Qwen-Audio-3.1-TTS vs Qwen-Audio-3.0-TTS: What Two Months Bought

Qwen-Audio-3.0-TTS shipped on July 20, 2026 and went straight to the top of the independent speech leaderboards — the Plus tier reached 1,238 Elo on Artificial Analysis's Provider Voice Arena leaderboard, second on the board. Nine weeks later, on September 23, 2026, the vendor announced Qwen-Audio-3.1-TTS at the Apsara Conference in Hangzhou with a roughly 70% price cut attached. That timing makes this an unusual comparison: the model being replaced is not obsolete, it is benchmarked, proven and still near the top. So the real question is not which one is better. It is what specifically changed, whether any of it matters to your workload, and what happens to the score you already trust when the successor has not been scored at all.

Two months, five endpoints, one family

Alibaba did not ship a single model. The 3.1 release covers five: Qwen-Audio-3.1-ASR, Qwen-Audio-3.1-ASR-Next, Qwen-Audio-3.1-TTS, Qwen-Audio-3.1-TTS-Next and Qwen-Audio-3.1-Realtime. For anyone currently calling Qwen-Audio-3.0-TTS, only one of those is a drop-in — the plain TTS tier — and one is a genuine new product rather than an upgrade.

That second one is worth understanding even if you never call it. Qwen-Audio-3.1-TTS-Next is what Alibaba calls an "Audiogen" model: a single pass that produces voice, sound effects and background ambience together from text, timestamps and reference audio. Multi-speaker dialogue, podcast assembly, scene soundscapes, 48 kHz output. Alibaba Cloud's Model Studio documentation lists its limits plainly — 3,000 characters of input, 240 seconds of generated audio for podcasts and 120 seconds otherwise, 3 requests per second, WAV, MP3 or PCM out, and it accepts up to three reference clips of 30 seconds each in WAV, MP3 or OGG Opus. Raw PCM reference input is not supported. It is priced at $0.848 per million input tokens and $1.696 per million output tokens on the Beijing node, promotional rates excluded, and it handles Chinese and English only.

If you are on 3.0-TTS and your job is "turn this script into a voice," none of that changes your integration. If your job is "assemble a two-host podcast with music beds," 3.1-TTS-Next is a different product category and possibly the reason to look at this release at all.

The upgrade, line by line

• Languages — Qwen-Audio-3.1-TTS: vendor-reported 16 languages, with seven added over the previous generation. Qwen-Audio-3.0-TTS: 16 languages as listed in published coverage of the July release.

• Chinese dialects — Qwen-Audio-3.1-TTS: 20 dialect regions, carried over. Qwen-Audio-3.0-TTS: 20 dialect regions.

• Cross-language timbre transfer — Qwen-Audio-3.1-TTS: yes, one voice carried across Mandarin, Cantonese, English and Japanese. Qwen-Audio-3.0-TTS: not offered.

• Delivery control — Qwen-Audio-3.1-TTS: instruction-based control of emotion, pace and style. Qwen-Audio-3.0-TTS: 86 inline delivery tags.

• Noisy reference input — Qwen-Audio-3.1-TTS: handles noisy or echoed reference audio directly, no separate denoise mode. Qwen-Audio-3.0-TTS: clean reference audio expected.

• Independent score — Qwen-Audio-3.1-TTS: none yet. Qwen-Audio-3.0-TTS-Plus: 1,238 Elo on Artificial Analysis's Provider Voice Arena leaderboard as of August 2026.

The language count does not add up, and that matters

This is a small thing that will get smoothed over everywhere else, so it is worth stating. Alibaba's launch material says the 3.1 TTS tier adds seven languages. Published coverage of the July 3.0 generation — including our own comparison pieces from that month — listed 16 languages for the 3.0 tier. Seven added to sixteen is twenty-three, not sixteen, and Alibaba has not published a reconciliation of the two figures.

Possible explanations are unglamorous: the 3.1 number may count a different set, the 3.0 figure may have been overstated at launch, or "adds seven" may describe the ASR tier rather than TTS. We do not know which. What we do know is that the language count on both sides of this comparison is a vendor-reported number that has never been independently audited, and that anyone quoting "16 languages" as a hard fact about either model is quoting a marketing slide. Check the specific languages you need against the Model Studio documentation before you plan around a count.

Instruction control is the change that actually shows up in code

Of everything in the 3.1 TTS release, this is the one that alters how you write your prompts. The 3.0 generation controlled delivery through 86 inline tags — a fixed vocabulary you had to learn, insert at the right positions, and that did exactly what it said and nothing more. The 3.1 tier replaces that with natural-language instruction: you describe the emotion, the pace, the style you want, and the model interprets it.

The practical difference is expressiveness versus predictability. A tag vocabulary is a contract — the same tag produces the same delivery every time, which is what you want in a production pipeline where a voice has to be consistent across ten thousand clips. Free-form instruction is broader but looser, and it inherits the usual prompt-sensitivity problem: two phrasings of "warm but not sentimental" will not land identically, and neither will two calls of the same phrasing on different days.

If your current pipeline is built on those 86 tags, the upgrade is not free. You are trading a deterministic control surface for a probabilistic one, and the work of porting is not the API call — it is re-validating every prompt template you own against the new model's interpretation. Budget for that before you budget for the savings.

