Edit the words.
The audio follows.
KokoroGUI is a text to speech editor where the transcript is the project. Every line belongs to a character, every character has a voice, and the timeline underneath is generated from the text. Change a word and only that clip generates again. Everything runs on your machine, free and open source.
Four moves, and the text stays in charge
[Marta]: in front of it. Each character has a voice, an engine and an FX, and its text is tinted in its color.A TTS editor built on the script, and the only one built this way
Text to speech apps give you a text box and a file. Audio editors give you a timeline and no text. KokoroGUI joins the two at the clip: the words decide what each clip says, and the timeline decides where it plays. Edit either one and the other keeps up.
| Typical TTS app | Typical audio editor | KokoroGUI | |
|---|---|---|---|
| Where you work | A text box | Waveforms on a timeline | The script, with a multitrack timeline built from it |
| Change one word | Generate the whole file again | Re-record it, or splice by hand | That line regenerates. Nothing else does |
| Several speakers | One voice per run, stitched by hand | One track per recording | A character per line, each with its own voice and engine |
| Effects | Baked into the output | Non-destructive | Applied at playback, so a slider move never costs a regenerate |
| Versions | Overwritten | Saved copies | Takes per clip, with review status and notes |
| The project | Loose audio files | A session folder | One .tbaw file holding text, audio, voices and presets |
One project format, four kinds of work
Pick what you're making. The same transcript, timeline and .tbaw file handle all of them; what changes is which tools you reach for.
Drop in an EPUB and get one subproject per chapter, with the same cast reading every one.
- A book is a project, a chapter is a subprojectFile menuNew from text file splits an EPUB, or a PDF with a table of contents, at its chapters. Generate works through every stale chapter in turn.
- One cast for the whole seriesCharactersThe character library holds the narrator and every regular. Change a voice there and every book that uses it follows.
- Sentence-sized piecesSettingsText is generated about 40 words at a time and cut at paragraph, sentence or pause boundaries, never mid-word, and never after "Dr.".
- Names said rightLexiconFind-and-replace rules fix a name everywhere it appears, and mark only the clips they change as stale.
- Chapter files and subtitlesExportExport a file per chapter, or just the stretch between two markers, with an
.srttimed to the audio.
Import the subtitles, hear the original under each line, and fit every new line into its slot.
- Subtitles in, clips outImport SubtitlesSRT, WebVTT and ASS/SSA. Each cue becomes a clip locked to its time, and named speakers map to characters in one dialog.
- Fit to slotTimelineEach clip shows how full its slot is. Kokoro regenerates at a new speed until it fits within 3%; Audio8 time-stretches on playback.
- Hear the originalTransportImport the source track and switch between Dub, Original and Both mid-play. The original under each clip is its cue's slice.
- Picture to dub againstVideoLoad an MP4, MOV, MKV or WebM; it plays muted and follows the transport, with an offset for leaders.
- Words land on the cueProject settingsA locked clip starts early by the breath before its first word, so the word hits the cue's in-time. Timecode on the ruler, drop-frame included.
Cut a recording by deleting its words, then score it with music that ducks when anyone speaks.
- Edit a recording as textImport AudioWhisper or a caption file puts the words in the transcript. Delete words and the audio closes up; cut, paste and drag carry the audio with them.
- Music beds that duckTrack headerImport WAV, FLAC, OGG, MP3 or AIFF as a bed, trim or loop it, and press
Don its track to drop it by 12 dB while voices play. - A real mixerTimelineMute, solo, fader, pan and a volume automation lane per track; fade handles on every clip; stereo playback and export.
- Rooms and phonesAudio FXConvolution reverb takes your own impulse responses; the GSM codec makes a phone call. Seventeen Pedalboard effects in all.
- Conversations on shared lanesTrack layoutUnified layout alternates a few lanes every time the speaker changes, instead of a track per character.
Hundreds of short lines, several takes each, and a sheet that tells everyone which ones are approved.
start,end,character,source_text,text,status,note 00:00:00:00,00:00:01:12,Guard,,"Halt! Who goes there?",approved, 00:00:02:03,00:00:03:05,Guard,,"Did you hear that?",needs rewrite,too calm
- TakesClip menuRegenerating keeps the old version. Switch back from the clip's menu; a take from text you've since changed says "old text".
