AssemblyAI-TTS-1 vs Breeze TTS 2: One Elo Point Apart, 87 Points Apart

Put AssemblyAI-TTS-1 and Breeze TTS 2 on Artificial Analysis's Provider Voice board and they are indistinguishable: 1,227 against 1,228, one Elo point, one place on the ladder. Put them on the Controlled Voice board — where every model speaks the same eight cloned voices — and the gap opens to 87 points, 1,108 against 1,021, with AssemblyAI tenth and Breeze twenty-eighth. Both numbers are real, both were read on 30 October 2026, and the distance between them is the entire argument of this comparison. One of these models has a strong voice catalogue wrapped around an ordinary acoustic model. The other has a strong acoustic model wrapped around the cheapest per-character price in the category.

What follows is not a coin flip dressed up as analysis. The two products are aimed at different buyers. AssemblyAI-TTS-1 is a hosted API from a speech-recognition company that has just added a mouth to its ears — proprietary, $15 per million characters, six languages, unlimited concurrency. Breeze TTS 2 is a 3-billion-parameter open-weights model from BreezeBlue, a fifteen-person startup, with a hosted endpoint at $34 per million characters and a weight licence that forbids commercial self-hosting. The scores are close enough to be interesting and the commercial models could hardly be more different.

Two numbers before any of the rest of it

Artificial Analysis runs its blind listening tests on a rolling basis, and the boards move under the models. Breeze TTS 2 debuted in August at 1,215 Elo and sixth overall, and was reported as the top open-weights entry at that point. Two months and a dozen launches later it reads 1,228 and thirteenth — a higher score and a worse rank, because the field filled in above it. AssemblyAI-TTS-1 arrived on 27 October at 1,227 and twelfth, and read fourteenth two days later at the identical score. Anyone quoting a rank from a launch post without a date is quoting a stale fact.

• Provider Voice (native voices) — AssemblyAI-TTS-1 1,227, 14th; Breeze TTS 2 1,228, 13th

• Controlled Voice (eight shared cloned voices) — AssemblyAI-TTS-1 1,108, 10th; Breeze TTS 2 1,021, 28th

• Price per million characters — $15 vs $34

• Voices — 4,000+ library voices plus instant cloning vs designed voices from a text description, plus a 15,000-voice companion library

• Weights — none, API only vs downloadable but research/non-commercial only

• Languages — six at launch vs 50 claimed, model card tagged English and Chinese

• Throughput — 162 characters per second vs about 45

The controlled-arena split is the one to internalise, because it is the only measurement in this comparison that holds one variable constant. Provider Voice lets each model pick its own voices, so a model with 15,000 hand-designed characters can win on casting rather than acoustics. Controlled Voice strips that away. Breeze loses 207 Elo when its own voices are taken off the table; AssemblyAI loses 119. When the voice is fixed, Breeze is not in the conversation — 28th — and AssemblyAI is still in the top ten.

Why the blind-listening tie is a mirage

The provider board is a popularity contest between models that are each allowed to bring their best voice, and Breeze brings a lot of them. BreezeBlue's pitch is voice design: you describe a character in natural language and the model synthesises a timbre to match, then voice direction separates delivery from identity so you can ask for "hushed" or "rushed" while the speaker stays recognisably the same person. Its companion Voice Galaxy library carries more than 15,000 character voices. That is the kind of catalogue that wins a blind preference test, and it is also the kind of catalogue that tells you nothing about what happens when you hand either model a reference clip and ask it to speak in that voice.

Hand both models a cloned voice and the picture inverts. That is what the controlled arena does, and it is where the vendor's own numbers agree with the independent ones. Breeze reports 0.67 speaker similarity for its clones; AssemblyAI reports 0.64. Neither is impressive, and the difference between them is inside the noise. What is not inside the noise is the 87-point Elo gap on the same board. If your product needs a distinctive voice you designed yourself, Breeze's design tools are exactly the feature that matters and its controlled-arena weakness is exactly the risk — the model can cast a voice it made, and it is less reliable at holding a voice you made.

What the price actually buys

$15 against $34 is a 2.3x gap at list, and it widens or narrows depending on how you bill. AssemblyAI charges per input character including spaces and punctuation and throws in every library voice, every supported language, instant cloning, word timestamps and US-or-EU processing at one rate, with no concurrency ceiling on pay-as-you-go. Breeze's hosted endpoint is $34 per million characters on the same character basis, and the free weights do not change that for commercial users — the licence attaches to self-hosted outputs as well as to the weights themselves.

Run the arithmetic on a real script and the gap is about a dollar per finished hour of audio at normal speaking pace: roughly 81 cents on AssemblyAI against roughly $1.84 on Breeze, using the standard 15-characters-per-second estimate. For a company generating a hundred thousand hours of agent speech a month, that is the difference between $81,000 and $184,000 — and it is a difference you pay for voice design tools you may not use and an open-weights story you may not be allowed to act on.

