Fish Audio Drama 3 vs Inworld Realtime TTS-2: One Reads a Script, the Other Follows a Conversation

The word "Realtime" in Inworld Realtime TTS-2 is a latency claim, not a description of what the model does. Inworld's speech-to-speech product for live conversations is a separate line item on its own pricing page; Realtime TTS-2 is a text-to-speech model fast enough for interactive use. That distinction matters the moment you line it up against Fish Audio Drama 3, which is built for the opposite job — a 35-billion-parameter model whose entire design is that you write stage directions into the script before a single line is spoken. Put them side by side on Artificial Analysis and the split is unusually clean: Drama 3 leads the Provider Voice Arena by 11 points, and Inworld leads the Controlled Voice Arena by 61. The direction of that gap says more about what each one is for than either number does alone.

Two products that were never aiming at the same thing

Inworld Realtime TTS-2 began as a research preview on 2026-05-05 and reached general availability on 2026-09-02. The pitch at GA was not fidelity: it was that the model takes the prior turns of a conversation, as audio, and reasons over tone and pacing before deciding how the next line should sound. That is a conversational-agent feature. A support bot that has just heard a frustrated customer does not need a stage direction — it needs to infer one from what was just said.

Fish Audio Drama 3, published on 2026-10-13, goes the other way. Every capability it advertises is about explicit authorship: bracketed directions before the phrase they govern, sound and pause events placed mid-line, paired tags that change only the wrapped words (whisper, emphasis, soft, fast, slow, stress, forceful), phoneme pinning for pronunciation, a language-switch tag for a single phrase, directions written in a language different from the script, and multi-speaker dialogue in one request via a reference_id array with per-turn speaker tags. Nothing in that list is inferable from context. All of it is something a writer types.

So the honest framing is not "which is better" but "which failure you cannot tolerate". A drama pipeline that has to render a twelve-character scene with emotional continuity will not get there on a model designed to infer tone from a chat history. A voice agent that has to answer in under a second will not get there on a model that wants a stage direction and 48 GB of VRAM. The overlap where both are genuinely viable is narrow: scripted, high-quality narration delivered through an API for interactive content — audiobook previews, NPC barks that are pre-rendered, marketing VO that has to be regenerated on a schedule.

The boards, read in the right order

Both figures below were read from Artificial Analysis on 2026-10-29:

• Provider Voice Arena — Fish Audio Drama 3 at rank 6 with 1,278 and a board-recorded $15.00 per million characters; Inworld Realtime TTS-2 at rank 8 with 1,267 and $20.80. Inworld Realtime TTS-2 Flash sits at rank 19 on the controlled board with 1,065.

• Controlled Voice Arena — Fish Audio Drama 3 at rank 12 with 1,085; Inworld Realtime TTS-2 at rank 7 with 1,146. Here Inworld is 61 points ahead.

• The asymmetry — Drama 3's 11-point provider advantage shrinks and inverts the moment the comparison is run on identical reference voices instead of each vendor's own catalogue. Its lead is largely a lead among the voices Fish Audio ships; Inworld's is not.

• The other asymmetry — Inworld's two entries straddle the board, flagship at 1,146 and Flash at 1,065, an 81-point gap between tiers of the same family. Drama 3 has no tier below it; you get the 35B model or you use something else.

This is the opposite pattern to the same two boards under Google Gemini 3.8 Flash TTS, where the closed model won on provider voices and lost on controlled. Neither pattern is noise; both are reminders that a single arena number describes the model-plus-voice-library bundle, and only the controlled board isolates the model.

Context awareness versus direction, concretely

The GA announcement for Inworld Realtime TTS-2 describes a model that reads conversation history — prior turns as audio — and uses it to shape pacing and tone. That is a vendor claim, not an independently scored capability, and Artificial Analysis' arenas do not test it; they test single lines against reference voices. So the feature Inworld is actually selling is the one no public benchmark currently measures. Treat the claim as promising and untested, and test it against your own dialogue rather than against a leaderboard.

Drama 3's direction capabilities are measurable in the narrow sense that you can see the output change when you change a bracket, and Fish Audio documents each of them with sample audio. The feature that no competitor in this comparison matches is the language-independent direction: a direction written in Chinese in front of an English line, documented and demonstrated. For an international post-production team, that is an operational unlock rather than a nice-to-have, and it is the sort of thing that shows up in a pilot within an hour.