- Review status and notesClip fieldsTo do, generated, approved or needs rewrite, with a note. The timeline's Show filter dims everything else.
- Voice variantsCharactersAn Audio8 character can carry "angry" and "whisper" References; the transcript's Variant box picks one per line.
- Files for the engineExportPer-clip files named
<base>_001_Guard.wav, a cue sheet CSV, and subtitles with one cue per word if you want lip-sync data. - Frame-accurate timesTimecodeThe ruler and clock show
HH:MM:SS:FFat your frame rate, from any start time.
Everything is a layer on the text
Voices, clips, takes, FX and word timings all hang off the transcript. Edit it however you like and the layers keep up.
Transcript
Each run of text is tinted by its character. The gutter names the speaker when it changes and puts a play button beside every stale clip, so you can regenerate one line without touching the rest. During playback the spoken word lights up; Ctrl+click a word to jump there.
Transcript reference →Timeline
Clips follow the text in order with a pause between them, or sit where you drag them. When a regenerated clip comes back longer, the clips after it ripple. Markers, loops, timecode, and Shift+drag inside a clip to replace a few words with fresh speech.
Timeline reference →Mixer
Mute, solo, fader, pan, a volume automation lane and ducking on every track. Fades on every clip. Playback mixes in stereo in the audio callback, and export uses the same mixer, so what you hear is what you get.
Playback reference →FX
Seventeen Pedalboard effects, a convolution reverb among them, set for the project, a character or one clip. They're applied when a clip plays, so a slider move is heard on the next pass and never marks anything stale.
FX reference →Characters
A character is a name, a color, a voice, an engine and an FX. Keep them in a library every project shares, or local to one file. Kokoro and Audio8 characters can talk to each other in the same scene.
Characters reference →Tags
Type [Marta]: at the start of a line and the line is Marta's when you press Enter. [pause:1.5] puts a second and a half before the next clip. Auto-split turns a whole tagged transcript into clips in one go.
Two local engines, picked per character
Each character chooses its engine in Edit > Characters, so a Kokoro narrator can introduce an Audio8 clone. The Settings and Voices tabs follow whichever character's clip you've selected.
| Kokoro hexgrad/kokoro · default | Audio8 Audio8-TTS-Preview-0.6b | |
|---|---|---|
| Voices | Built-in voices, plus your own Mixes from the Voices tab | Cloned from a Reference: a short WAV and its transcript |
| Languages | American and British English, Spanish, French, Italian, Portuguese, Japanese, Chinese | Whatever the reference speaks |
| Speed | 0.5x to 2x, and Fit to slot regenerates at a new speed | Its own pace; Fit to slot time-stretches instead |
| Word timing | Reported by the engine in English | From a Whisper pass after generating |
| Sample rate | 24,000 Hz | 44,100 Hz |
| Parallel generation | One pipeline per thread | One shared model, one segment at a time |
| Licensing | Apache-2.0 | Check the model card. Transcribing a reference uses Whisper (MIT) by default; Vosk is the offline Apache-2.0 option |
Raw on disk, FX on the way out
A generated clip is saved as the model made it. These six steps run, in this order, whenever it's played, drawn on the timeline or exported.
Trim
Leading and trailing silence, if on
Volume
A plain gain
Pitch
By resampling, so length changes too
Stretch
Only when Fit to slot set one
FX chain
Every enabled Pedalboard effect
Normalize
Peak to 98% of full scale
The format is named after the idea
.tbaw stands for text-based audio workflow. It's a zip with a manifest: the text, the clips, every generated segment, and the voices and presets the project uses.
- Opens on another machine. Mixes, References, saved FX and impulse responses ride inside it.
- Never half-written. Save runs in the background into a temp file and swaps it in. Autosave writes to a working copy, not your file.
- Recovers after a crash. Reopen the project and it offers the unsaved session, even if the file was renamed or moved.
- Small when you want it. Leave the generated audio out and the file is just text and settings; clips regenerate on open.
manifest.json format, version, stats, hashes document.json text, clips, tracks, characters project.json export and bundle options audio/generated/ 3f9a…c2_0.wav one file per segment audio/imported/ 9b1e…04.flac music beds, recordings fx/ Radio.json saved FX ir/kitchen.wav impulse responses projects/ chapter-2.tbaw embedded subprojects engines/kokoro/voices/ narrator_mix.pt Kokoro Mixes engines/audio8/refs/ marta.wav Audio8 Reference marta.txt and its transcript
The latest release
The short version. Every change in full, back to 3.1, is on the changes page.
