Fish Audio Drama 3 ships editable speech, with license limits

Fish Audio's Drama 3 makes editable speech generation available through downloadable weights and local serving tools. The 35B mixture-of-experts model supports multi-speaker scenes and revisions of earlier takes. Its community license also imposes commercial and hosting restrictions that a self-hosted deployment does not remove.

Key facts

• Fish Audio describes Drama 3 as a 35B model with 3B active parameters, using a hybrid-attention text backbone and an audio decoder.
• The model card documents speaker tags for up to 16 characters and chunked streaming for long text.
• The community license permits covered commercial use only below $500,000 in aggregate annual revenue, subject to its other conditions.
• Fish Audio reports 88.9 milliseconds median time to first audio at single-request concurrency on an H200; this is a published project measurement, not a DeAI test.

What changed in the speech workflow

DeAI's October 15 RSS collection caught the Drama 3 release. The current project documentation identifies multi-turn director revision and local infilling as additions over S2-Pro. Instead of discarding a take and starting again, an application can provide earlier output as context and request a different delivery or a textual change.

The model card demonstrates this with an edited deadline: the revised text changes the day, then a further revision changes the time. Earlier takes remain in the request history. These are supplied examples of the interface, not evidence that arbitrary edits will preserve every surrounding sound without artifacts.

The deployment surface is also tangible. Fish Audio links launch support in SGLang-Omni and vLLM-Omni, with commands for serving the checkpoint. An operator can inspect the generation path and retain control of the machine running inference. That creates a local alternative to sending every script and reference recording to a hosted voice service.

The commercial threshold is only one condition

The decisive document for product planning is the license text. Its revenue test counts the licensee and affiliates together, across all sources, over the preceding year. It is not limited to money earned from generated audio. Some work done for a particular third party also requires that party to meet the threshold, while the agreement defines exceptions for delivering finished output and certain generally available products.

The agreement separately restricts hosted services regardless of revenue or purpose. Its definition covers making the materials available so third parties can supply text, audio, or voice samples and generate speech through covered services. A small company's general-purpose speech API therefore cannot rely on the revenue threshold alone.

The exceptions are material. The license distinguishes generating speech for content an operator controls, a production service delivering finished audio without self-serve access, and qualifying free research or demonstration services. Whether an application fits one of those descriptions depends on its actual workflow. The blanket descriptions “commercial use allowed” and “all hosting prohibited” both lose important terms.

The same agreement requires attribution, restricts certain uses of outputs for training other models, and sets conditions around consent and provenance. This article reports those conditions; it does not establish that a particular business qualifies under them. Teams need the full agreement when choosing a deployment model.

Performance depends on the serving layout

Fish Audio's serving table provides more useful context than a single latency claim. It specifies H200 hardware, English evaluation prompts, a pre-encoded reference, sampling settings, and a generation limit. At increasing concurrency, throughput rises while the first-audio wait grows. Those are different operating points, not a promise that every customer receives the fastest single-stream result.

The card also compares placing the vocoder on the inference GPU with moving it to another GPU. The reported improvement under heavy load indicates that the audio pipeline can bottleneck outside the text backbone. Capacity planning therefore needs the full generation and playback path, including reference processing and the vocoder, rather than just active parameters.

A local workstation should be evaluated on its own terms. The published H200 results cannot establish latency on a consumer GPU, and the smaller active count does not mean only those active weights need storage. The usable configuration depends on the checkpoint, precision, available memory, and serving engine.

Background and what to verify

The distinction between downloadable weights and unrestricted permissions is central to DeAI's open-weight versus open-source guide and the trust hub. Drama 3 makes that distinction concrete: local execution is technically available while the licensing still constrains how a service can be offered.

Privacy is another separate question. As the private-inference explainer describes, control of the model host does not prove that recordings avoid every external system. Reference uploads, object storage, application logs, and support tooling need their own data-flow review.

For builders evaluating this release, the next useful evidence is an end-to-end run on the intended hardware: representative scripts, permitted reference voices, edits across successive takes, and the concurrency the product expects. Compare audible continuity and failure recovery alongside first-audio latency. Preserve the exact license version with the checkpoint. The release offers a new editing workflow, but operating it successfully requires both a compatible serving stack and permission for the intended use.