The practical test for a team choosing between them is one page of your own dialogue. Render it with Inworld's conversation context set up as it would be in production, and render it with Drama 3 and explicit directions. Whichever produces fewer lines you have to hand-fix is the answer, and it will be a different answer for a chat agent than for a radio play.

Cloning, latency, and the numbers you can plan on

Inworld's instant voice cloning accepts 5 to 15 seconds of reference audio, and its professional-cloning beta takes about 10 minutes of audio for a higher-fidelity clone. Those are vendor figures and they are specific enough to test. Fish Audio's cloning sits inside the same credit system as generation, and Drama 3 uses the same credits as S2.1 Pro on subscription plans — but the licence is the part that differs, since Fish Audio's community licence explicitly bans cloning a real person's voice without consent, while Inworld's terms are the ordinary API terms.

Latency is where Inworld is clearly the right tool and the comparison should not pretend otherwise. Realtime TTS-2 and its Flash tier were built for interactive products, with a Flash tier priced 40% below the flagship for exactly the workloads that need it. Fish Audio's published speed figure for Drama 3 is 271 frames per second on a single H200 against 177 for S2 — a throughput number for batch rendering, not a time-to-first-byte for a live stream. If your requirement is "the voice must start inside a conversational turn", you are in Inworld's product, and no amount of Drama 3's 11-point provider lead changes that.

Price, at the rate you will actually pay

Inworld Realtime TTS-2 is $25.00 per million characters on demand, stepping to $20.00 on the $25 Creator plan, $17.50 at Builder, $15.00 at Developer and $12.50 at Growth, with Flash at $15.00 on demand falling to $7.00 at Growth. Fish Audio Drama 3 is $50.00 per million UTF-8 bytes at list and $10.00 during the launch promotion through 2026-11-12 09:00 PT. On a 100,000-character English script that is $2.50 at Inworld's on-demand rate against $5.00 for Drama 3 at list, or $1.00 during the promotion.

Three caveats belong in that arithmetic rather than under it. Drama 3 bills per UTF-8 byte, so a CJK script costs multiples of what an English one does for the same character count, while Inworld's character rate does not move. Drama 3's direction brackets are billed as input text, which is a 10-25% uplift on a heavily directed script — the exact scripts where you need the model. And Fish Audio's fallback is quiet: a request whose model header is missing or not an exact match is served and billed as s2.1-pro at $15.00 per million bytes rather than rejected.

The comparison that survives all three caveats is shorter than the price tables suggest. At list, Inworld is about half of Drama 3 per English character and better than it on controlled voices. During the promotion, Drama 3 is a third of Inworld's price. Anyone deciding on price alone is deciding on a date, not on a product.

Neither is on our catalogue, and what that changes

Fish Audio Drama 3 and Inworld Realtime TTS-2 are both reached through their own vendors' APIs and third-party platforms; neither is one of the models on OrcaRouter, and the routing argument here is about the rest of the stack rather than about these two. A voice product rarely calls one speech model and nothing else — it calls an LLM for the line, a speech model for the audio, sometimes a second speech model for a different segment, and it has a deadline. Consolidating the non-speech half behind one key at provider list price with no markup, with automatic failover so a single upstream hiccup does not stall the render, is what keeps the speech choice a configuration decision rather than an architecture one.

Where the routing DSL earns its keep in this specific matchup is the mixed pipeline: pre-rendered dramatic lines through Drama 3 and live conversational turns through Inworld, addressed as two routes on one key with the choice made per segment. Model fusion covers the case where you cannot decide — render the same line through both, score it against a reference, and keep the winner. Neither of those is a way around either vendor's price. They are a way around maintaining two integrations to save a line in a config file.

For a reader deciding one: Inworld Realtime TTS-2 if the speech has to react to a conversation, if latency is the constraint, if you want a Flash tier to absorb volume, and if a closed API is acceptable. Fish Audio Drama 3 if the speech is authored rather than reactive, if pronunciation of names and codes has to be reliable, if the direction has to be written down, if you want the weights, or if the brief will be written in a language other than the script. The 11-point provider lead is the least important number in that list; the 61-point controlled deficit is the one to check against your own reference voices before you commit.