Dubbing, a mixer and a shared cast
Import subtitles as clips locked to their cues, fit each line to its slot, and dub against the original dialogue with a reference video playing along. Tracks get mute, solo, faders, pan, automation and fades; clips get takes, pauses between them and ripple on regenerate. Characters can live in a library every project shares and pick their own engine, and a project can hold subprojects.
Also new: music beds with ducking, editing a recording by editing its text, convolution reverb, word-level highlighting, markers, loops and timecode, Whisper as the default transcriber, and a Settings window under Options that holds the program and project settings. FX changes now render in the background, so a long project no longer freezes.
A document, not a text box
The rebuild: a PySide6 app over a document with clips, tracks and characters, a timeline in seconds, non-destructive FX, undo and redo, projects in one .tbaw file, and Audio8 voice cloning beside Kokoro.
Questions people ask first
What is KokoroGUI?
A free, open-source desktop app for text to speech that works as a text-based audio editor. You write or import a script, assign each line to a character, and KokoroGUI generates the audio onto a multitrack timeline you can mix and export. It started as a GUI for the Kokoro TTS model and now runs Audio8 voice cloning beside it.
Is there a GUI for Kokoro TTS?
Yes, this is one. KokoroGUI runs hexgrad/kokoro locally with every built-in voice, voice mixing, speed and language settings, and a full editor around it. You don't need to write Python to use it.
What is a text-based audio editor?
An editor where you change audio by changing its words. In KokoroGUI that goes both ways: edit the script and the changed lines regenerate; import a recording and Whisper puts its words in the transcript, so deleting a word cuts it from the audio.
Is it free? Does it run offline?
It's free under Apache-2.0. The models download from Hugging Face the first time you use them, and after that generation runs on your own CPU or GPU with no account and no upload.
Can it clone a voice?
Yes. The Audio8 engine clones a voice from a short WAV and its transcript, and KokoroGUI can transcribe the WAV for you. A Kokoro character and an Audio8 character can speak in the same scene.
What can I make with it?
Audiobooks from EPUB or PDF, dubbed video from subtitle files, podcasts and radio drama with music beds, and game or animation dialogue with takes and a cue sheet. See the workflows.
Which systems does it run on?
Windows, macOS and Linux, anywhere Python 3.11 and eSpeak NG install. A GPU speeds things up but isn't required.
From clone to first line
Python 3.11 or newer and eSpeak NG. Models download from Hugging Face the first time you use them.
4.0 is a beta. 4.0.0-beta.2 is the latest release tag. --branch takes a tag name, and dropping it gets the default branch. The aim is for every 4.x release to open a .tbaw from any earlier 4.x.
Looking for 3.2.0, the old one-text-box app? It's frozen. Clone the 3.2.0 tag with --branch 3.2.0 and follow its own README. Only presets/ and custom_voices/ carry over.
# clone the latest beta tag git clone --branch 4.0.0-beta.2 --depth 1 https://github.com/CoffeeMethod/KokoroGUI.git cd KokoroGUI # virtual environment python -m venv .venv .venv\Scripts\activate # Windows source .venv/bin/activate # macOS / Linux # install and run pip install -r requirements.txt python main.py # or run.bat on Windows
# Windows: run the installer from https://github.com/espeak-ng/espeak-ng/releases # macOS brew install espeak-ng # Debian / Ubuntu sudo apt install espeak-ng # Fedora sudo dnf install espeak-ng
Python 3.11+
The only hard version requirement.
eSpeak NG
Kokoro needs it to turn text into phonemes. Install it before the first run.
A GPU is optional
Everything runs on the CPU. Options > Device picks CUDA when PyTorch finds one.
First use downloads models
Kokoro and Audio8 come from Hugging Face; Audio8 loads with trust_remote_code=True. Whisper is about 1.6 GB and asks before it downloads.
Smaller or offline transcription
Copy .env.example to .env. Set WHISPER_MODEL to small, base or tiny, or VOSK_MODEL_PATH to use Vosk.
Trouble with torch?
Follow pytorch.org for your OS and GPU before pip install -r requirements.txt.