Stable Audio

Stable Audio production guide

Stable Audio is Stability AI's generative-audio family for music and sound-design outputs, not speech, voice cloning, dialogue, or final mix/mastering. Use it when a project needs instrumental music, loops, stems/ideas, sound effects, foley-like assets, ambience, stingers, transitions, or style/continuation edits from rights-cleared audio. Route speech, dubbing, voiceover, transcript alignment, and loudness mastering to the appropriate speech/audio-post tools instead.

All volatile facts in this guide were verified on 2026-07-10 from Stability AI official pages, the Stability API OpenAPI specification, official model cards/repositories, and Stability policy pages.

Capability map

Documented facts:

• Hosted Stability API exposes two Stable Audio route families:
  • Stable Audio 2 endpoints under /v2beta/audio/stable-audio-2/..., with model choices stable-audio-2 and stable-audio-2.5.
  • Stable Audio 3 endpoints under /v2beta/audio/stable-audio/..., with model=stable-audio-3.
• Supported hosted workflows are text-to-audio, audio-to-audio, and inpainting/continuation.
• Hosted outputs are 44.1 kHz stereo, with mp3 or wav output format.
• Stable Audio 2/2.5 hosted endpoints return audio synchronously on success. Stable Audio 3 hosted endpoints return HTTP 202 with a generation id; poll GET /v2beta/audio/results/{id} until HTTP 200 returns the audio.
• Stable Audio 2/2.5 hosted duration limits are 1-190 seconds. Audio uploads for audio-to-audio and inpaint must be 6-190 seconds, mp3 or wav, and requests are capped at 50 MB.
• Stable Audio 3 hosted duration limits are 1-380 seconds. Audio uploads for audio-to-audio and inpaint must be 6-380 seconds, mp3 or wav, and requests are capped at 100 MB.
• Prompt length is capped at 10,000 characters.
• seed accepts 0 or omitted for random generation; explicit seeds may be recorded for reproducibility.
• steps differs by model: Stable Audio 2 accepts 30-100 steps and defaults to 50; Stable Audio 2.5 and Stable Audio 3 accept 4-8 steps and default to 8.
• cfg_scale ranges 1-25. API docs describe it as prompt-adherence strength; defaults are 7 for Stable Audio 2, 1 for Stable Audio 2.5, and 1 for Stable Audio 3. Treat it as a secondary control after prompt, duration, model, steps, and source-audio strategy.
• strength in audio-to-audio ranges 0-1 and controls source-audio influence. Stability describes 0 as identical to the input and 1 as equivalent to no audio influence.
• Inpaint uses mask_start and mask_end in seconds to choose the replacement/continuation segment.
• Hosted pricing, verified 2026-07-10: Stable Audio 2.0 is credits = 17 + 0.06 * steps per successful generation, usually 20 credits at the default 50 steps and 23 credits at 100 steps; Stable Audio 2.5 is 20 credits per successful result; Stable Audio 3.0 is 26 credits per successful result. Stability states failed generations are not charged. Stability pricing uses 1 credit = $0.01 and is subject to change.
• Stable Audio Open 1.0 is an open-weight text-to-audio model that generates variable-length stereo audio up to 47 seconds at 44.1 kHz.
• Stable Audio 3 open-weight releases include small-music, small-sfx, and medium; the official repository lists Small music/SFX as CPU-capable 433M-parameter models with 120s max length and Medium as a 1.4B CUDA model with 380s max length. Stable Audio 3 Large is API-only in the official repo.

Sources: Stability API reference / OpenAPI spec, Stable Audio 3 page, Stable Audio 2.0 announcement, Stable Audio 2.5 announcement, Stable Audio 3 repository, Stable Audio Open 1.0 model card, Stable Audio Open paper, Stable Audio 3 paper, Stability pricing page.

Choose the route before prompting

Use hosted Stable Audio 3 when the project needs current highest-capability hosted Stable Audio, up to ~6 minute beds, async generation is acceptable, or inpainting/continuation is part of the plan. It is the cleanest choice for longer music beds, adaptive game/music cues, sonic-branding variations, and complex ambience where 190 seconds is not enough.

Use hosted Stable Audio 2.5 when you need synchronous API behavior, 190 seconds is enough, and the brief is commercial music/sound production where Stable Audio 2.5's documented improvements in speed, musical structure, mood/genre prompt adherence, and inpainting matter.

