I’ve been using AI tools in my audio workflow for about a year now, and the results aren’t what I expected. I thought I’d be replacing myself. Instead, I’m mostly just… deleting weird artifacts from generated samples and spending more time explaining what I want than doing it myself.
That’s not to say it’s useless. But the gap between the demos you see online and actual production work is significant.
What Actually Works
Source separation is the clear winner. I had a client send me a music track with the vocals baked in—no stems, no instrumental, just the final mix. They needed the voice removed for a trailer edit. A few years ago, that would’ve been a polite “no.” Now I run it through a separation model, get decent-ish instrumental and vocal tracks, and do cleanup on the artifacts. It’s not perfect, but it’s workable. The artifacts are predictable—swirly high-end stuff, some phase weirdness around the vocals—and I know how to massage them now.
Dialog cleanup is another one. I’ve been using AI denoisers on location audio that would have been borderline unusable. Not the fancy stuff from the big post houses—the open-source models running locally. They’re not magic, but they turn “we need to ADR this” into “I can fix this in an hour.”
The key is knowing when to stop. These tools have a confidence threshold where they start hallucinating. You learn to recognize the sound of an AI trying to reconstruct audio it doesn’t understand—it gets this glassy, over-smooth quality. When I hear that, I back off the settings and do the rest manually.
What’s Still Awkward
Generative audio is the obvious disappointment if you were hoping for “prompt to final asset.” I’ve tried every tool that promises to generate sound effects from text descriptions. The results are either impressively wrong or blandly generic. “Metal impact” gets you something that sounds like someone hitting a frying pan with a spoon, recorded through a phone. Every time.
The image generation people dealt with this already—you can’t just type “epic dragon” and get production art. You iterate, you guide, you composite. Same with audio, except the iteration cycle is slower because you have to listen to each attempt in real-time.
Where it gets interesting is variation generation. I’ve had success taking a sound I recorded—say, a mechanical switch click—and using AI to generate variants. Same character, different timbre. It’s not replacing my recording session, but it extends what I captured. That’s a genuinely useful workflow, though it took me months to figure out the prompting that gets consistent results.
The Workflow Shift
The bigger change isn’t in the tools themselves—it’s in how I think about pre-production. I used to do a lot of exploratory recording, capturing way more than I needed and sorting it later. Now I’m more likely to capture a representative sample and generate the variations I need. It’s faster, but I worry about the long-term effect on my ear. There’s a difference between selecting from recordings you made and curating generations from a model.
I’m also spending more time on metadata. These tools need precise tagging to work well. My sound library was always reasonably organized, but now I’m maintaining parallel taxonomies—one for human browsing, one for machine parsing. That’s overhead I didn’t anticipate.
The Integration Problem
Most AI audio tools exist outside the DAW. You export, process, import. Some have VST wrappers now, but they’re buggy—latency issues, crashes when you change projects, UI that clearly wasn’t designed by people who mix for a living. The stable ones tend to be the simpler tools: one-knob denoisers, basic stem separation. Anything ambitious requires leaving your session.
I’ve ended up building a lot of small utilities—scripts that batch process through AI tools and re-import with proper naming. It’s not glamorous work, but it saves hours on large projects. I keep thinking someone will release a proper integrated suite, but the economics favor standalone apps with subscription pricing.
What I’d Tell Myself Six Months Ago
Don’t buy the “AI will 10x your productivity” narrative. It’s more like a 1.3x on specific tasks, with significant upfront investment in learning the tools and building workflows around them. The real value is in expanding what’s possible—solving problems that were previously intractable—rather than speeding up work you could already do.
Also, keep your original recordings. The temptation to let AI “fix it in post” and delete the source is strong, especially when storage is tight. But these models improve, and what sounds like a reasonable compromise today might be fixable properly in a year. I’m glad I kept those noisy location tracks instead of just the cleaned versions.
If you’re curious about getting started, I’d suggest source separation tools first. They’re the most mature, the most immediately useful, and they fail in predictable ways. Everything else is still figuring out what it wants to be.
What about you? Are you using AI tools in your audio work, or holding off until they settle? I’ve found the community knowledge-sharing around specific workflows more valuable than any official documentation—there’s a lot of undocumented behavior that only comes up in practice.