AssemblyAI-TTS-1 vs Qwen-Audio-3.0-TTS: The Model in the Title Is Not the One to Buy

This comparison names Qwen-Audio-3.0-TTS, and that is a problem worth fixing before any numbers are quoted. Alibaba retired 3.0 as its flagship speech model on 23 September 2026, when the Qwen-Audio-3.1 series launched at the Apsara Conference and put Qwen-Audio-3.1-TTS-Plus at the top of its own TTS line. So the honest version of this matchup is three models, not two — AssemblyAI-TTS-1, the closed $15-per-million-character endpoint that went generally available on 27 October 2026, against both generations of Qwen's speech stack — and the answer changes depending on which Qwen you actually meant. If your contract, your integration or your procurement note says 3.0, read the next section. If it says 3.1, skip it.

First, the correction: 3.0 is not the current Qwen

Qwen-Audio-3.0-TTS shipped on 20 July 2026 in two tiers. Flash was built for real-time use with a first-packet latency around 300 ms; Plus was the quality tier at $27.59 per million characters through Alibaba Cloud Model Studio. The family covers 16 languages plus 20 Chinese dialect regions, supports 86 inline control tags such as [angry] and [giggles] alongside free-form natural-language style direction, does noise-robust zero-shot cloning, outputs at 48 kHz, and synthesises up to three minutes in one pass. It is closed and hosted only — no open weights, which was a deliberate departure from Alibaba's earlier open-weight Qwen3-TTS.

Ten weeks later the 3.1 stack replaced it across ASR, TTS, TTS-Next and realtime. The 3.1 TTS tier carries the same 8-voice lineup, lists at $19.3 per million characters on Artificial Analysis's price column, and is the entry most Qwen comparisons since late September have actually meant. Practically: 3.0 is still callable if you are already integrated, 3.1 is what a new buyer gets quoted, and any article that puts 3.0's July score next to an October launch without saying so is selling you a stale comparison.

The board, read once, for all three

Artificial Analysis's Provider Voice Arena, read on 30 October 2026, so one instrument for every row below:

• Qwen-Audio-3.1-TTS-Plus — third, 1,305 Elo, 2,281 samples, released September 2026.

• Qwen-Audio-3.0-TTS-Plus — seventh, 1,269 Elo, 2,794 samples, released June–July 2026.

• AssemblyAI-TTS-1 — fourteenth, 1,227 Elo, 1,632 samples, released October 2026.

• ElevenLabs Eleven v4 Turbo and Eleven v4 sit first and second at 1,336 and 1,334, for the reference point this article is not about.

The 78-point gap between AssemblyAI-TTS-1 and the current Qwen TTS tier is real but is not the widest gap in this matchup — ElevenLabs is 31 points above Qwen and 109 above AssemblyAI on the same reading. What matters more is that a 1,632-sample row for a model four days old is thin evidence, and the intervals on that board are wide enough that a re-run could move AssemblyAI-TTS-1 a few places without anything about the model changing.

Sixteen characters per second is the number that decides this

Here is the measurement almost nobody puts in a Qwen comparison, and it is the one that matters for the job AssemblyAI-TTS-1 was built for.

Artificial Analysis measures AssemblyAI-TTS-1 at 162 characters per second of synthesis. Qwen-Audio-3.0-TTS-Plus, on the same style of measurement, generates about 16 characters per second — the slowest model anywhere near the top of the arena. The 3.1 tier improved on that, but generation speed is still the family's weak axis rather than its strength.

Now convert to the only unit a live agent cares about. Ordinary speech runs at roughly 900 characters per minute, which is 15 characters per second. A model that generates 16 characters per second is, within a rounding error, real-time factor 1.0 — it can just about keep ahead of one stream of playback, and it cannot serve two concurrent conversations without more hardware behind it. A model at 162 characters per second has roughly ten streams of headroom per unit of the same compute.

That reframes the Elo gap. A 78-point blind-preference deficit is a quality question you can answer with your own corpus. A ten-times throughput deficit is a capacity question that arrives as a bill, because the fix is more GPU-seconds per concurrent caller, and on Alibaba's hosted pricing that is not a lever you control.

Both caveats, stated plainly. These are two vendors measured on one instrument, not a controlled head-to-head; generation throughput depends on hardware, batching and sequence length. And Qwen's 12.5 Hz speech tokenizer — deliberately low compared with the usual 50–75 Hz — is what keeps its first-packet latency near 300 ms on the Flash tier. Qwen optimised for latency-to-first-sound. AssemblyAI optimised for characters-per-second once the turn starts. Those are different bets and a phone agent feels both.