Use hosted Stable Audio 2 only when a pipeline already depends on its behavior or when you intentionally need its 30-100 step range and cfg_scale behavior. Do not default to it just because the endpoint default is stable-audio-2; explicitly set the model.

Use Stable Audio 3 open weights when local or self-hosted generation, fine-tuning/LoRA experimentation, data-control review, or offline iteration matters. Choose:

• small-music for quick CPU music drafts up to 120s;
• small-sfx for quick CPU sound-effect drafts up to 120s;
• medium for higher-quality local generation up to 380s when CUDA and dependency constraints are acceptable.

Use Stable Audio Open 1.0 only when the 47s limit is acceptable and the workflow needs the older open model, diffusers/stable-audio-tools compatibility, or research comparison.

Do not use Stable Audio for:

• lyrics or sung vocal performance unless the current product/tool explicitly supports the needed vocal workflow;
• voiceover, narration, voice cloning, or lip-sync;
• exact recreation of a known artist, track, label sound, or copyrighted recording;
• uploads that the user has not rights-cleared;
• final legal/music-supervision clearance.

Rights, consent, privacy, and upload gate

Before any audio-to-audio or inpaint request, require the user or project brief to establish that every uploaded sound is rights-cleared for transformation. Stability's API docs and Stable Audio announcements state that copyrighted content is not allowed to be uploaded and describe content recognition/compliance scanning for uploads. Treat "I found this song online" as blocked until replaced with licensed, public-domain, commissioned, self-recorded, or otherwise authorized material.

For commercial media, record:

• provider/model/version or repository id;
• endpoint/workflow;
• prompt;
• duration, seed, steps, cfg_scale if used, strength, mask times, and output format;
• source-audio filename, owner, license/permission, and whether it contains third-party music, vocals, identifiable voices, brands, or location ambience;
• generation timestamp, request id/generation id if available, and credit cost;
• post-processing chain and final exported path.

Stability's Terms of Service state that users are responsible for inputs and must have rights, licenses, and permissions for inputs; as between the user and Stability, Stability assigns any right it has in outputs subject to compliance and applicable law. The same terms also say outputs may be similar across users, users must verify legality/appropriateness before use, and Stability may use inputs/outputs to improve services unless the user opts out where available. Privacy/security policy claims should not be overstated: Stability states it uses organizational and technical safeguards, but no internet transmission/storage is guaranteed to be fully secure.

For confidential brand libraries, unreleased music, celebrity voices, minors, medical/legal content, or contractual IP, prefer enterprise-approved settings or local/open-weight routes only after the project's data-handling requirements are known. Do not upload sensitive stems just because the API is convenient.

Sources: Stability Terms of Service, Stability Acceptable Use Policy, Stability Privacy Policy, Stable Audio 2.0 announcement, Stable Audio 2.5 announcement, Stability SOC 2/SOC 3 announcement.

Prompting for production

Stable Audio responds best when the prompt describes what the audio should sound like, not what the video should show. Translate visual intent into sonic parameters.

Documented prompt elements from Stability guidance include genre/subgenre, style, tempo/BPM, mood, and instrument type. Production heuristics below are not documented guarantees; use them because they make review and iteration easier:

• State the deliverable role first: "15-second UI success stinger", "90-second product-launch bed", "6-second cloth-rustle foley", "loopable 120 BPM tech-house drum loop".
• Include tempo/BPM when synchronization matters; include "loopable" only when you will QA the loop seam.
• Include key only when the composition must sit under other music; verify by ear or audio analysis because key adherence is not guaranteed.
• Use concrete instrumentation and mix language: "muted palm-picked electric guitar", "dry close-mic footsteps on wet concrete", "subtle analog tape hiss", "wide stereo pad, no lead melody".
• Use section language for music beds: intro, lift, breakdown, outro, button ending, no hard drop before 20s.
• Use avoidance language sparingly inside the positive prompt when no negative prompt parameter exists in the hosted API: "no vocals, no recognizable melody, no heavy drums".
• For sound effects, describe object, material, gesture, space, distance, intensity, and tail: "ceramic mug placed gently on a wooden desk, close perspective, small room reflection, short natural tail."
• For cinematic ambience, separate bed and events: "continuous low city night ambience with occasional distant siren and soft tire hiss; no foreground speech."

Keep prompts internally consistent. A single prompt that asks for "minimal ambient corporate piano" and "aggressive drum & bass festival drop" will produce less controllable review targets.