AssemblyAI-TTS-1 vs Fish Audio Drama 3: A Checkout Button Against a 41.5 GB Download

The first thing to understand about AssemblyAI-TTS-1 against Fish Audio Drama 3 is that one of them is a file. Drama 3, released on 13 October 2026, is a 35-billion-parameter mixture-of-experts speech model published on Hugging Face as fishaudio/drama-3 — roughly 41.5 GB of weights before you add a codec, and about 48 GB of VRAM to serve it in FP8. AssemblyAI-TTS-1, generally available since 27 October 2026 at $15 per million characters, is an HTTPS endpoint in US-East or EU-North that answers in under a tenth of a second and asks you for nothing but text. Every comparison between them that starts with Elo scores is starting in the wrong place, because the two products are separated by an operational decision before they are separated by a quality one.

What $15 buys, and what 41.5 gigabytes buys

The AssemblyAI-TTS-1 side is short to describe. A flat $15 per million input characters covering all 4,000-plus library voices, instant cloning from a 24-second sample, word timestamps, six launch languages (English, Spanish, French, German, Italian, Portuguese), US or EU processing, WebSocket and HTTP endpoints plus an OpenAI-compatible route, a cancel message for barge-in, and unlimited concurrency with no fee attached to it. No SSML, no speed parameter, no pronunciation dictionary, no emotion tags, no professional cloning tier.

The Drama 3 side is a different shape of product altogether. It is a 35B MoE with about 3 billion parameters active per token, initialised from Qwen3.5-35B-A3B, 40 layers, 8 of 256 routed experts plus one shared — which is the reason it can sound expressive and still fit on a single accelerator. It covers 83 languages, clones voices, handles up to 16 speakers in one pass, and accepts bracketed natural-language stage directions such as [sighs] that the model acts on. It is also, per Fish's own model card, built for directed performance: dubbing, narration, multi-character scenes. The sub-100 ms telephone loop is not what it was tuned for.

That distinction is not a knock on either. It is the reason the matchup exists at all: Drama 3 is the only model in this series you can own outright, and AssemblyAI-TTS-1 is the only one in this series you can put in front of a caller this afternoon without buying a GPU.

The concurrency math is the real argument

Both vendors publish a latency number and neither is directly comparable, so treat the following as two separate claims rather than a head-to-head.

• Drama 3, vendor model card — 88.9 ms median time to first audio at a single stream on one H200, real-time factor 0.131, 271 frames per second. Serving support landed within days of release and the same hardware holds 255 ms median time to first audio at 128 concurrent streams, or about 222 seconds of audio per wall-clock second.

• AssemblyAI-TTS-1, vendor lab — 47 ms median time to first byte; measured independently by Cekura at 159 ms median streaming startup and 296 ms at p90, with 162 characters per second of throughput on Artificial Analysis's measurement.

Look at what happens to the Drama 3 number as load rises: 88.9 ms alone, 255 ms at 128 streams. That is a perfectly good curve for a dubbing pipeline and a marginal one for a call centre, where every millisecond before the caller hears something is a millisecond of dead air. AssemblyAI's counter-claim is not a faster number, it is the absence of a curve — unlimited concurrency at one flat rate, because the scheduling problem is the vendor's and you never see it. Whether that promise survives your own traffic is exactly the thing to test, since an unlimited-concurrency claim with no published queueing behaviour is a claim about someone else's cluster.

Two bills that are not in the same units

Fish lists Drama 3 at $50.00 per million UTF-8 bytes, which the vendor's own documentation equates to roughly twelve hours of speech — about $4.17 per hour of audio. There is a launch promotion at $10.00 per million bytes, about $0.83 per hour, running through 12 November 2026. AssemblyAI-TTS-1 at $15 per million characters works out to roughly $0.81 per hour of speech, using the same charitable assumption of about 900 characters per minute of ordinary conversational audio.

So during the promotion the two are priced within a rounding error of each other, and at list price Fish is about five times more expensive per hour of output — and that is before you buy the H200. Self-hosting does not remove the cost, it converts a per-character bill into a GPU-hour bill plus a person who knows how to serve it.

Two traps are worth naming because they are cheap to fall into.

• Characters are not bytes. Fish bills by UTF-8 byte, AssemblyAI by character. English text is close to one byte per character, so the two look comparable; Japanese, Arabic or emoji-heavy scripts inflate the byte count and quietly raise the Drama 3 bill by 30–100% on the same script.

