Qwen: Qwen-Audio-3.0-TTS Plus

Qwen-Audio-3.0-TTS-Plus converts written text into expressive speech, with the Plus edition focused on synthesis quality, naturalness and detailed vocal delivery. Qwen names content creation, audiobooks, film and video dubbing and professional speech production as its main applications. The model’s updated profile describes broader low-resource-language and Chinese-dialect support, improved dialect authenticity, and stronger control of emotion, tone, character, speaking rate, volume and style. The producer also reports improved clarity, resolution and robustness under noisy or reverberant acoustic conditions.

Plus is the quality-focused edition of this text-to-speech family; Flash is separately positioned for real-time interaction. Prepare a narration as a spoken script, with the intended audience, pace and emotional direction identified. For an audiobook passage, an illustrative brief might use clear articulation, a measured pace and restrained warmth; for a video voice-over, it might ask for a serious opening followed by a more energetic explanation. These are delivery choices for supplied words. The model is a speech synthesiser rather than a transcription service or a generator of a complete visual scene.

The official system voices for Plus include Long An Ling Xin, described as warm and empathetic, and Long An Lu Feng, described as bright and cheerful. Both are listed for Mandarin and English. The producer also supplies more than 500 base voices for this edition, with previews to assess before choosing one. It recommends system voices for more consistent synthesis. Select a voice from the exact Plus list and match the script language to that voice’s supported languages; a voice from Flash is not interchangeable merely because it belongs to the same broader family.

The model supports sample-based voice cloning. Its documented cloned-voice languages include Chinese, English, Japanese, Korean, Russian, French, German, Portuguese, Thai, Indonesian, Vietnamese, Spanish, Italian, Malay, Filipino and Arabic. Chinese variants include Mandarin, Cantonese, Chongqing, Northeastern, Gansu, Guizhou, Zhejiang, Hebei, Henan, Hubei, Hunan, Jiangxi, Ningbo, Ningxia, Qingdao, Shaanxi, Shanxi, Shandong, Shanghainese, Sichuan and Yunnan, with dialects controlled through instructions. The producer’s model table distinguishes cloning from voice creation from a text description: cloning is supported for Plus, while that separate voice-design mode is not.

A suitable cloning reference is clear speech at normal speed, without music, background noise, competing speakers or singing. The producer recommends 10–20 seconds and permits up to 60 seconds, in a file no larger than 10 MB. Reference formats are 16-bit WAV, MP3 and M4A, with a minimum 16 kHz sample rate. At least five seconds of continuous clear speech is required, and remaining pauses should be no longer than two seconds. For stereo, only the first channel is processed. The created voice identifier is bound to Plus and used for synthesis when ready; reproducing the same reference through Flash requires a separately created Flash voice.

Natural-language instructions can shape the tone, pace, emotion and timbre of both system and cloned voices. Use concrete audible descriptions instead of a vague request for a good voice. Inline emotion tags can set a style within the script, and rich-language tags can insert a vocal effect at a chosen point. For example, [serious] changes the following delivery, while [sighing] inserts a sigh without changing the surrounding emotional style. Control tags apply until the next control tag or automatic sentence segmentation. This tag-based feature is documented for unidirectional streaming only, so it has a different scope from instructions used with bidirectional text input.

The producer’s output options are PCM, WAV, MP3 and Opus, at listed sample rates of 8,000, 16,000, 22,050, 24,000, 44,100 and 48,000 Hz. MP3 and 22,050 Hz are the documented defaults. Volume can be set from 0 to 100, and rate and pitch from 0.5 to 2.0; MP3 and Opus additionally allow a bitrate from 6 to 510 kbps. Choose the representation needed for playback or editing, then keep those settings consistent across the sections of a longer piece. Streaming output includes audio frames and sentence metadata, with word timestamps available for eligible voices, including cloned voices.

The documented WebSocket text-input flow allows 20,000 characters per message and 200,000 characters cumulatively, and requires a submission interval no greater than 23 seconds. It can receive a complete script or ordered text segments and finish when the input is complete. The producer also documents cancelling the current synthesis turn. These are limits and behaviours of that protocol, not a universal generated-duration cap. For a long narration, organise the script into coherent sections and retain their order and text alongside the resulting audio so the piece can be assembled and reviewed consistently.

For pronunciation, the producer provides target-language hints, pinyin annotations and text replacements to address unwanted readings of digits, abbreviations, symbols or words. SSML is a separate markup mode with eligibility constraints: the Qwen documentation supports it for cloned voices in non-streaming or unidirectional-streaming calls, with only one text submission when enabled. It does not establish SSML for every system voice or for bidirectional streaming. A practical Plus workflow starts with a compatible voice and a finished script, then adds precise delivery and pronunciation choices for the passage being narrated.