These articles come from lessons learned while building Eterna Clarity and the operating system I use to run it.
This week I used a Qwen text-to-speech model running locally to make narration. I have been paying ElevenLabs for voice production, so I had a pretty good reference for what I wanted to hear. The local result genuinely surprised me. To my ears, the narration was every bit as good as work I had been getting from the paid service.
That is my experience with the material I tested, not a controlled comparison or a claim that every local voice model beats every paid service. But it was exciting. I had a piece of software running on my own machine producing something I would be happy to use, with far more room to control how it fit the rest of my work.
The discovery didn't stop at narration. Eterna can now generate music and sound effects locally, transcribe speech, handle edits and carry out much of the production work on the same computer. I've been impressed by the quality across those early uses. Some results are comparable to, or better than, paid tools I've used, especially when I factor in how much control I have over the process.
That has made me look at local AI a little differently. The first result proved to me that the quality was there. Now I'm interested in what happens when I can use those capabilities repeatedly and keep improving the way the work gets done.
The impressive output is only one piece
The wider interest in running AI on personal computers makes sense to me. In October, Hugging Face introduced Llama Windows, making it possible to download and run an open model from a Windows tray app. Microsoft announced experimental local-model support in Windows ML, and NVIDIA launched new PCs designed with local AI among their main uses. This is becoming easier for people to try without assembling a specialist development environment first.
The appeal isn't complicated. If a model on your own computer can produce a natural voice or a usable piece of music, there's a real creative option sitting in front of you. You can test an idea without moving every part of the job into a separate paid service.
I'm finding the different specialities just as interesting as the big general-purpose models. One local tool can speak a script. Another can produce a soundtrack or a specific effect. Another can turn speech back into words. They don't all need to be brilliant at everything. Each needs to do its own job well enough that I can judge the result and use it.
Then come the ordinary parts of making something. The voice has to fit the picture. The music can't drown out the words. Captions need to match what was actually said. A video has to make sense on a phone as well as a large screen. Getting a strong model output solves one piece of that job, and I want to keep the good pieces instead of starting the entire production over every time one part changes.
The tools around the models do a lot of the work
Eterna already had a growing collection of video-production tools before this recent local audio work. During an October Workspace demo project, the editing utilities started becoming a reusable Eterna Motion Toolkit instead of scripts that only knew how to finish one video.
Some tools inspect recorded speech for noise and help clean it up. Others work with transcripts, cut timings and captions, including the names that speech recognition can easily get wrong. There are tools for lowering music when somebody speaks, checking how the mix translates to a phone speaker, picking usable thumbnail frames, and making proper landscape and vertical versions. Eterna can even track a hand in a video so a graphic can be positioned in relation to the person's movement.
None of that sounds quite as magical as asking a model to make a song. It's also the work that determines whether the song becomes useful in a real production.
Now Eterna's local generation capabilities are working alongside those established production tools. I can make a narration track, produce music and effects, work with the words and timing, and edit the pieces in a more connected process. I still need to decide what the audience should hear and see. I still need to reject a sound that doesn't belong, change a line that reads awkwardly, or decide that a scene needs less going on. But I can make those decisions while keeping far more of the production under my control.
Consider a small change to a line of narration. In a disconnected setup, that change can send you through a voice service, an editor, a caption tool and another export. With a reusable Eterna workflow, the steps are clearer: regenerate the line, check where it sits, update the captions against the actual audio, adjust the music mix and inspect the new video. I can keep the other pieces that were already working. That's the kind of control I want more of, without pretending every revision is automatic or that no judgment is needed.
The same distinction applies to the final export. Software can check that a video uses the intended format, that the sound levels are within the chosen range, and that a caption file exists. It cannot decide whether a shot is boring, whether a musical change feels right, or whether the whole thing is something I want to put my name on. I've seen work pass technical checks and still fail that last test. Those are different kinds of quality, and I want the system to respect both.
The next production should inherit something
What I keep coming back to is how much of the work can survive the project that taught it to me.
A production reveals that certain names need to be handled carefully in transcription. That can become a shared vocabulary rather than a correction made from memory every time. A music mix reveals how much room a voice needs. The next edit can start from a tested mixing procedure instead of finding those levels again. A vertical version exposes a framing problem. The next project can use an established review step before the mistake reaches a finished file.
Some of that learning becomes code. Some becomes a checklist, a preset, a production recipe or a better way to explain the job to whichever AI is helping. When a technique is genuinely useful, Eterna can keep it where the next project will find it.
There is an important distinction here. I don't mean that every music model or voice model is secretly training itself as I work. The models can stay exactly as they are while the production system gets better at using them. A correction made once can become a tool or procedure that applies again. The next project starts with more of the hard-won knowledge already in place.
This isn't instant. A reusable tool takes work to build and maintain, and a complicated workflow can be worse than a simple one if you're only going to use it once. I don't want to turn every creative choice into a rigid rule, either. A repeatable process should preserve room to make a different choice when the next piece needs it.
The benefit is that I can choose what deserves to be repeated. A good default can handle the predictable steps while I spend my attention on the story, the look, the sound and the things that actually change from project to project.
It reaches beyond video
The same approach shows up in Eterna's non-creative work. A useful file, an accepted decision or the current state of a project has to be available again when the next AI conversation begins. Otherwise the next session spends its time asking me to explain what was already settled, rather than helping with the next job.
One October demonstration connected five different consumer AI applications to the same live Eterna environment. Each began in a fresh conversation and could retrieve the same current work. That wasn't a test of a small local audio model, and I don't want to mix up those capabilities. It showed why the surrounding system matters when different forms of intelligence are doing different jobs.
The creative tools are making that idea much easier to see. It's one thing to say that an AI can access a file, or that software can preserve a workflow. It's another to hear narration generated on your own computer, add locally made music and effects, watch the edit come together, and know that the useful steps can be carried into the next production.
I still use strong hosted AI services, and I expect to keep using them where they give me something valuable. Local production isn't free of hardware costs, electricity, maintenance, or the time it takes to make good decisions. Output quality also varies. A voice that works beautifully for one script may need more direction on another, and creative work still needs actual listening and review. I also have to respect the rights and permissions attached to source material and voices, regardless of where the model runs.
What has changed is the range of choices available to me. A paid service no longer has to be the default for every piece of narration or sound. A local model doesn't have to be forced into a job that ordinary software handles better. And the work doesn't have to be spread across separate places that forget what the last project taught me.
If you want to try the same idea without building a system like Eterna, start with one real project. Make something you actually need, even a short narrated video. Keep the narration, music and effects separate so you can change one without destroying the others. Save the corrections and the steps that worked, and see which ones are worth making repeatable. A useful toolkit can begin with a few well-understood steps.
When I first heard that Qwen narration, the quality was the surprise. Now I'm more interested in the production I can build around it, and in how much of that production will already be ready when I start the next one.
Top comments (0)