Cross-language timbre transfer, and what it is actually for

One reference voice, carried across Mandarin, Cantonese, English and Japanese, holding its identity as the language changes. On paper this reads like a feature bullet. In practice it solves a specific and expensive problem: a single presenter who has to exist in more than one language without sounding like a different person in each.

The traditional workaround is to cast a separate voice actor per language and accept that your brand voice is now four voices that share a script. The alternative has been to clone one speaker into each language, which is exactly the workflow where cloned voices tend to drift — the accent carries over, the timbre thins, and listeners notice within a sentence. Cross-language timbre transfer is the claim that neither compromise is necessary.

It is also, right now, entirely unverified. There is no third-party listening test on Qwen-Audio-3.1-TTS in any language, and Alibaba's own demos are the only evidence that the transfer holds. If this capability is the reason you are considering the upgrade, test it on your own reference audio before you commit — that is a five-minute experiment and it is the only one that answers the question.

What the price cut does to a 3.0 bill

Alibaba cut speech pricing across the line: synthesis by about 70%, realtime by about 85%, recognition by as much as 95%. The adjustment took effect on Alibaba Cloud Model Studio on September 22, 2026 at 00:00 Beijing time, covering the Beijing and Singapore regions, and it applies to the 3.1 TTS and 3.0 realtime endpoints. Post-adjustment, the TTS Flash tier lists at 1.5 yuan per million input tokens and 12 yuan per million output tokens; the realtime Flash tier lists at 1.5 for text input, 6 for audio input, 4.5 for text output and 12 for combined text-and-audio output, per million tokens.

Here is the part that matters if you are already running 3.0-TTS. The 3.0 Plus tier was tracking near $27.60 per million characters on Artificial Analysis through August, with the Flash tier near $15. If the ~70% reduction applies to the tier you use, your bill on the same workload drops to roughly a third of what it was — without you changing a line of code. That is the unusual thing about this release: for a 3.0 customer, the price cut is worth more than the model upgrade, and it arrives whether or not you migrate.

Two caveats worth holding onto. Percentage cuts are quoted against rates that differ by region and by tier, so the headline 70% is not a promise about your specific traffic. And Alibaba Cloud moved realtime transcription billing on the Singapore node from duration-based metering to token-based metering — $0.93 per million input tokens and $0.70 per million output tokens — which is a less predictable basis when silence, noise and multi-turn context all inflate token consumption. Read the new rate card against your own audio, not against the press release.

Where Qwen-Audio-3.0-TTS is still the right call

Three cases, and they are not edge cases.

When you need a number you can defend. Qwen-Audio-3.0-TTS-Plus has an independent Elo of 1,238 — the same board reads 1,259 for that model in September, still second, so the score has drifted upward rather than decayed — from a blind-listening arena with thousands of votes behind it. Qwen-Audio-3.1-TTS has no arena score at all. If your sign-off process requires third-party evidence, the older model is the one that has it, and a price cut does not change that.

When determinism beats expressiveness. If your production pipeline is built on the 86-tag control surface, you have a system whose output you can reason about. Instruction-based control is more capable and less predictable, and the migration cost is real.

When nothing in the upgrade touches your workload. If you synthesise Mandarin and English at moderate volume with a clean reference voice and a fixed set of delivery tags, then cross-language timbre transfer, noisy-input robustness and seven additional languages are all solving problems you do not have. Take the price cut, keep the model.

Running both without running two integrations

The awkward part of a generation-over-generation switch is that you often want both models live at once — 3.0 for the production path with the score behind it, 3.1 for the new languages and the transfer feature, and a way to move traffic between them without a redeploy. Both are closed hosted APIs on Alibaba Cloud Model Studio, so that is two endpoints, two rate limits and two failure modes behind one feature.

An honest note on where we sit: OrcaRouter does not route Qwen-Audio-3.1-TTS or any Qwen-Audio model, and we do not route Alibaba's speech endpoints at all. What we provide is the text-inference layer around a pipeline like this — one API key across 200-plus models at 0% markup, provider list price passed straight through, so a vendor price cut lands on our side the same day rather than after a repricing cycle. If the service driving your speech pipeline is a script generator, a summariser or a content planner, that means one contract instead of several, automatic failover when an upstream degrades, and a routing DSL for composing models into a single call. It does not mean you call Qwen's voice through us, and any page suggesting otherwise is wrong about our own catalogue.

The honest summary

Qwen-Audio-3.1-TTS is a real upgrade with three additions that matter — instruction-based delivery control, cross-language timbre transfer, and robustness to bad reference audio — plus a price cut that is worth more than any of them. Qwen-Audio-3.0-TTS is not superseded in the way a two-month-old model usually is: it still holds an independent benchmark score that its successor has not earned, and it still covers the Chinese dialect range that most of this family's differentiation rests on.

So the decision is narrower than the marketing suggests. If you are generating at volume, the pricing change is already yours. If you need the new languages or the timbre transfer, migrate and validate the feature on your own audio first. If you need a defensible quality number, wait until Qwen-Audio-3.1-TTS appears on a blind-listening arena — that is the single event that would settle this, and until it happens, the honest position is that the newer model is the cheaper one and the older model is the proven one.