Gemini 3.8 TTS vs Qwen Audio 3.0 TTS: Two Different Answers to "Make It Sound Right"

One Elo point separates these two models on the Artificial Analysis Provider Voice Arena, and they get there by solving completely different problems. Qwen-Audio-3.0-TTS-Plus is at rank 3 with an Elo of 1,259 across 1,447 samples and 15 arena voices, released September 2026 and listed at $27.60 per million characters. Gemini 3.8 Flash TTS is at rank 2 with 1,260 across 1,999 samples and 8 arena voices, released September 23, 2026 and listed at $33.00 per million characters. Statistically indistinguishable, twenty percent apart on price, and built on opposite theories of what makes synthetic speech good.

The theory is the interesting part. Qwen's answer is inline control: 86 delivery tags you sprinkle through the text so the model knows where to breathe, stress and shift register. Google's answer is a direction layer: line-by-line instructions, two-speaker staging, and a small set of non-verbal cues. If you have already built a pipeline that emits SSML-like annotations, one of these will fit and the other will not, and that compatibility matters more than a single Elo point.

Where the two actually sit on the board

Reading the Provider Voice Arena honestly requires looking at the intervals, not the ranks.

• Qwen-Audio-3.0-TTS-Plus — rank 3, Elo 1,259, 1,447 samples, 15 arena voices, September 2026, $27.6 per 1M characters.

• Gemini 3.8 Flash TTS — rank 2, Elo 1,260, 1,999 samples, 8 arena voices, September 2026, $33.0 per 1M characters.

• Cartesia Sonic 3.6 — rank 1, Elo 1,273, 1,757 samples, 8 arena voices, August 2026, $49.0 per 1M characters.

The top three are one band. The published intervals on the top two run seventeen points either side, which is wider than the entire spread between first and third, so the ordering among them is not stable evidence. What the board does establish is that all three are in the leading group and that nothing below them is close — rank 10, Gemini 3.1 Flash TTS, is at 1,199.

The one asymmetry worth noting is the arena voice count. Qwen fields 15 voices to Gemini's 8, so the Qwen score is an average over a wider sample of the vendor's own product. That cuts both ways: it is a fairer estimate of what Qwen's catalogue sounds like overall, and it gives Qwen more opportunities to field a weak voice that drags the average down. Neither reading changes the conclusion that the two are tied.

86 delivery tags against a direction layer

This is the substantive difference and it decides most integrations.

Qwen-Audio-3.0-TTS ships 86 inline delivery tags — annotations embedded in the text that control how a span is spoken. It is a markup approach: your script carries the performance. That is familiar to anyone who has worked with SSML, and it composes well with tooling that generates text programmatically, because the control surface is a string transformation rather than an API call.

Gemini 3.8 Flash TTS takes the other route. You direct a line the way you would direct an actor — pace, emphasis, emotional register — as a separate instruction rather than as inline markup. Two-speaker scenes can be staged inside a single call. And a small set of non-verbal cues is written into the script: <laughs>, <sigh>, <gasp> as inline cues, plus |mhm| and |yeah| for back-channel noises.

So both models accept in-text annotation; the difference is the size of the vocabulary and where the rest of the control lives. Qwen's is broader and flatter — 86 tags, all in the text. Google's is narrower in the text and wider outside it, with delivery direction and speaker staging as call-level parameters. If you are porting a system that already emits rich inline markup, Qwen is the shorter port. If you are building the direction logic in application code anyway, Google's split is cleaner.

Language coverage versus dialect coverage

The coverage numbers are not comparable, and comparing them naively is a mistake.

Gemini 3.8 Flash TTS covers 130 languages, auto-detected from the text; Flash-Lite TTS covers 101. That is breadth across language families, which is what you want for a localisation pass that has to reach a long tail of markets.

Qwen-Audio-3.0-TTS covers 16 languages plus 20 Chinese dialect regions. That is depth inside one language family. Twenty dialect regions is not a smaller number than 130 languages in the way it looks — it is a different kind of claim, and for a product serving Chinese-speaking users across mainland regions it is the more useful one. A model with 130 languages and one Mandarin voice does not serve a user in Sichuan the way a model with 20 dialect regions does.

Which matters depends entirely on your audience. If your requirement is "at least passable in twenty markets", Gemini. If it is "actually right in the Chinese market", Qwen.

