QwenCloud added three distinct audio products in October: Qwen Audio 3.1 TTS Plus on October 21, TTS Next on October 22, and Realtime Flash on October 27. The entries address premium speech synthesis, unified audio generation, and full-duplex conversation. Their interfaces and capabilities should be evaluated as separate products.
What the October dates actually establish
The exact model IDs are qwen-audio-3.1-tts-plus, qwen-audio-3.1-tts-next and qwen-audio-3.1-realtime-flash. Alibaba Cloud’s Singapore release table independently corroborates the October 21, 22 and 27 dates. However, its Beijing table dates TTS Next to September 21. October 22 therefore should not be described as that model’s first worldwide announcement; it is the dated QwenCloud entry and Singapore availability milestone.
These are documented hosted-service releases with API integration instructions, not merely previews of a possible future capability. The cited release records do not label these entries as previews or establish a separate general-availability certification. Nor do they announce downloadable weights for these exact models. Treat API access, preview status and a weight release as separate claims; the verified news here concerns the hosted services.
TTS Plus: the voice layer, not the whole conversation
The October 21 TTS Plus entry emphasizes professional speech quality, multilingual and dialect support, streaming synthesis, natural-language direction and tags for emotion, persona, pace and volume. Qwen also claims improved cloning robustness to noisy or reverberant reference recordings; that is a vendor claim, not a result from our own testing.
The synthesis overview lists WebSocket and HTTP access for TTS Plus and positions it for brand voices, audiobooks and dubbing. Streaming speech output is not the same as understanding a caller or managing a conversation. Our operational reading: Plus fits an approved-script narration workflow or the speech-output component of a larger assistant. Keep script approval and pronunciation checks separate from voice selection.
TTS Next: generate the scene, not just the speaker
Despite its TTS name, Next is an AudioGen model: it combines speech, sound effects and ambience into a generated scene. The guide explicitly distinguishes this from converting a fixed script into speech. Its documented limits are narrower than the broad multilingual launch language: the supported-model table lists Chinese and English, a 3,000-character input ceiling, and maximum outputs of four minutes for podcasts or two minutes for other scenarios.
Reference conditioning accepts up to three clips, each at most 30 seconds and 10 MB, in WAV, MP3 or OGG Opus—not raw PCM. Results can vary for identical prompts. For a marketing or training team, our inference is that this could consolidate scene prototyping, but it does not remove editorial review: check spoken wording, speaker identity and whether generated background sounds obscure the message.
Realtime Flash: conversation introduces different constraints
The October 27 Realtime Flash release supports text and audio input/output, voice cloning, function calling and web search, with a listed 262,144-token context window. Its guide describes full-duplex streaming through WebSocket, WebRTC or AOQ, but web search and function calling cannot be enabled simultaneously. Audio history also has a separate retention ceiling of 50 turns or 300 cumulative seconds, with older history discarded.
For a booking assistant, our practical interpretation is to choose the tool-enabled session around scheduling operations rather than assume search and transactional tools coexist. Store confirmed booking state outside audio history. Interruption handling also belongs in the client: Qwen instructs developers to clear queued playback when new user speech starts.
What businesses should verify next
The regional record matters: Singapore lists all three October milestones, while Realtime Flash’s guide specifies Singapore or Beijing credentials. Do not infer US-region deployment from international availability. Our recommended next step is a narrowly scoped evaluation: narration approval for Plus, scene-level review for Next, or interrupted tool-driven conversations for Realtime Flash. The release news supports three distinct adoption paths—not a blanket claim that one new voice model replaces an entire production workflow.