STABLE AUDIO

There is a moment that tells you what a tool is really for. With Stable Audio it came when I asked for a "moody synthwave track, full arrangement, three minutes" and got back something that started promisingly — a detuned analog pad, a clean kick around 110 BPM — then drifted, lost the plot of its own chord progression, and faded out like it had somewhere better to be. Then I asked for "metallic impact, reverb tail, designed for a sci-fi door slam, 48kHz" and got a usable asset on the first try. That gap is the whole Stable Audio review in one session.

The short version

Stable Audio is a text-to-audio generator with open-weights options, clean 48kHz WAV export, and an unusual strength in sound design and loops. Our score: 7.8, best for sound design and open models. The honest one-liner: this is research-grade generation built for tinkerers and texture-hunters, not for anyone who needs a finished pop song with vocals by Friday.

If you score, foley, or build loops for games and edits, keep reading. If you want a verse-chorus-verse song with a singer, you are about to learn why this is the wrong room.

How the field came to believe a thing

Worth tracing, because the belief shapes what you expect when you open the tool.

Somewhere in the last few years, "text-to-music" quietly fused with "text-to-song" in the public imagination. You type words, you get a complete track with a hook and a singer — that is the picture most people carry now. It is a strong belief. It is also thinner at the source than it looks.

The picture comes mostly from a handful of consumer-facing demos that optimized hard for the wow moment: a full song, vocals included, from one sentence. Those demos were real, and they were good at the narrow thing they showed. But the leap from "a model can generate audio from text" to "any such model makes finished songs" was an audience inference, not a technical guarantee. Different systems made different bets. Some chased the song. Stable Audio, with its roots closer to research and open weights, bet on audio in the broader sense — texture, timbre, sound design, the raw material of production rather than the finished radio edit.

So when you bring the songwriter expectation to it, the tool underdelivers against a promise it never actually made. Bring the sound-designer expectation and it delivers against the thing it was genuinely built for. The disappointment is real, but it is a mismatch of belief and source, not a failure of the tool to be what it is.

Where it shines

Clean WAV you can actually drop into a session

It exports 48kHz WAV. That sounds small until you have spent an afternoon resampling lossy MP3 exports from a tool that thinks 44.1 is plenty and "WAV export" is a premium upsell. A 48kHz WAV drops straight into a video timeline or a game audio pipeline without a conversion step, and 48k is the working sample rate for picture. This is a tool that respects where its output lands.

Sound design and loops, which is the real story

The capability table marks sound design and foley as partial, and that is fair and earned. You will not get a perfect, foley-artist-grade footstep every time. But for the connective tissue of a project — a rising tonal whoosh, a granular drone under a tense scene, a four-bar percussion loop you can chop and layer — it is genuinely useful. Loops are where generative audio is at its strongest right now, because a two- or four-bar idea does not have to sustain a three-minute argument with itself. Short, modular, textural: that plays to the model's strengths instead of exposing its weakness.

Open weights and a research community

This is the quiet differentiator. Open-weights options mean you are not entirely renting access to a black box that can change terms or vanish. For tinkerers, that is the difference between a tool and a dependency you can actually inspect, run, and build around. There is API access for pipeline work, and a research community around the open models that keeps the surface area larger than the polished web app suggests. If you like to take the lid off, there is a lid to take off.

Where it falls short

One honest concession up front: the full-song output is the weak link. That synthwave render that lost its own thread is not a one-off — full arrangements are the hardest thing to ask of this class of model, and Stable Audio is not the one that cracked it. Treat song-length coherence as a maybe, not a feature.

The structural gaps matter more than the occasional mushy render:

• No stem separation. What it generates, it generates as a mixed file. If your workflow depends on pulling drums, bass, and melody apart to remix or rebalance, that capability is not here. You are mixing what you get.
• No live voice steering. There is no real-time vocal-driven control. The table marks vocal work as instrumental focus — this is not a vocal tool, and you should not buy it expecting one.
• Plan-dependent ownership. Whether you own your outputs depends on your plan. This is the licensing trap that bites people after they have already cut the asset into a client deliverable, so read the terms for your specific tier before you ship commercial work. Do not assume; verify.

On price: there is a free tier, and paid plans start around $12/month as of writing. Run the twelve-month math before you commit — roughly $144 a year if the entry plan holds. That is not expensive for a tool you use weekly, but it is real money for one you open twice and forget, and the ownership-by-plan wrinkle means the cheapest tier may not be the one that actually lets you use the output the way you need.

Who should use it, who should skip it

Use it if you are a sound designer, a game audio person who needs adaptive loops and texture beds, or a video editor hunting for risers, drones, and impacts that do not sound like the same three stock library cuts everyone else pulls from. Use it if open weights and API access matter to how you work — if "I can run and inspect this" is a feature you will actually exercise. The 48kHz WAV and the loop strength are built for exactly your job.

Skip it if your need is a finished song with a singer, or if your remix workflow lives and dies on stem separation. The instrumental focus and the missing stems are not oversights you can prompt your way around; they are the shape of the tool. Prompt-roulette will not conjure a verse-chorus structure the model was not built to hold.

Before you pay either way, take the free tier and run it on a real task — the actual loop your build needs, the actual impact your edit is missing — not a toy prompt. The free render against your own deadline tells you more than any score, including this one.

The rule of thumb for tonight: if your task is a sound, audition it; if your task is a song, look elsewhere.