Answer first: use Qwen-Audio-3.0-Realtime when the product must listen, reason, interrupt, remember, or call tools during a live conversation. Use Qwen-Audio-3.0-TTS when the text is already known and the job is to produce controlled, multilingual speech.
The two products solve different problems
| Decision factor | Qwen-Audio-3.0-Realtime | Qwen-Audio-3.0-TTS |
|---|---|---|
| Input | Live audio and conversational context | Prepared text and delivery direction |
| Output shape | Two-way, interruptible dialogue | Generated speech asset or stream |
| Agent tools | Provider documents knowledge-base, MCP, and OpenAPI connections | Not the core product role |
| Memory | Context-aware dialogue | Script and generation context |
| Provider examples | Customer service, education, entertainment, companionship | Video voiceovers, audiobooks, dubbing |
| Key risk | Wrong action or poor recovery during a live interaction | Mispronunciation, artifacts, rights, or unusable delivery |
The shared Qwen-Audio name does not make the products interchangeable.
Choose Realtime for a live agent
Use Qwen-Audio-3.0-Realtime when at least one of these is essential:
- the user can interrupt;
- the system must detect turns in noise;
- tone changes during the conversation;
- the next response depends on session memory;
- the model calls a tool or retrieves current information;
- a human handoff may happen mid-session.
Define tool permissions before connecting a live user. Read-only retrieval is a safer first pilot than account updates, purchases, or irreversible operations.
Choose TTS for a produced asset
Use Qwen-Audio-3.0-TTS when:
- an approved script already exists;
- editors need reproducible takes;
- language, dialect, pacing, tone, or emotional tags shape delivery;
- the output enters a video, audiobook, lesson, advertisement, or announcement workflow;
- a human can review the complete asset before publication.
TTS Flash and Plus are tier choices inside this product. Use the Flash vs Plus language guide for a matched test.
Use both only with a clear boundary
A service can use Realtime for a live support session and TTS for a prepared queue message or follow-up asset. Keep their state and controls separate:
- Do not let a generated asset inherit live-session data without authorization.
- Do not let the live agent speak a TTS asset that has not passed the required content policy.
- Record which product produced each output.
- Use separate fallback paths for connection failure and audio-quality failure.
Compare with two scorecards
For Realtime, measure time to first audio, turn detection, interruption success, task completion, tool accuracy, handoff quality, and recovery.
For TTS, measure pronunciation, language accuracy, style adherence, artifacts, generation time, edit minutes, and accepted-output rate.
Both require policies for consent, disclosure, recordings, scripts, voices, likenesses, retention, and commercial distribution. Capability is not permission.
FAQ
What is the difference between Qwen-Audio-3.0-Realtime and TTS?
Realtime is for live two-way conversation with interruptions, memory, and tool connections. TTS converts prepared text into controlled multilingual speech for voiceovers, audiobooks, and other produced audio.
Which model should I use for a customer-service agent?
Start with Realtime when the system must listen, respond, handle interruptions, retain conversational context, or call tools. Use TTS only when another system already determines the exact response text and live dialogue behavior is unnecessary.
Which model should I use for video voiceovers?
Start with TTS. Alibaba positions its Flash and Plus variants for generated speech, including voiceovers and audiobooks, with language and expressive controls.
Can one application use both?
Yes. A product can use Realtime for live sessions and TTS for prepared announcements or edited assets, but each path needs separate latency, quality, consent, rights, and failure controls.
Official source
Source checked July 29, 2026: