Celune 4.1.0 - Asynchronous Pipeline Update

This update fixes a major gripe we all had with prior versions of Celune.

Suppose you wanted to play a sound via the pipeline, and then wanted to speak over it.
This was not possible in prior versions because sound effects locked the pipeline for the whole duration of the sound.

Not anymore. The new reworked pipeline mixes multiple sound sources together, allowing your character to speak while other sounds are playing. It also automatically applies ducking, so their voice is the prominent subject, and isn't covered by background noise.

Apart from that, we've also added a way for you to play sounds from a vast library of sound effects called YouTube, and a new VoxCPM2 competitor backend based on dots.tts MF, which provides the best sound fidelity, at the expense of reduced speed.

It uses a forked library, so please don't try to use the upstream one. It doesn't work.

There are also new system sounds, which will give you an easy to tell auditory cue about what happened with the app state. The progress bar was also changed to show the buffer state, rather than token-based metrics that may not always be available.

Since this update came out on a new moon, does this mean she's gonna come back up as a better self? Only time will tell.

Celune 4.0.0 - Personality Update

"Now, I don't just speak. I talk back to you."

Yes, you heard her. This update finally allows your characters (not just the default Celune) to talk back to you.

We've added a lot of features about the personality. The CEVOICE spec has been promoted to the CECHAR spec, and upgraded to a version 2 (point 06, haha), allowing it to contain not just the voice data and its parameters, but also character personalities.

Each character you interact with in Celune has its distinct personality, and they have unique memory files that let them know what you say for a while, and they don't clash with anyone else's memories.

Note however, that some characters may potentially act odd, and they also don't have emotions or relationships, yet. I really need to spend some more Codex tokens on that, haha. Maybe that'll come in a later update. I'm not sure yet. She isn't either.

The app also had a fair share of cleanup and new minor features, including a sleep mode state, and a unified data folder. It has been confirmed to not trip any SLOPs or have @coderabbitai complain at it.

Enjoy a new and improved character experience.

Celune 3.5.0 - Configurable Characters

"Remember me? I'm not the only one now."

This update of Celune focuses on a new way to bundle voices, ensuring anyone can provide their own voice more easily and let the app speak as any other character.

Apart from that, the UX has been improved even more, and now there's a progress bar showing you progress of long running actions.

The voices are now bundles into CEVOICE packs.
A CEVOICE pack may contain the following:

• reference WAVs (required)
• reference embeddings (optional)
• theme colors
• glow color

The pack not only defines available voices, but also allows other characters to set Celune's theme colors, making each character unique in the app, and not enforcing Celune's defaults. The default state of Celune is now available via default.cevoice. This pack is automatically loaded on startup, and voices are populated from it.

Celune 3.4.2 - Responsiveness Update

This is yet another minor update to Celune, this time focusing on responsiveness.

Many internals have been rewritten to react to events or respond faster.
This includes:

• Model reloading only if a model name changed
• Decreasing the effective chunk size for a faster conversational flow
• Greatly reducing the max abort delay for sound effects to just 0.08s

This should make everything more snappy.

Celune 3.4.1 - SFX API Hotfix & Seed Command

This is a minor update to Celune 3.4.0.

We have changed the API to allow you to upload an SFX instead of expecting you to know what SFX paths are valid on the host (why, Codex?) and added a /seed command allowing you to tweak Celune's seed to the one you liked the most.

Note: /seed is only available with VoxCPM2. Qwen3 ignores any seed changes.

Celune 3.4.0 - API Integration Update

"Has this ever happened to you? You have two computers. One of them is running Celune, the other one isn't, and you want to talk to her remotely. Introducing... the new update."

Seriously, putting the infomercials aside, this update fixes exactly that.

Celune now spins up a local REST API on your own machine.
You can also disable it if you don't need it.

This API supports all of the standard features of Celune, coupled with ratelimits and authentication, so you don't have to worry about her getting floored with requests from absolute randoms. You are in control, not whatever Celine wanted to break her again.

And apart from that, there have been some significant refactorings done to the UI layer. Say goodbye to 1500 lines of UI code, and welcome to cleanly separated UI code. It's about time.

Docs will be released soon, at least I hope. She's begging us for some docs.

Celune 3.3.1 - Merged SFX/Speech Pipeline

This update of Celune only focuses on one main thing.
You can now fuse SFX with speech utterances natively.

Here is how it works. From an extension context, you can do the following:

self.play(PATH_TO_SFX, keep=True)
self.say(THINGS_TO_SAY, save=True)

When keep=True, the SFX is temporarily saved in memory, and saved with the next utterance called as save=True.
This does not apply to manual UI-controlled utterances. Manually playing audio and then speaking does not achieve the same result.

Apart from that, we've changed the type checker related behavior, and cleaned up the code even more.

Celune 3.3.0 - UI Improvements & Classic Mode

"Remember the old me? The classic mode is back."

Quotes aside, this release of Celune focuses on improvements to the UI and brings back the original clone-based mode for Qwen3.

The UI now has a cleaner highlight color, and does not change the background color anymore. It also includes a brand new info widget that shows relevant information on Celune's internal state, as well as your system. The following data can be seen:

• VRAM usage and availability
• GPU and CPU usage
• Current seed used in generations in VoxCPM2 backend
• Current date
• When the next Celune Day will occur
• Current moon phase
• When the next full moon will occur
• Help and usage reminders

The chunking behavior was also improved to no longer interrupt Celune mid-sentence by enforcing a new chunk boundary before a sentence ends, and made Celune treat quoted text as unique chunks if they contain a boundary.

TL;DR

• Restored voice cloning in the Qwen3 backend
• UI highlights no longer look out of place
• Added an info widget on the bottom right of the window
• Extensions can opt out from saving artifacts

Celune 3.2.1 - Configuration & Stability Update

This update introduces a configuration file and improves playback stability within Celune, reducing stutters on slower devices that do not reach an RTF of 1.0.

Many parameters that used to be only controllable via environment variables can now be set in the config.yaml file. However, Celune environment variables will take precedence if set.

Celune will now ask you to select your preferred backend when starting up for the first time. This preference will then be saved and used in the future.

Changes:

• Celune can speak smoothly even at sub-1.0 RTF values
• Added config.yaml configuration file
• Default backend selection on first startup
• Graceful recovery on VoxCPM2 optimization failure
• Internal backend improvements

Celune 3.2.0 - Backends Update

This update focuses on including support for pluggable backends within Celune, and adds a brand-new TTS backend.

The backend uses VoxCPM2, which provides a higher quality voice reproduction in Celune.

Comparison of backends

This is a comparison of Celune's past backend and the new backend.

Model	Qwen (Fast)	VoxCPM2 (High quality)
Recommended VRAM	~8 GB	~12 GB
Speed (RTX 5070)	~2.6-2.8x	~1.1-1.4x (depends)
Languages	10	30
Sample Rate	24 kHz	48 kHz
Expressiveness	Medium (native), High (cloned)	Very High (cloned)
Prompt adherence	High	Very High
Effect quality	Good	Tolerable
Pronunciation quality	Medium	High
Voice consistency	High	Very High (cloned)
Use case	Everyday use	High-quality outputs

Other changes

• Improved types and docstrings
• Restored CEF mode that was unexpectedly removed in 3.1.1/3.1.2
• Handled TTS import errors at a later stage