Simba 3.5 vs Qwen-Audio-3.0-TTS: Two Ways to Solve the Language Problem, Only One of Which Is Documented

Both SpeechifyAI Simba 3.5 and Alibaba Qwen-Audio-3.0-TTS will read your script in languages other than English. They get there by opposite routes. Simba 3.5 publishes a list: 48 languages, one model id, every catalogue voice speaks all of them, and Estonian and Cantonese are named as the exceptions. Qwen-Audio-3.0-TTS publishes an instruction field, and the documented way to make it speak Dutch is to send it the sentence "Speak in Dutch, slowly and clearly." One of those is a spec you can build a locale matrix from. The other is a prompt, and the difference between them is the most useful thing in this comparison.

Age matters here too, in a way it usually does not. Qwen-Audio-3.0-TTS shipped on 20 July 2026 and has since been superseded at the top of Alibaba's range by Qwen-Audio-3.1-TTS-Plus, which is what currently occupies third place on the Controlled Voice board. The 3.0 Plus tier is still listed and still the one with the stronger measured position on Provider Voice. Simba 3.5 arrived on 21 October 2026 and is three weeks into a pricing schedule it has already revised once.

The numbers, side by side where they exist

Artificial Analysis scores both models on both boards, which makes this one of the few matchups in the speech space where you can compare like with like rather than extrapolating from a predecessor.

• Controlled Voice — SpeechifyAI Simba 3.5 rank 1, Elo 1,192, plus or minus 17, 1,783 samples

• Controlled Voice — Alibaba Qwen-Audio-3.0-TTS-Plus rank 10, Elo 1,130, plus or minus 12, 3,596 samples

• Provider Voice — SpeechifyAI Simba 3.5 rank 5, Elo 1,294, plus or minus 17, 1,732 samples

• Provider Voice — Alibaba Qwen-Audio-3.0-TTS-Plus rank 10, Elo 1,264, plus or minus 13, 3,056 samples

• Controlled Voice — Alibaba Qwen-Audio-3.1-TTS-Plus rank 3, Elo 1,176, plus or minus 12 — the newer Alibaba model, and the one to watch

• Board price — Qwen-Audio-3.0-TTS-Plus $19.31 per million characters on both boards; Simba 3.5 listed at $6.58, a plan-embedded rate that does not match Speechify's published list

Simba 3.5 leads by 62 points on the controlled board and 30 on the provider board, and the intervals separate cleanly in both cases. That is a straightforward quality result in Simba's favour, and it is worth noting what does not explain it: sample count. Qwen-Audio-3.0-TTS-Plus has twice as many controlled-voice votes behind its estimate, so this is a well-measured 1,130 rather than a noisy one.

The interesting wrinkle is Alibaba's own line. Qwen-Audio-3.1-TTS-Plus sits 46 points above Qwen-Audio-3.0-TTS-Plus on the controlled board, which means the current-generation Alibaba model is much closer to Simba 3.5 than the 3.0 is. Anyone comparing against the 3.0 tier is measuring a model that has already been improved once.

Forty-eight listed languages against an instruction field

Speechify's documentation is unusually concrete for a speech API. Simba 3.5's 48 languages are enumerable, each catalog voice — and each paid-plan cloned voice — speaks all of them, and the model takes a standard BCP-47-style language tag such as ja-JP, with any regional variant of a listed language in scope. The pair that is excluded, Estonian and Cantonese, is named. Speechify also flags which voices are staff picks in each locale and ranks them with a staff_pick_rank field, because a voice keeps the accent of its source recording and its own recommendation is to generate with a voice native to the language.

Alibaba's approach is flexible in a different direction. Qwen-Audio-3.0-TTS accepts a natural-language instruction that describes delivery — pace, emotion, persona — and the documented technique for a target language is to say so in the instruction, with the instruction required to match the language of the text. It also supports explicit pronunciation control through language_hints and hot_fix, SSML and LaTeX input, and inline emotion and non-verbal tags such as [excited], [sad], [laughing] and [sighing]. Non-streaming requests return an audio URL valid for 24 hours; streaming returns base64 chunks. The model is served from Singapore and Beijing.

That tag vocabulary is a real advantage and Simba 3.5 does not match it. Speechify applies break, prosody rate, prosody pitch, prosody volume and sub alias; emphasis and emotion tags are accepted and not applied. If your product needs a laugh in the middle of a line, or a sigh at the end of one, Qwen-Audio-3.0-TTS does that today and Simba 3.5 does not.

