Fish Audio Drama 3: direct a voice and budget each take

The English launch announcement quotes USD $10 per million characters from October 13 to November 12, 2026, and $50 at regular price. It does not state a cutoff timezone. Near that boundary, budget at the regular rate and check the price shown for your account.

For 100,000 billable characters, one submission costs $1 during the offer or $5 at regular price. Three full submissions of that workload cost $3 or $15. Re-rendering changes the submitted workload; a short final recording does not erase the cost of discarded takes.

The creator-plan discount is a separate offer on eligible new subscriptions. App credits are not a character-to-minute conversion for API billing. Use the calculator’s API comparison to compare metered speech fees, then add any subscription or separate licensing costs that apply to your workflow.

Keep the POST /v1/tts endpoint and send the model header drama-3. Fish’s guide warns that an absent or misspelled header falls back to s2.1-pro rather than failing. Check that header when comparing output or billing.

Write performance directions inside the text and use a voice you are authorized to use. Start with a short line containing the names, numbers and emotional transitions that matter in your script. This request produces a speech file; it requires your own API access and incurs the provider’s charges when run.

Bracketed directions can set a mood, add a breath or change pace at the point where they appear. Paired tags such as <whisper> and <emphasis> apply to a span of words. Phoneme tags give another way to pin down a difficult pronunciation.

The docs describe experimental language switching within a line. Try the specific voice and language pair you need; a multilingual catalog does not establish that every pronunciation or switch will work equally well.

The model card supports speaker tags numbered 0 through 15 and chunked long-text output. For a scene, keep the mapping from speaker tags to reference voices consistent. Interactive editing and audio inpainting shown in the launch post remain in invite-only testing, so the ordinary workflow is to revise the direction and generate another take.

The official card describes a 35B MoE text backbone with about 3B active parameters per token, plus an audio decoder. Those active parameters are not the full weight memory requirement. The repository contains FP8 expert weights alongside BF16 tensors; plan for the actual files and runtime allocations.

Fish’s card points to SGLang-Omni for serving. Its installation example enables the fishaudio extra, then serves fishaudio/drama-3. Use the matching cookbook for reference-audio and streaming requests. Download size alone is not a tested minimum VRAM figure.

The current vLLM recipe adds Drama 3 through /v1/audio/speech, using vLLM 0.31.0 or later and a vLLM-Omni build with Drama 3 support. It requires a CUDA GPU with at least 48GB; the reference run uses one H100 80GB. Its 36.9GiB stage-0 weight allocation is only one part of the two-stage service.

Install the recipe’s matching runtime, use its Drama 3 launch configuration and wait for both the autoregressive and codec stages to load. The speech API guide covers the request body, speaker tags and reference-audio inputs. Send a short authorized line first and inspect the saved audio before extending the script.

The actual LICENSE file is Fish Audio Community License v1.0, dated October 13. It permits research and non-commercial use and grants limited commercial rights while combined annual revenue is below $500,000. It defines revenue across you and your affiliates; installing the weights for a third party also has conditions for that party.

Offering a self-service speech-generation, voice-cloning or inference service to others requires a separate written license, regardless of revenue. The license distinguishes that from delivering finished recordings, your own controlled content and certain free non-commercial demos. Read those definitions for your product. Cloning a real person’s voice also requires their consent.

Fish’s announcement cites an Artificial Analysis pronunciation result of 93.2% for Drama 3, versus 88.1% for S2.1 Pro. That is a pronunciation benchmark attributed in the release, not a measure of emotional acting, cloning similarity or performance in all 83 advertised languages.

Compare names, dates, codes, pauses and speaker consistency on the same script. For clone creation and access conditions, the HeyGen Voice guide covers a different workflow; YuE2 instead generates songs and belongs in music generation. Neither a music tariff nor an unspecified minutes-to-credits price belongs in a character-rate comparison.