• A missing model header does not fail loudly. Fish's own documentation notes that an absent or misspelled model identifier falls through to s2.1-pro at $15.00 per million bytes — a 50% surcharge over the promotional Drama 3 rate, for a different model, with no error in the response.

There is a third number in circulation and it is worth naming because you will meet it on a leaderboard. Artificial Analysis's price column records Drama 3 at $15.0 per million characters, which does not reconcile with Fish's $50 per million bytes or with the promotional $10. Three figures — list, promotion, board normalisation — and only the first two come from the vendor. The board's number is a comparison unit, not a quote.

The license gate at the bottom of the post

Drama 3 is not open source in the sense most teams assume. It ships under the Fish Audio Community License v1.0: commercial use is free below US$500,000 in annual revenue, a separate licence applies above that, and reselling the model as a hosted TTS API requires its own agreement regardless of size. Artificial Analysis records it as open weights without full commercial-use rights, which is the accurate description.

AssemblyAI-TTS-1 has no such gate because it has no weights — you cannot self-host it, fine-tune it, or keep running if the vendor changes its pricing. That is the trade nobody puts in a comparison chart: Drama 3's licence limits what you can do with the model, and AssemblyAI-TTS-1's architecture limits what you can do without the vendor.

The quality gap, stated once

On Artificial Analysis's Provider Voice Arena — read on 30 October 2026, so the same run for both — Drama 3 ranks sixth at 1,278 Elo and AssemblyAI-TTS-1 fourteenth at 1,227. On the Controlled Voice Arena Drama 3 is twelfth at 1,085 and AssemblyAI-TTS-1 tenth at 1,110, which flips the order and tells you the two boards are measuring different things: stock voices in the first, cloning fidelity against a fixed reference in the second. Drama 3 is the stronger ranker of the two on natural voice, and the two are within noise of each other once the reference recording is held constant.

The number that separates them most sharply is pronunciation robustness, where Drama 3 leads the board outright — sources put the figure anywhere between roughly 93% and 95% depending on when the page was read, so treat the ranking as the fact and the decimal as decoration. AssemblyAI-TTS-1 scores 81.7% on the same instrument, sixteenth, with its weakest sub-score on expanding shorthand at 73.8%. If your input is full of abbreviations, Drama 3 handles them without a dictionary. If your input is a phone number read aloud, Cekura's independent test puts AssemblyAI-TTS-1 first in its field at 91.7% on exactly that category.

Neither vendor's own benchmark should be read as independent. AssemblyAI's 99.0% word accuracy, 3.85 predicted MOS and 85.5 TTSDS score come from its own harness on its own held-out sets; Fish's 88.9 ms and 271 frames per second come from its own model card. The third-party measurements — Artificial Analysis, Cekura, Coval — are the only ones in this article that were produced by someone with no product to sell.

Pick by what you can operate, not by what ranks higher

There are exactly three situations in which Drama 3 is the right answer. You need expressive, directed performance with stage directions and multiple speakers in one pass — no argument, AssemblyAI-TTS-1 has no equivalent. You need a language outside the six AssemblyAI launched with — 83 against 6, and the comparison ends there. Or you need the model on your own hardware for data-residency reasons that a US or EU vendor region does not satisfy — which is a real requirement for some regulated workloads and the only one that justifies the GPU.

Everyone else is buying a phone agent, and a phone agent wants a flat bill, no queue behind someone else's H200, and a pronunciation problem it can fix. AssemblyAI-TTS-1 gives you the first two and does not yet give you the third. That is the honest shape of it: Drama 3 is the better model and the worse product for this job.

The practical compromise most teams will land on is both — Drama 3 for the generated audio you ship as content, AssemblyAI-TTS-1 for the turn-by-turn speech in the live loop, with one fallback path between them. Keeping a second TTS vendor wired up is the same discipline OrcaRouter applies to the language model half of that stack: one API for 200-plus models at provider list price with 0% markup, automatic failover when a provider degrades, and a routing DSL that lets one endpoint serve several models instead of several contracts. Neither of these two speech models is on the OrcaRouter catalogue today — the speech routes we carry are OpenAI's tts-1 family and Google's Gemini TTS previews — so read that as the shape of the argument rather than a suggestion that you can buy Drama 3 here.

What would change this verdict: Fish publishing a concurrency curve that holds under 150 ms at call-centre volume, or AssemblyAI adding a pronunciation dictionary and four more languages. Watch for the second one first. It is a configuration change, and the company that shipped a flat $15 rate card has already shown it prefers competing on the bill.