The trade is straightforward once stated: Qwen gives you expressive control and leaves language support to prose; Speechify gives you a countable language list and a narrower expressive surface. For a localisation pipeline, the countable list is worth more, because a locale matrix is a thing you can test, quote in a statement of work, and fail a build against. "Speak in Dutch, slowly and clearly" is a thing you have to listen to.

Pricing, converted into a number you can compare

Alibaba bills synthesis by input character in yuan, with regional rates that differ between the Beijing and Singapore deployments. Qwen-Audio-3.0-TTS-Plus is listed at ¥1.4 per 10,000 characters in Beijing; the Flash tier is ¥1.0. Artificial Analysis normalises the Plus tier to $19.31 per million characters, which is the figure to use if you are comparing against a dollar-denominated vendor rather than building a regional budget. Each listed Qwen-Audio model carries a one-million-token free allowance in the Beijing region for 90 days from activation or release, whichever is later.

Speechify publishes in dollars per million characters and has changed its mind once already:

• Simba 3.5, list from 22 November — $15.00 per million on Starter, $12.00 on Pro, $10.00 on Scale

• Simba 3.5, through 21 November — $10.00, $8.00 and $6.00, the Simba 3.2 and 3.0 rates

• Simba 3.5, original launch list on 21 October — $20, $16 and $12, superseded by the 9 November cut

• Simba 3.2 and Simba 3.0 — $10, $8 and $6 per million characters, unchanged

At list, Simba 3.5 lands between roughly half and three-quarters of the Qwen-Audio-3.0-TTS-Plus board rate depending on plan, and during the offer window it is a third to a half. Neither is expensive in absolute terms — five million characters a month is under $100 on either — so price is unlikely to be the deciding factor unless you are running a very large catalogue. Language coverage and expressive control probably are.

What the 21 November date does to this decision

Speechify's Simba 1.6 ids — simba-english and simba-multilingual — were withdrawn from selection at API version 2026-09-21. On 21 November 2026 both are served by Simba 3.5 in every language and billed as simba-3.5, on any API version that can still name them. An integration that has been running simba-multilingual since the summer will change voice and change price on that date without a code change.

Alibaba has no equivalent scheduled event in this comparison. Qwen-Audio-3.0-TTS remains listed next to the 3.1 SKUs, and the 3.1 line is billed on a different unit — tokens rather than characters — so the two generations cannot be compared by dividing one rate card cell by another. If you are on 3.0 today, the migration to 3.1 is a real project rather than a one-line model id change, and the 46-point controlled-board gap between the two tiers is the reason to schedule it.

Where each one is the right call

Pick Simba 3.5 if your problem is breadth under a single voice id: a product that ships one brand voice across dozens of locales, a back catalogue that has to be re-rendered in languages nobody on the team speaks, or an accessibility read where the same cloned voice has to follow the user from English into Spanish without sounding like a different person. The counted language list, the staff-pick ranking and the zero-shot cloning without approval are the features that make that work, and the 62-point controlled-board lead means the quality holds up when the voice is not one the vendor chose.

Pick Qwen-Audio-3.0-TTS if your problem is performance rather than coverage: audiobook and podcast work where an [excited] tag earns its place, Mandarin-first content where Alibaba's voice catalogue and pronunciation controls are deepest, or a deployment that has to sit in a Singapore or Beijing region. The instruction field is genuinely more expressive than what Speechify exposes, and the non-streaming URL and SSE chunk paths cover both batch and live rendering.

OrcaRouter does not carry either line — Simba is reached through Speechify's own API, Qwen through Alibaba Cloud Model Studio — and nothing here should be read as a claim otherwise. What we do route is a smaller synthesis catalogue that works as a neutral control when you test: openai/tts-1 at $15.00 per million characters and openai/tts-1-hd at $30.00, both at provider list price with no markup added, alongside google/gemini-3.1-flash-tts-preview. Rendering the same three-minute script through those and through both candidates is the cheapest way to settle a question that two leaderboards can only narrow.

The part both vendors leave you to work out

Neither model will tell you whether its non-English output is good enough for your audience, and the arena will not either: the Controlled Voice board is an English-language evaluation on eight English reference voices, four American and four British. A 62-point lead measured in English says nothing direct about Japanese, Polish or Brazilian Portuguese. Speechify's answer is to point at staff picks per locale; Alibaba's is to point at instructions and a voice list. Both are asking you to listen.

That is the honest close on this pair. One is newer, cheaper at list, better measured in English, and documented down to the language count. The other is more expressive, more established in Chinese-language work, mid-way through a generational transition, and controlled by an instruction field rather than a spec. Test both on the two locales that actually carry your revenue, and the rest of this comparison becomes a tiebreaker rather than the decision.