Fish Audio Drama 3 Preview: Controllable TTS Without Tags

Introduction
Fish Audio just previewed a new control layer for expressive speech. On September 23, 2026, the company posted Drama 3 (preview) with the tagline “A voice you can direct.”
The pitch: describe performance in language you already use, then steer individual words. Fish’s demo card stacks three claims—add a little [performance], describe it in your own words, and choose exactly which words change—then shows multilingual takes that actually follow those directions.
Access is early and invite-gated: comment DRAMA on the announcement for an API key. The closing card says Drama 3 will land in Fish Audio Creative Studio + API, with early access rolling out. Public docs still list S2.1 Pro / S2.1 Pro Free as the production stack; Drama 3 is a named preview, not a documented GA model ID yet.
Pro Tip
Watch the control surface in the demo carefully. Drama 3 mixes scene-level brackets like [breathless, speaking fast, urgent] with word-level XML-style wraps such as <emphasis>…</emphasis> and <whisper>…</whisper>. That is a different directing grammar than S2.1’s open-domain [bracket] cues alone—practice both layers on a short script before you port a full dialogue tree.
What the Preview Demo Shows
The official X note (Sep 23, ~17:01 UTC / ~01:01 MYT Sep 24) introduces Drama 3 with:
- Plain-language direction for tone, pacing, and character
- Mid-sentence voice shifts
- Multi-character / multi-language scenes
- Surgical single-word (or phrase) fixes
The ~40s demo then proves it with three speakers:
CharacterLanguageWhat changes mid-line
Mila
EN
Breathless urgency → terrified whisper → slow, emphasized “do not open that door”
Haruto
JA
Surprised breath-catch → pause → soft smile on the return line
Thérèse
FR
Conspiratorial lean-in, with only entre nous dropping into a true whisper
UI cues in the video show language chips (EN / JA / FR), inline translations for non-English lines, and the dual tag styles side by side. Fish has not yet published a matching blog entry, changelog line, or docs.fish.audio model page for Drama 3—so the announcement thread plus this demo remain the primary sources.
How This Fits Beside S2.1 Pro
Fish’s current documented TTS ladder is still:
ModelRole
s2.1-pro
Recommended production TTS (83 languages, bracket natural-language control, multi-speaker)
s2.1-pro-free
Same model at $0 under Fair Use through Nov 30, 2026 (no TTFA/DPA SLA)
s2-pro
Prior open generation with [bracket] cues
s1
Earlier flagship with (parenthesis) emotion presets
Drama 3 sits above that ladder as a controllability preview aimed at Creative Studio directors and API builders. Do not swap production agents off s2.1-pro until Fish publishes a stable model string, pricing, and API examples.
Why Creators Should Care
Expressive TTS usually forces a trade: pick a voice and hope, or learn a private cue language. Drama 3 aims at writers who already think in stage notes—“breathless,” “softening into a smile,” “whisper only these two words”—and want that language to drive the synthesizer.
If the preview holds up, the practical wins are:
- Faster iteration on dialogue takes (change one phrase, keep the rest)
- Multi-language character reads without stitching three clones by hand
- Less folklore around which tags “really” work
Caveats stay honest: this is comment-to-get-key early access, metrics and full language lists for Drama 3 are not published yet, and there is no public migration guide from S2.1.
How to Try the Preview
- Open the official announcement:
- Comment
DRAMAas Fish requests for an API key - Expect Creative Studio + API once your early-access invite lands
- Keep a parallel
s2.1-pro/s2.1-pro-freepath for anything that must ship while Drama 3 stays invite-only
Bookmark Fish’s models overview and watch for a Drama 3 model string—until it appears there, assume preview semantics.
Conclusion
Drama 3 is Fish Audio’s bid to make TTS feel like directing a cast: bracketed performance notes, word-level wraps, mid-line shifts, and multilingual demos that actually obey the stage directions. The catch is access—preview badges, early-access rollout, X-gated keys, and still undocumented on the public model list.
If you write dialogue for games, podcasts, or agents, request the key, A/B it against S2.1 Pro on the same script (especially whisper/emphasis boundaries), and keep notes for when Fish promotes Drama 3 out of preview.
—Anabel ♡
