Adobe Firefly Audio Goes Public—but Every Generation Still Uses Credits

Adobe made Generate Music, Generate Speech and Generate Sound Effects generally available in its browser-based Firefly studio on August 20, 2026. In its August 20 launch announcement, Adobe described the release as commercially safe AI audio for creating music matched to video, scripted voiceovers and effects synchronized with onscreen action.

The tools have left beta on the Firefly website, but generation remains metered. Music is charged by duration, speech by script length and sound effects by each generation, making rejected alternatives and corrective reruns part of the production cost rather than free experimentation.

One Firefly workspace now covers three audio jobs

Generate Music accepts a written brief or uploaded video and produces original background tracks around a requested mood, purpose, energy, tempo and duration. In the current web workflow, creators can compare alternative tracks and download either the music by itself or a version combined with the uploaded video.

Generate Speech converts a written script into downloadable voice audio. The model selector includes the Firefly Speech Model and ElevenLabs Multilingual v2, with controls for the voice, pace, emotional delivery, pauses, tone and pronunciation. Model choice matters because credit rates and Adobe’s contractual protection are not identical for Firefly and partner models.

Generate Sound Effects creates individual effects from written prompts or recorded vocal imitations. When video is present, generated clips can be placed against specific actions, moved along the timeline, trimmed and adjusted for volume, giving this part of Firefly a more editor-like workflow than a simple prompt-and-download tool.

Commercial-use protection depends on the model, plan and surface

The broad commercial-use message is clearest for Firefly’s own models, but it should not be treated as unconditional coverage for every account or model choice. “Commercially safe” describes the intended use of Firefly outputs; IP indemnification is a narrower contractual benefit tied to eligible features, qualifying plans, approved surfaces and specified export events.

Adobe’s Firefly product description lists Generate Music, Generate Speech, Text to Sound Effects and Voice to Sound Effects as eligible features, while excluding non-Adobe-trained models and features or surfaces marked beta or trial; it also states that Content Credentials are applied when Firefly-generated assets, or projects containing them, are downloaded or exported.

• Generate Music: Tracks made with the Firefly Music Model are offered as original and licensed for commercial projects, with music-only and music-plus-video export options in the tested web workflow.
• Generate Speech: Firefly Speech and ElevenLabs Multilingual v2 are separate model choices. Output from the ElevenLabs option does not automatically receive the protection reserved for Adobe-trained models.
• Generate Sound Effects: Both text- and voice-driven generation appear among the eligible Firefly features, subject to the customer’s agreement, plan, production surface and export conditions.

Content Credentials provide provenance information about the use of generative AI; they do not expand a licence or turn an ineligible partner model, beta surface or account into an indemnified workflow. A rights check therefore has to identify both the selected model and the application from which the finished asset is exported.

Each correction can consume another block of credits

Adobe’s current generative-credit table lists Generate Sound Effects at 10 credits per generation, Firefly Speech at 10 credits per 1,000 characters, ElevenLabs Multilingual v2 at 15 credits per 1,000 characters and the music function—still named Generate Soundtrack in the table—at 20 credits per minute.

Those units create different revision costs. Replacing an unsatisfactory effect consumes another fixed generation charge, while revising a long narration is costlier than correcting a short line and extending a music track increases the duration-based charge. An output that is auditioned and discarded still used the credits required to generate it.

Monthly credit allowances vary by subscription and unused balances do not roll over. Some subscriptions include unlimited standard generations, but partner models and compute-intensive audio features are treated as premium generation categories. The practical budget is therefore determined not only by the final asset’s length, but also by how many alternatives, prompt changes and delivery corrections the project requires.

Independent testing found useful music and awkward speech controls

A launch-day TechRadar hands-on test found that the first four Generate Music results were usable but generic, with more specific genre, purpose and energy terms improving the next set; the same evaluation required manual pause adjustments and a pronunciation correction in Generate Speech, while an initial sound-effect request returned four harsh variations.

Music was the strongest part of that evaluation because even the basic results could function as background audio for routine video work. Its weakness was sameness rather than outright failure, so the workflow still depended on refining the brief and comparing several generated tracks.

Speech exposed a more direct interface constraint. Previewing the full passage required selecting the text and opening a context menu, after which pause timing and pronunciation could be corrected. Those controls make repair possible, but each substantive rewrite or regeneration can add another credit-consuming cycle before a voiceover is ready to deliver.

Sound Effects provided useful placement and timing controls, yet prompt interpretation remained inconsistent. The test’s poor first set and more usable response to a simpler replacement prompt show why generated effects still need auditioning rather than being treated as finished assets on arrival.

Premiere integration remains a separate beta workflow

Generate Music also arrived in Premiere Beta 27.0 build 22 on August 20. The Premiere beta release details document text and video-based prompting, seamless loops, adjustable tempo, durations from five seconds to five minutes and four alternatives per generation; every run draws from the user’s generative-credit balance, generated tracks carry Content Credentials, and no stable-release date is specified.

That surface distinction is material to rights planning. The three audio generators are generally available on the Firefly website, while Generate Music inside Premiere is still explicitly beta and therefore does not meet the product description’s eligibility rules for beta features or beta surfaces.

For now, Firefly offers a consolidated route to music, narration and timed effects with commercial-use assurances around its own models. What remains unresolved is how quickly Adobe will improve speech preview and correction, when Generate Music will reach stable Premiere, and when the credit table will adopt the released product’s current name.

Also read:

• Avast Premium Security Covers 10 Devices, but Mobile Setup Now Uses Avast One
• AI’s Black Box Is Opening—but Researchers Still See Only Fragments
• Google Goes Beyond LLMs, but SensorFM and TabFM Remain Research Tools