Stable Audio

Stable Audio is Stability AI's generative audio product. It creates music tracks and sound effects from text prompts, producing stereo audio that can be used in video, games, podcasts, advertising and product interfaces. It sits alongside the company's image and text models and shares their commercial structure: a hosted web tool for creators, and an API for developers who want generation inside their own products. What distinguishes Stable Audio from much of the generative music field is not only the model quality but the provenance of its training data, which is the single most important commercial question in this category.

The training data position deserves to come first, because it determines whether the output is safe to monetise. A large part of the audio generative space was built on datasets scraped from wherever recordings could be found, which produced a legal exposure that recording artists and labels have begun to test in court. Stable Audio was trained substantially on audio licensed from production music libraries, notably AudioSparx, whose catalogue consists of recordings created specifically for commercial licensing. That does not make every legal question disappear, and buyers with serious exposure should still read the vendor's terms and indemnification language, but it is a materially different starting position from a model trained on unlicensed material. For an advertising agency or a game studio, this is often the deciding factor.

Functionally, there are three main modes. Text to music generates an instrumental track from a written description. The prompt is where control lives: you specify genre, instrumentation, tempo in beats per minute, mood, energy and structure, and the model produces a stereo track. Track length is a tier-dependent parameter, with paid plans allowing substantially longer output than the free tier, historically up to around three minutes for a single generation. Text to audio effects generates short sounds, which is a genuinely useful capability for game developers, interface designers and video editors who need a specific texture that does not exist in their sample library and cannot justify commissioning one. Audio to audio takes an uploaded recording and transforms it according to a prompt, which enables restyling, changing instrumentation or extending a piece while preserving some of the original character.

The web tool provides generation history, download in standard formats, and a prompt interface that lets you iterate. There is no piano roll, no stem separation, no mixing console and no arrangement timeline, and that absence is deliberate. Stable Audio is a generation tool, not a digital audio workstation. The expected workflow is that you generate candidates, pick one, and take it into a real editor or a video timeline for trimming, layering, ducking under dialogue and mastering. Anyone hoping to compose a structured song with verses, a chorus and specific chord changes will be disappointed; the model is better at producing a coherent bed of music with a given feel than at obeying detailed musical architecture.

The API exposes the same generation capabilities programmatically, which is where the commercial value concentrates. Applications include adaptive audio for games that needs a specific mood at a specific length, automated music beds for video at scale, sound design tooling, and product interfaces where a user describes what they want. API usage is billed separately from consumer subscriptions, typically per second or per generation, and the economics need modelling against the alternative of licensing production music per track.

Against competitors, Stable Audio faces a strong field. Suno is the most popular consumer option for full songs with vocals, and its ability to generate a complete track with lyrics and singing is something Stable Audio does not attempt; Suno's training data provenance has been the subject of major litigation from record labels. Udio is the other vocal-capable rival, similar in ambition and similarly litigated. AIVA focuses on symbolic composition and gives far more musical control, exporting MIDI so a composer can actually edit the notes, which is a fundamentally different and in some ways more professional proposition. Soundraw and Mubert generate royalty-free background music with a licensing model built for creators and are less about prompt fidelity than about safe, usable beds. Beatoven and Epidemic Sound's tools serve video editors directly. ElevenLabs dominates the speech side and has expanded into effects. For sound effects specifically, ElevenLabs' effects generation and dedicated game audio libraries are the comparison. Stable Audio's position is: serious prompt-based instrumental music and effects, with a cleaner data provenance story than the vocal-first consumer products, and an API for developers.

Pricing follows a freemium model. The free plan provides a monthly allowance of credits for generation, with commercial use restricted or excluded, which makes it an evaluation tier rather than a working one. Paid plans are monthly subscriptions that increase the credit allowance, extend maximum track duration, improve generation speed and priority, and, critically, grant commercial usage rights on the generated audio. Higher tiers exist for professionals and studios with larger volumes. API usage is metered separately, generally by generation or by seconds of audio produced. The licensing terms attached to each tier are the part that matters most: what you are allowed to do with output generated on a free plan differs from output generated on a paid plan, and anyone monetising audio should read those terms rather than assume them. Current figures are published on the vendor pricing page.

The limitations should be stated plainly. There is no vocal or lyric generation, which rules out songwriting use cases entirely. Musical structure is not controllable at the level a composer needs; you cannot ask for a key change at bar thirty-two and expect compliance. Output is a rendered audio file rather than an editable score or MIDI, so corrections mean regenerating rather than fixing. Consistency across generations is limited, which makes it hard to produce a coherent set of tracks for one project without a lot of iteration. Duration caps on lower tiers constrain long-form work. The model can produce generic results when prompts are vague, and prompting well for music takes practice and vocabulary. Legal certainty about generative audio is still evolving, and the licensing position depends on the vendor's agreements with its data providers as well as on the law in your jurisdiction. And because Stability AI as a company has gone through periods of financial and leadership turbulence, buyers making a long-term commitment should consider continuity risk.

Stable Audio fits video creators and editors who need instrumental beds and effects without a licensing search, game developers building adaptive or placeholder audio, product teams generating interface sounds, advertising and marketing teams that need music quickly with a defensible provenance story, and developers integrating audio generation through an API. It does not fit songwriters wanting vocals, composers who need editable musical material, anyone requiring exact structural control, or organisations that cannot tolerate any residual legal ambiguity about generated content.

In the closing comparison, choose Stable Audio when you need instrumental music or sound effects from a prompt, when API access matters, and when training data provenance is part of your decision criteria. Choose Suno or Udio when you need a full song with vocals and accept the legal debate. Choose AIVA when you want editable composition and real musical control. Choose Mubert, Soundraw or a production music library when you need safe, plentiful background beds and care more about availability than about prompt fidelity. Choose ElevenLabs when the requirement is speech. For a video or game team that needs specific instrumental textures on demand with an API and a cleaner data story, Stable Audio is one of the more defensible choices in a category where defensibility is not common.