Alibaba Cloud Model Studio Adds Four Qwen Audio 3.1 Models

Alibaba Cloud shipped four speech models in about two weeks, and one of them can generate a voiceover and the sound effects under it in a single pass. Qwen Audio 3.1 landed in Alibaba Cloud Model Studio (international, Singapore) between Oct. 12 and Oct. 27, 2026, per the Model Studio new models page. If your team makes voiceovers, localizes campaigns or runs a voice agent, this is worth an afternoon.

What shipped in Qwen Audio 3.1

| Model ID | Added | What the docs say | | --- | --- | --- | | qwen-audio-3.1-tts-flash | Oct. 12 | Streaming speech synthesis with system voices, for voice assistants and customer service | | qwen-audio-3.1-tts-plus | Oct. 21 | High-quality text-to-speech with multiple languages and dialects, streaming, free-style instruction following, and tags for emotion, tone, persona, speaking rate and volume. Voice cloning copes better with noise and reverberation in the reference audio | | qwen-audio-3.1-tts-next | Oct. 22 | An "AudioGen" model: text, timestamps and reference audio in, speech, sound effects and ambient sound out in one pass. Covers multi-speaker dialogue and podcasts | | qwen-audio-3.1-realtime-flash | Oct. 27 | A real-time duplex speech model with text and audio input and output, function calling, web search and voice cloning. Context window: 262,144 tokens |

Duplex means the model can listen while it talks, which is what lets a caller interrupt it mid-sentence.

Why Qwen Audio 3.1 matters for voiceovers and localization

The TTS overview points audiobooks, dubbing, brand voices and content production at tts-plus, and voice assistants and customer service at tts-flash, built for low latency. That split is your first decision.

For localization, the overview lists cloned-voice support for Mandarin, Cantonese, English, Japanese, Korean, Spanish, French, German and Arabic, among others, plus about 20 Chinese regional dialects steered through written instructions rather than separate voices.

tts-next is the one to try for video and podcast work: if speech and ambient sound come out of one call, you may skip a mixing step. That is my read, not a docs claim, so test it on a real spot.

realtime-flash is the voice agent option. Function calling means the agent can look up an order or book a slot mid-call, which is what separates a demo from a support line.

Check Qwen Audio 3.1 pricing before you pick a model

The four models are not billed the same way. The pricing page, updated Nov. 2, 2026, says: "qwen-audio-3.1-tts-flash is billed separately for input and output tokens. Other models are billed by the number of input characters, with no charge for output."

For qwen-audio-3.1-tts-next, the Singapore table lists USD 0.93 per million input tokens and USD 1.87 per million output tokens, with a free quota of 1 million tokens valid for 90 days. Tokens and characters do not compare at a glance, so run the same 60-second script through each candidate and compare what your usage report shows.

The docs keep separate Singapore and China (Beijing) price tables, and these four launches sit under the international scope, so check your region before you budget.

Try the realtime model from a terminal

Per the Realtime API WebSocket guide, the Singapore endpoint takes the model name as a query parameter and an API key as a bearer token:

export MODEL_STUDIO_API_KEY="<your-api-key>"
npx wscat -c "wss://<your-workspace-id>.ap-southeast-1.maas.aliyuncs.com/api-ws/v1/realtime?model=qwen-audio-3.1-realtime-flash" -H "Authorization: Bearer $MODEL_STUDIO_API_KEY"

This only proves your key, workspace ID and model name work; the session events to send next are in the same guide. Turn detection offers server_vad, smart_turn and smart_turn_v2 (the recommended one), plus push-to-talk when turn_detection is null.

What to do on Monday morning

1. Bake off two models. Run one real script through tts-plus and tts-flash, then compare quality and reported usage.
2. Test tags and dialects on live copy. Try the emotion, tone and speaking-rate controls on a script you ship, with a native reviewer.
3. Get cloning consent in writing. Before you upload reference audio, get the speaker's written consent covering languages, channels and duration, and read Alibaba's voice cloning guide for its own terms.
4. Mind the transport. The realtime models support WebSocket, WebRTC and AOQ (AI over QUIC), but real-time TTS lists WebRTC as unsupported, so a WebRTC browser demo will not work with the TTS models.

For anyone deploying AI in production, the useful part of this release is the spread: a cheap low-latency voice, a premium voice, an audio generator and a voice agent are four separate bills and four separate test plans.