The licence is the fork in the road, and it points the wrong way for Breeze

Open weights are usually the argument that ends a comparison like this: download it, run it on your own metal, stop paying per character. That argument does not survive contact with Breeze TTS 2's licence. The inference code is Apache 2.0; the weights and any outputs from a self-hosted deployment fall under a research-and-non-commercial licence that extends to derivatives, LoRAs, merges and quantisations. Commercial self-hosting requires authorisation from BreezeBlue. In practice the hosted API is the commercial route, which means the open-weights badge buys you inspection, experimentation and an escape hatch you have to negotiate — not a cheaper production path.

Compare that with AssemblyAI-TTS-1, which offers no weights at all and no pretence of openness. It is a hosted service with a published rate card and a data-residency choice. For a buyer deciding between the two, the question is not "open or closed" but "do I want to sign a research licence and negotiate for the right to run a 3B model myself, or do I want a per-character bill I can forecast?" The 3B size is genuinely practical — about 7.7 GiB of VRAM in the eager configuration, a 12GB GPU as the stated minimum, a real-time factor of 0.32 on an H100, so roughly three times faster than real time on the hardware BreezeBlue measured. It is the smallest credible self-host in this category. It is just not a self-host most companies can legally use.

The thing the Elo boards do not measure

Throughput is where AssemblyAI's advantage stops being a rounding error. Artificial Analysis recorded 162 characters per second for AssemblyAI-TTS-1 against roughly 45 for Breeze TTS 2 — a 3.6x difference in how fast the model turns text into audio. For a batch job, that is a queue-length difference. For a live agent, it is the difference between a model that can keep up with a fast talker and one that cannot. Breeze's answer to this is its streaming path, which it says reaches under 40 ms time to first audio; independent measurements put p50 time to first byte near 119 ms and time to first audio near 134 ms, with a 95th percentile around 163 ms. AssemblyAI's comparable independent figures are thinner — one lab, 159 ms median streaming latency — so treat both latency stories as closer to each other than either vendor's headline suggests.

On hard-string accuracy the gap is not close. AssemblyAI's vendor eval puts entity accuracy at 93.7% on phone numbers, URLs, codes and drug names, against 82.5% for Cartesia Sonic 3.6 and 68.0% for ElevenLabs Flash v2.5 — no Breeze figure exists in that table. Breeze has never published an equivalent, and its model card does not claim one. If your agent confirms order numbers, reads back email addresses or spells out confirmation codes, that is an unmeasured risk on the Breeze side and the vendor's best-documented strength on the AssemblyAI side.

Where Breeze is simply the better tool

None of the above makes Breeze a bad model; it makes it a specialist. If the product is a game with forty characters, a narrative audio app, or a companion that needs to sound like a specific person described in a prompt rather than a voice picked from a dropdown, Breeze's design-and-direction pipeline is a capability AssemblyAI does not have in any form. AssemblyAI's 4,000 voices are a library you choose from; Breeze's are voices you author. Add inline vocal events — a laugh, a sigh, a breath written into the script — and you have a tool for performance rather than for announcements. Its 50-language claim is also far beyond AssemblyAI's six, though the model card's own tags stop at English and Chinese and nobody has independently verified the wider list.

The counterargument is the one that keeps coming back: 45 characters per second is slow, 1,021 Elo on controlled voices is ordinary, and the licence blocks the obvious reason to want open weights. Breeze TTS 2 is a research artifact with a commercial on-ramp. AssemblyAI-TTS-1 is a commercial product with no research story. Choose according to which of those sentences describes your problem.

If you expect to need both

Most voice-agent teams end up with two TTS vendors, not because they want to but because quality, language coverage and price rarely land on one model. Neither AssemblyAI-TTS-1 nor Breeze TTS 2 is in the OrcaRouter catalogue today, so this is not a pitch for either; the honest advice is to wire the second one in before the first one has an outage. That is the case automatic failover exists for, and it is also the cheapest way to A/B two TTS models against your own traffic rather than against a vendor's prompt set — one key, a routing rule, and the traffic split you choose, instead of two contracts and two SDK integrations.

We route more than 200 models behind one API at 0% markup, passing provider list prices through unchanged, which matters here because TTS pricing is moving monthly: AssemblyAI entered at $15, Cartesia sits at $49, and the board that ranks them reshuffles every few weeks. The TTS legs we do carry are OpenAI's tts-1, tts-1-hd and gpt-4o-mini-tts and Google's Gemini 3.1 Flash TTS Preview — usable as the second leg while you decide whether the $15 rate or the voice-design toolkit is the thing your product actually needs.

The decision, stated plainly

Take AssemblyAI-TTS-1 if you are building a high-volume, mostly-English-plus-five agent that confirms numbers and reads structured data aloud, and if a 162 characters-per-second throughput advantage plus unlimited concurrency plus a rate card you can forecast matters more than having the prettiest voice. Take Breeze TTS 2 if the voice is the product — if you need to design characters, direct delivery, and hold a performance across a script — and if you can live with a non-commercial weight licence, roughly 45 characters per second, and a controlled-voice score that sits in the bottom third of the board.

The one point worth carrying away is that these two are not really rivals. They tie on a board that rewards casting and separate sharply on a board that rewards acoustics, and the model that looks weaker on the tie is the one that wins the measurement holding everything else constant. Price and licensing then decide it for most commercial buyers before the Elo argument gets a hearing.