SI Agent News
DevelopingOpen models

Simba 3.5 and Qwen-Audio-3.0-TTS offer different paths to multilingual speech

SpeechifyAI and Alibaba provide distinct methods for global voice synthesis, choosing between a strict language list and flexible natural language prompts.

48languages supported by Simba 3.5

The short version

Why it matters

For developers, the choice depends on whether they need a predictable, testable locale matrix or high-fidelity expressive control. Simba is better for consistent global branding, while Qwen is superior for content requiring specific emotions like sighs or laughter.

Two distinct technical paths

SpeechifyAI's Simba 3.5 and Alibaba's Qwen-Audio-3.0-TTS both enable non-English speech synthesis but use opposite architectural philosophies. Simba 3.5 operates on a documented list of 48 languages, where every catalog voice is capable of speaking every language on that list. It uses standard BCP-47-style language tags, such as ja-JP, to identify specific regional variants.

Alibaba takes a different approach by using an instruction field for language and delivery. To make Qwen-Audio-3.0-TTS speak a specific language like Dutch, users are instructed to send a natural language prompt such as "Speak in Dutch, slowly and clearly." This method allows for more flexible delivery descriptions, including pace, emotion, and persona, but lacks a concrete list of supported locales.

As OrcaRouter reported, the difference is primarily between a technical specification and a prompt. Simba provides a countable list that developers can use to build a locale matrix for testing and quoting in statements of work. Qwen provides a flexible instruction field that requires the user to listen to the output to ensure it meets the desired quality.

Performance and expressive control

In terms of measured quality, Simba 3.5 currently holds a lead on the Controlled Voice board with an Elo of 1,192. This is compared against Qwen-Audio-3.0-TTS-Plus, which sits at rank 10 with an Elo of 1,130. These scores are based on thousands of samples, meaning the results are well-measured rather than noisy. Simba leads by 62 points on the controlled board and 30 points on the provider board.

While Simba leads in measured English performance, Qwen-Audio-3.0-TTS-Plus offers a wider range of expressive controls. It supports explicit pronunciation hints, SSML, and LaTeX input, along with inline tags for non-verbal sounds like [excited] or [sighing]. Simba 3.5 supports tags for break, prosody rate, and volume, but it does not support emotion or non-verbal tags like laughter.

The trade-off for the user is straightforward based on the project needs. Qwen provides expressive control and leaves the language support to prose, which is ideal for audiobooks or podcasts. Simba provides a countable language list and a narrower expressive surface, which is better suited for products needing a consistent brand voice across many different countries.

Pricing and regional availability

The cost of these models varies significantly depending on the provider's pricing structure. Qwen-Audio-3.0-TTS-Plus is normalized to $19.31 per million characters for comparison with dollar-denominated vendors. Alibaba also offers a Flash tier and regional rates in Beijing and Singapore, including a one-million-token free allowance for 90 days in the Beijing region.

Simba 3.5 pricing is tied to specific plans and has undergone several revisions since its launch in October 2026. Depending on the plan, the rate for Simba 3.5 ranges from $6.58 to $15.00 per million characters. This makes Simba generally cheaper than the Qwen-Audio-3.0-TTS-Plus board rate, though neither model is considered expensive in absolute terms for typical monthly volumes.

For large-scale operations, the cost difference may be less important than the technical requirements of the deployment. A user running five million characters a month would spend under $100 on either platform. Therefore, the decision usually rests on whether the user needs the specific expressive tags of Qwen or the documented language coverage of Simba.

Model transitions and updates

Both models are currently in states of transition or have been recently updated. Qwen-Audio-3.0-TTS was released in July 2026 and has already been superseded by the Qwen-Audio-3.1-TTS-Plus model. The newer 3.1 version sits closer to Simba 3.5 in quality, but it is billed on a different unit of tokens rather than characters, making direct price comparison difficult.

Simba 3.5 has also seen significant changes to its integration paths. On November 21, 2026, the old simba-english and simba-multilingual IDs were merged into the Simba 3.5 ID. This means existing integrations will automatically switch to the new model and pricing without requiring any code changes from the developer. This shift ensures that the 3.5 model becomes the standard for all previous Simba versions.

Users looking to migrate from Alibaba's 3.0 tier to the 3.1 tier may face a more complex project than a simple model ID swap. Because the 3.1 line uses tokens, developers must re-calculate their budgets and adjust their logic. In contrast, the Simba transition is designed to be a seamless update for those already using the Speechify API.

Choosing the right tool

The final decision depends on which specific problem the developer is trying to solve. Simba 3.5 is the better choice for a product that needs to ship one brand voice across dozens of locales. It is also ideal for accessibility projects where a single cloned voice must follow a user from English into Spanish without changing the persona.

Qwen-Audio-3.0-TTS-Plus is the preferred choice for performance-heavy content like podcasts or Mandarin-first content. Because it supports non-verbal tags and has deep pronunciation controls, it is better for content where the delivery must be highly expressive. It is also the better option for deployments that must reside specifically in the Singapore or Beijing regions.

Ultimately, neither model can guarantee that the non-English output will be perfect for every audience. The Controlled Voice board focuses on English reference voices, so a lead in English does not automatically translate to high quality in Japanese or Polish. Developers should test both models on the specific languages that generate the most revenue to see which one performs best in practice.

What to watch

Try it

Render the same three-minute script through both Simba 3.5 and Qwen-Audio-3.0-TTS-Plus to hear the difference between a prompt-based delivery and a list-based one.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by OrcaRouter · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news