Fish Audio Drama 3 vs Eleven v4: One Model Wins the Preference Vote, the Other Wins the Words

ElevenLabs Eleven v4 and Fish Audio Drama 3 disagree about what a text-to-speech score is for. On Artificial Analysis' Provider Voice Arena, Eleven v4 sits second at an Elo of 1,334 and Drama 3 sixth at 1,278; on the Controlled Voice Arena, where both are scored against the same reference voices, Eleven v4 is fourth at 1,156 and Drama 3 twelfth at 1,085. Those are 56- and 71-point gaps, both comfortably outside the noise, and both say the same thing: listeners prefer Eleven v4. Then the same lab asks human reviewers whether each hard word in a sentence was actually pronounced correctly, and Drama 3 scores 93.2% against Eleven v4's 91.7%, with the two at opposite ends of the category that decides whether a brand name, a figure or a part code comes out intact. The preference gap is real and so is the pronunciation gap, and no single number reconciles them.

The gap that does not close, and the gap that does

Start with what Eleven v4 has not lost. It went to general availability on 2026-09-28, four weeks after Fish Audio's model shipped, and it has held the top two provider-voice slots through every board update since. The current spread over Drama 3 is 56 Elo on provider voices and 71 on controlled voices — and the controlled board is the harder test, because it removes each vendor's own voice catalogue from the comparison. Whatever Drama 3's lead among open-weights models means, it does not mean Drama 3 sounds better than Eleven v4 to a blind listener, and any comparison that implies otherwise is selling something.

What Drama 3 has is the category-by-category breakdown on Artificial Analysis' English Pronunciation Robustness benchmark, where 31 models are scored and the top is shared:

• Overall — Drama 3 93.2%, joint first with Ming-3.0-tts; Eleven v4 91.7%; Eleven v4 Turbo 90.1%.

• Standalone terms (brands, places, technical names) — Drama 3 95.9%, first on the board; Eleven v4 below it.

• Expanding shorthand (numbers, dates, units, notation) — Eleven v4 94.1%, first; Drama 3 93.6%, second.

• Preserving exact sequences (codes, paths, emails) — Drama 3 86.3%, second on the board behind Ming-3.0-tts at 90.6%; Eleven v4 78.8%, which is its weakest category by a wide margin.

• Contextually appropriate — Drama 3 95.3%; Google Gemini 3.8 Flash TTS leads the whole board at 98.0%.

The pattern is usable. If your script is mostly prose with a handful of numbers, Eleven v4's shorthand handling is the better one and the overall difference is under two points. If your script contains SKUs, case numbers, URLs, part codes or email addresses read aloud, Eleven v4 is the weakest model in the top group on exactly that category — 78.8% means roughly one in five exact sequences comes back wrong — and Drama 3 at 86.3% is still imperfect, but it is the difference between a support agent that reads a ticket number once and one that reads it twice.

Two syntaxes for the same direction

Both models let you write the performance into the script, and the notations are close enough that a team can port a script in an afternoon — which makes the differences that remain worth knowing.

ElevenLabs uses inline audio tags — [laughs], [whispers], [sighs], [said angrily in French accent], even [light rain] — dropped into the text where the delivery should change. The vendor documents audio tags for Eleven v4 Turbo, and Eleven v4 follows the same syntax; both went GA on 2026-09-28. Eleven v4 caps input at 10,000 characters per request, roughly ten minutes of audio, so a long-form script has to be chunked at sentence boundaries and stitched, with the stitching logic becoming your problem. Multi-speaker dialogue is supported, and the low-latency sibling quotes a median inference latency of about 100 milliseconds excluding application and network overhead.

Fish Audio's notation is a bracket convention with a paired-tag layer on top, documented in its own Drama 3 pages:

• Bracketed directions before the words they govern — [barely holding it together, forcing a smile] — written the way a director would brief a voice actor.

• Sounds and pauses in the same brackets — [shaky breath], [small laugh] — placed exactly where they occur.

• Paired tags that change only the wrapped words: whisper, emphasis, soft, fast, slow, stress, forceful. There is no equivalent per-word control documented on the ElevenLabs side.

• Directions written in any language, independent of the script's language. A Chinese direction bracket in front of an English line is documented and demonstrated. This is the capability I would test first, because it is the one that changes who can operate the tool: a director who does not write English can still direct an English performance.

• Phoneme pinning per language, with documented pinyin and romaji paths for Chinese and Japanese, plus a language-switch tag that flips one phrase into another language mid-line — marked experimental in the docs.