Languages and control tags are where Qwen is not close to being beaten

If the throughput argument decides the call-centre case, the language argument decides everything else, and it is not subtle:

• Coverage — Qwen-Audio-3.0-TTS ships 16 languages and 20 Chinese dialect regions; AssemblyAI-TTS-1 launched with six (English, Spanish, French, German, Italian, Portuguese) and multiple accents within them.

• Direction — Qwen exposes 86 inline control tags plus free-form style prompts; AssemblyAI-TTS-1 has no emotion tags, no SSML, no speed parameter and no pronunciation dictionary at launch, and its own post says written disfluencies are read as written rather than acted on.

• Cloning — Qwen does noise-robust zero-shot cloning; AssemblyAI does instant cloning from a 24-second sample with a speaker-permission confirmation step. Neither offers a trained professional clone tier, which is an ElevenLabs product, not a Qwen or AssemblyAI one.

• Output — Qwen synthesises at 48 kHz with a three-minute single-pass ceiling; AssemblyAI streams 8–48 kHz PCM, µ-law or A-law with word timestamps and a barge-in cancel message, which is the plumbing a telephony stack actually needs.

• Openness — neither publishes weights. Both are hosted-only, in-region, vendor-priced.

That table has an obvious reading: AssemblyAI-TTS-1 is not a multilingual product and does not pretend to be. If your agent needs to speak Japanese, Korean or Mandarin, the Qwen family is the answer and this article ends there. If your agent speaks English to English callers and occasionally Spanish, six languages is enough and the Qwen capability list is paying for features your prompts will never use.

Price per hour, three ways

Using the same estimate as everywhere else in this series — about 900 characters of speech per minute, so roughly 54,000 per hour — the hourly cost of synthesis is:

• AssemblyAI-TTS-1 — $15 per million characters, all voices and languages and cloning included, unlimited concurrency with no fee: about $0.81 per hour of audio.

• Qwen-Audio-3.0-TTS-Plus — $27.59 per million characters on Alibaba Cloud Model Studio: about $1.49 per hour.

• Qwen-Audio-3.1-TTS-Plus — $19.3 per million characters as recorded by Artificial Analysis: about $1.04 per hour.

Note the discrepancy you can see on the board yourself: Artificial Analysis lists both Qwen generations at $19.3 per million characters, while Alibaba's Model Studio card for 3.0-Plus says $27.59. Those are two instruments — the vendor's list price and the board's normalisation — and neither is the invoice you will receive once committed volume pricing enters the conversation. The hourly figures above use each number as labelled, so treat the 3.1 rate as a comparison unit and the 3.0 rate as the closest thing to a list price.

AssemblyAI is the cheapest of the three, by 22% against the current Qwen tier and by 45% against the one in this article's title. But the price column is not where the money is. The money is in the throughput column, because a model running at real-time factor 1.0 needs roughly ten times the hardware per concurrent caller that a model running at 162 characters per second does — and hosted per-character pricing hides that multiplier inside the provider's capacity planning until it surfaces as throttling or latency at peak.

What a router actually adds here

The realistic deployment for a team facing this decision is neither model alone. It is English turns on the cheap, fast endpoint and non-English turns on the Qwen tier, switched by the language of the caller — which is a routing problem, not a model problem. That is the layer OrcaRouter covers for the rest of the 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 composes several models behind a single endpoint instead of several contracts and several SDKs.

To be precise about what that does and does not mean: neither AssemblyAI-TTS-1 nor any Qwen-Audio TTS tier is on the OrcaRouter catalogue today. The speech models we route are OpenAI's tts-1, tts-1-hd and gpt-4o-mini-tts and Google's Gemini TTS previews. So the routing argument applies to the language model that writes the sentence your TTS reads, and to any speech model you add later that we do carry — not to these two.

Who this comparison is actually for

Buy Qwen-Audio-3.1-TTS-Plus — not 3.0 — if your agent has to speak more than six languages, if you need directed performance rather than plain reading, or if you are already inside Alibaba Cloud and the contract is the constraint. It is third on the blind board, it costs $1.04 an hour, and its generation speed is the thing to test against your own concurrency target before you commit.

Buy AssemblyAI-TTS-1 if the workload is English-first telephone conversation at volume: one flat $15 rate, unlimited concurrency with no fee, 162 characters per second of headroom, word timestamps and a barge-in cancel built in, and an independent latency measurement of 159 ms to weigh against a vendor claim of 90 ms. Accept that you are giving up ten languages, 86 control tags, and any ability to fix a mispronounced abbreviation with a dictionary — and that its weakest independent sub-score, 73.8% on expanding shorthand, is exactly the failure a caller notices.

The mistake to avoid is the one the title of this piece invited: quoting a July model's score in an October decision, then discovering at integration that the tier you were quoted is the one that replaced it. Check which Qwen you are actually being sold before you compare it to anything.