. Output folder, filename and format moved from the Settings tab to Export.
Projects
-
.tbawproject bundlesA project is one zip file with the text and clips, every generated segment, and every Mix, Reference and saved FX it uses, so it opens on another machine with the same engines installed. Save writes the file in the background and never leaves a half-written project behind. Save As keeps the same working copy. Autosave writes only into the app's own working copy (
cache/projects/), so the file on disk is as new as your last Save.Closing with unsaved changes asks Save / Discard / Cancel, and after a crash the next open offers to recover the unsaved session, even if the file was renamed or moved. The window title names the project and shows
*while it has unsaved changes. A bundle only ever names audio inside itself, and Save never copies a file from outside the project's working copy into it..jsonprojects from the 4.0 previews still open and are converted on the spot, with a.tbawwritten next to the untouched.json. -
Generation writes once
A clip's audio lands straight in the project's working copy under a name derived from what produced it, instead of one copy in
cache/and another in the output folder. Regenerating an up-to-date clip (the gutter button) makes a fresh take under a new name. Opening a project made with another version of an engine keeps its clips clean and says so in the status line. -
Segment cache rekeyed
Cache entries are keyed on the voice's name and content rather than its path, so the first generate after upgrading from 3.2.0 misses the old
cache/entries.
Engines
-
Pluggable engines
kokoro_gui/engines/defines an engine interface (settings fields, voices, capabilities, project hooks) with three registered engines: Kokoro, Audio8 and a sine-tone Dummy that exists to prove the interface isn't Kokoro-shaped. Options > Engine swaps the Settings tab's fields and the Voices tab live.kokoro_engine.pyis a slim core backed by akokoro_gui/engine/package split by feature. -
Audio8, a voice-cloning engine
Audio8-TTS-Preview-0.6b clones a voice from a reference WAV plus a transcript of what's said in it. The Voices tab lets you browse a WAV, auto-transcribe it, edit the transcript and save it under a name that then shows up in the Voice dropdown. It needs
transformersandtorchaudio, loads withtrust_remote_code=True, and downloads from Hugging Face on first use. -
Auto-transcription for References
Auto-Transcribe has an engine picker, also runnable as
python -m kokoro_gui.engine.asr <wav>. In this release the default was Audio8-ASR-0.1B (CC-BY-NC-4.0, non-commercial). The alternative is Vosk, fully offline and Apache-2.0, which needs a model folder downloaded by hand. Its path lives inVOSK_MODEL_PATHin.env, which the tab's Browse / Save / Reload buttons edit. The reference is converted to the 16-bit mono PCM Vosk requires, whatever format it starts in.
Tag 3.2.03.2.0
The last CustomTkinter release. Frozen, no further fixes.
-
Cross-platform audio playback
Preview and JIT playback go through
sounddeviceandsoundfileinstead of the Windows-onlywinsoundmodule.
Tag 3.1.03.1.0
JIT generation
Streams audio as it's generated, so you start listening straight away.
Audio FX
Pedalboard reverb, compression and EQ.
Pronunciation lexicon
A custom dictionary that overrides how specific words or acronyms are pronounced.
Voice mixing
Combine existing voices into a new one, a Mix.
Tags
[Speaker:FX]: Textswitches voice and effects mid-transcript.Caching
Generated segments are cached so repeated runs skip them.
Windows quick start
run.batstarts the app with one click.