• Multi-speaker in a single request: one voice id per speaker in a reference_id array, each turn starting with a speaker tag such as <|speaker:0|>, with directions and paired tags working inside each turn.

The control surfaces are genuinely different in kind, not just in spelling. ElevenLabs gives you a tag vocabulary tuned to emotion and delivery. Fish Audio gives you emotion plus phonetic pinning plus language switching plus per-word scoping, which is closer to a mastering tool than a performance one. For audiobook and dubbing work, that second set is the more useful one. For conversational or expressive marketing copy, Eleven's tags read more naturally and the preference scores agree.

The money, on a script you actually have

Per-hour comparisons between these two are unreliable, because the vendors bill in different units and imply different speaking rates. Do the arithmetic on a fixed script instead. Eleven v4 lists at $80.00 per million characters, so a 100,000-character script costs $8.00. Drama 3 lists at $50.00 per million UTF-8 bytes, and English runs close to one byte per character, so the same script costs $5.00 — and $1.00 during the launch promotion, which runs to 2026-11-12 09:00 PT at $10.00 per million bytes.

That is the honest shape of the price difference: Drama 3 is 37% cheaper per English character at list and 87% cheaper during the promotion, not the 60% that a naive characters-versus-characters comparison of $80 against $50 would suggest if you forgot the units. Two further details belong in the budget rather than in the footnote. Bracket directions count as input text and are billed, on both platforms, so a heavily directed script costs 10-25% more than its word count implies. And at Fish Audio, a request whose model header is missing or misspelled does not fail — it falls back to s2.1-pro, which during the promotion bills at $15.00 per million bytes instead of $10.00. A typo in a header is a silent 50% price increase.

ElevenLabs' cost model is different in kind: the subscription tiers carry cloning, and instant voice cloning starts at the $6 Starter plan with professional clones from the $22 Creator tier. Fish Audio sells cloning through the same credit system as generation, and Drama 3 uses the same credits as S2.1 Pro. If cloning is the centre of your workflow rather than a feature you use twice, the comparison is a plan comparison, not a per-character one.

Licensing, and the clause that is not negotiable on either side

Neither vendor will let you clone a person without consent, and both say so explicitly. The difference is what happens to the model itself. Eleven v4 is a closed API: you licence the output, not the weights, and the practical constraint is the rate card. Drama 3 publishes weights on Hugging Face as fishaudio/drama-3 under the Fish Audio Community License v1.0 — free for research and non-commercial use, free for commercial use below US$500,000 in annual revenue for you and your affiliates, and requiring a separate licence above that threshold or if you want to offer Drama 3 itself as a hosted service. Artificial Analysis labels it open weights without full commercial-use rights, which is accurate.

That cuts both ways for a comparison. You cannot self-host your way out of ElevenLabs' price, and you cannot silently become a TTS provider on Fish Audio's weights. But you can, today, put Drama 3 on a 48 GB card and generate at a marginal cost of electricity — which is a different procurement conversation from $8.00 per 100,000 characters, and it is the one that decides long-running catalogue work.

Which one to put in front of the router

Neither Fish Audio Drama 3 nor ElevenLabs Eleven v4 is one of the models on OrcaRouter. Drama 3 is reached through Fish Audio's own API and several third-party platforms; Eleven v4 through ElevenLabs'. What sits in front of both, in the arrangement this blog keeps recommending, is the routing layer — one key across ~200 models at provider list price with no markup, automatic failover, and a routing DSL that can hold two voices behind one call site.

The reason that matters for this specific matchup is that the correct answer is "both, for different sentences". A production drama pipeline wants Drama 3 on the lines carrying proper nouns and codes, and it may still want Eleven v4 on the passages where listeners judge the voice itself. Model fusion and the routing DSL let a pipeline address that as one interface with a per-segment choice, rather than as two integrations, two rate tables and two sets of retry behaviour. When a vendor adjusts a price, the pass-through means the change lands on your side the same day rather than at the next contract renewal — which matters more here than usual, because Fish Audio's promotional rate expires on 2026-11-12 and the effective comparison changes overnight.

For a reader deciding only one: Eleven v4 if the deliverable is judged on how it sounds and the script is written in ordinary prose; Drama 3 if the deliverable has to be correct on names, numbers and codes, if the work is long-form and price-sensitive, if you want the weights, or if the person writing the direction does not write English. The arena says Eleven sounds better. The pronunciation board says Drama says it better. Both are the same benchmark suite, and the useful question is which of those two failures your production actually suffers from.