Fish Audio Drama 3
Preview — direct the performance in the script. No published price or time to first audio.
Comparative Scores
Architecture
The closest Fish fit for Où est Ava ? emotional direction. A PerformanceIntent can be written as a French brief and, when the line needs it, as a scoped tag, without Auto Tag and without a second LLM. Test French lines off-air first. Do not replace Gradium on the live path while price, time to first audio, take stability and the silent s2.1-pro fallback are unmeasured. One-pass multi-speaker is for a prepared scene, not the live single-character turn.
Analysis
Drama 3 is Fish Audio’s preview TTS for directing a performance, not only the words. You brief the line in your own words, place a sigh, a laugh or a pause where it happens, and limit whisper, emphasis, pace or force to the words you wrap. The same marks work in POST /v1/tts and in the live WebSocket when the model header is drama-3. Multi-speaker dialogue is one HTTP request: one voice id per speaker and <|speaker:N|> turns. Auto Tag can draft directions in the editor; the preview varies from take to take, so Fish tells you to generate several and pick one. The models overview still recommends s2.1-pro for production. This card does not copy an arena ELO: the 17 October audit mentioned 1244 before the Fish page was read, and this pass did not re-read the arena.
Strengths
- Free-form brief, including in a language other than the line
- Paired tags affect only the selected words
- Sounds and pauses sit where they are written
- Same HTTP API and live WebSocket, header drama-3
- One-pass multi-speaker dialogue on HTTP
Weaknesses
- Preview: takes vary, no production TTFA or DPA guarantee
- No price on the 18 October rate card
- A bad model header silently falls back to s2.1-pro
- Spoken-language count is not published; language switch is experimental
- No documented lip-sync timestamps or EU/CH residency
Voice Capabilities
La requête passe une voix par reference_id. La doc demande un clone bilingue quand la réplique change de langue. Les exemples publics sont Adrian, Delia et Sheila (EN-US).
You brief the line like an actor: a free-form direction, a sound or pause where it happens, and an effect limited to a few words.
How: `drama-3` header. `[bracket]` directions, including in a language other than the line. `<whisper>`, `<emphasis>`, `<soft>`, `<fast>`, `<slow>`, `<stress>` and `<forceful>` on the selected words. Auto Tag is an editor starting point.
Validate: Preview: takes vary. No published price or time to first audio. A missing header falls back to s2.1-pro without an error. Mid-line language switch is experimental. Test French off-air before the Où est Ava ? live path.
Le WebSocket live accepte l’en-tête model: drama-3. Aucun temps au premier audio n’est publié. Un modèle omis ou inconnu retombe sur s2.1-pro. Le paragraphe multi-locuteurs du WebSocket nomme s2-pro et S2.1 Pro, pas drama-3.
Pas de timestamps mot, phonème ou visème documentés pour Drama 3.
Pricing
The 18 October 2026 rate card lists s2.1-pro, s2.1-pro-free, s2-pro and deprecated s1. drama-3 is absent. No hourly price is inferred.
Sovereignty & Compliance
API cloud. Ce n’est pas le chemin open-source de s2-pro.
Data residency: Non indiqué pour drama-3. API cloud, sans résidence UE ou suisse nommée sur la page produit ni sur la grille.
This sheet's verification
Fish Audio Drama 3 product page and docs, 18 Oct 2026. Arena ELO not copied.
Update note: Page produit FR/EN et doc Drama 3 relues le 18 octobre 2026. Grille et WebSocket relus le même jour. Aucun ELO ni aucun TTFA inventés.
This date covers provider-specific prices, capabilities and notes. Comparative benchmarks follow the synchronization and methodology shown above.