Price, and what the per-character figure hides

• Qwen-Audio-3.0-TTS-Plus — $27.60 per 1M characters on the leaderboard's normalised basis; the Flash variant is around $15 per 1M characters with a claimed first-packet latency of about 300 ms.

• Gemini 3.8 Flash TTS — $33.00 per 1M characters normalised; billed by Google at $0.50 per million text tokens in and $9.00 per million audio tokens out through December 31, 2026, then $1.00 and $18.00.

• Gemini 3.8 Flash-Lite TTS — $22.10 per 1M characters normalised at rank 6 with an Elo of 1,235; billed at $6.00 per million audio tokens through December 31, 2026.

Both per-character figures are Artificial Analysis normalisations, not vendor list prices — Qwen bills through Alibaba Cloud Model Studio and Google bills in tokens, and neither publishes a per-character rate for these models. Use them for the ratio, which is roughly 1.2x in Qwen's favour, and not as quotes.

The tiers diverge in shape. Qwen splits into Plus and Flash, with the Flash tier trading quality for a ~300 ms first-packet claim. Google splits into Flash and Flash-Lite, and then further into Standard, Batch and Flex, and Priority — $4.50, $9.00 and $16.20 per million audio tokens on Flash TTS. Google's model rewards classifying your workload; Qwen's rewards choosing a quality tier up front.

Availability is the other real difference. Gemini 3.8 Flash TTS is generally available on the Gemini Developer API. Qwen-Audio-3.0-TTS is closed and hosted-only, served through Alibaba Cloud Model Studio with no open weights. If you need to run the model yourself, neither of these is for you — but if your infrastructure is already on Alibaba Cloud, the Qwen model is the one with no egress story to work out.

Vendor claims on both sides

Neither model has an independent benchmark beyond the arena, and both vendors have published claims that should be labelled as such.

Google's, for Gemini 3.8 Flash TTS: first on the Hume AI Voice Design Benchmark at 71.4, first on accent modelling at 60.8, first and second on the Hume Overall Quality Index for Flash and Flash-Lite, and top blind-preference placings claimed in Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi. The runs are not public.

Alibaba's, for Qwen-Audio-3.0-TTS: the 86-tag delivery system, the 20 dialect regions, and the ~300 ms first-packet figure on the Flash variant. Also unreproduced.

The arena Elo is the only independently collected quality number in this article, and it says the two are tied at the top of the board.

What Gemini adds that Qwen does not

Voice creation is the clearest gap. Gemini 3.8 Flash TTS supports voice design from a natural-language description and voice replication from a 30-second sample, with consent verification and a SynthID watermark plus C2PA credentials on the output. Custom voices come in two forms: 200 stateful voices per project with a one-year time to live, or stateless voices addressed by a voicekey_... handle that expire after seven days. Voice remixing is listed as coming soon.

Output format control is the second. WAV 24 kHz mono 16-bit signed PCM on the unary endpoint, headerless PCM (audio/l16) on the streaming endpoint, plus audio/mulaw and audio/alaw and a configurable sample rate. For telephony-adjacent work, mulaw support removes a transcoding step.

Which one to call

Pick Qwen-Audio-3.0-TTS if your users are Chinese-speaking and dialect coverage is the requirement, if you want a rich inline markup vocabulary your generator can emit directly, or if you are already on Alibaba Cloud and want the shortest path to a working endpoint. It is also the cheaper of the two on the normalised basis, and the Flash variant is cheaper again if ~300 ms first packet is acceptable.

Pick Gemini 3.8 Flash TTS if you need breadth across many languages rather than depth in one, if you need to create or replicate a specific voice, if you need mulaw or alaw output, or if "generally available rather than hosted-only" is a procurement requirement.

Neither is a bad choice and neither is clearly better — which is the point of a one-Elo-point gap. The decision is compatibility, not quality.

That is also the argument for not committing hard to either yet. OrcaRouter gives you 200+ models behind one key at provider list price with no markup, so a rate change from either lab reaches your bill the same day rather than at renewal, and automatic failover means a synthesis endpoint that falters under load falls through to another instead of dropping the request. We do not host Qwen-Audio-3.0-TTS — that is served through Alibaba Cloud Model Studio — and we do not host the Gemini 3.8 TTS endpoints, which come from Google's API. What we carry is the text layer and the routing above it, which is what lets you run both on real traffic for a month before you pick.

The number to watch is the next arena revision. With 1,447 samples on Qwen and 1,999 on Gemini, both intervals are still wide enough that a single month of votes could separate them — or confirm that the tie is the answer.