Qwen: Qwen-Audio-3.0-TTS Flash

Qwen-Audio-3.0-TTS-Flash converts a supplied script into speech, with the Flash edition focused on real-time interaction. Qwen identifies voice assistants, interactive dialogue and intelligent customer service as its main applications. Compared with the preceding version, the producer describes broader low-resource-language and Chinese-dialect support, more authentic dialect expression and stronger control of emotion, tone, character, speed, volume and speaking style. It also reports improved clarity and robustness in noisy or reverberant acoustic conditions. This is text-to-audio synthesis, rather than recognition of incoming speech or a voice-dialogue model.

Flash prioritises the start of speech in an interactive exchange. Qwen reports first-packet latency within 200 milliseconds for this model’s real-time synthesis; its usage guide also notes that the initial connection setup contributes to the first submission’s measured delay. Prepare the actual words to be spoken separately from the desired delivery. For a voice-assistant response, an illustrative brief could combine a concise answer with a calm, clear delivery instruction. The relevant result is the generated speech for that text, with the application handling playback and the surrounding conversation.

The producer’s system-voice list includes model-specific options such as Long An Feng Yue for a natural, friendly delivery and Long An Xiao Xin for a friendly, lively delivery, both listed for Mandarin and English. It also includes English voices such as loongmary with a British accent and loongjohn with a calm, friendly American accent. Each model additionally has more than 500 base voices. Qwen recommends starting with system voices for consistency and assessing previews when selecting a base voice. A voice must belong to the selected model, and the script must use a language that voice supports.

Cloned voices broaden the documented language range to Chinese, English, Japanese, Korean, Russian, French, German, Portuguese, Thai, Indonesian, Vietnamese, Spanish, Italian, Malay, Filipino and Arabic. The Chinese coverage includes Mandarin, Cantonese, Chongqing, Northeastern, Gansu, Guizhou, Zhejiang, Hebei, Henan, Hubei, Hunan, Jiangxi, Ningbo, Ningxia, Qingdao, Shaanxi, Shanxi, Shandong, Shanghainese, Sichuan and Yunnan varieties, with dialect selection through instructions. The model table supports sample-based voice cloning and distinguishes it from creating a new voice solely from a text description, which this edition does not support.

For cloning, the producer recommends a 10–20-second recording, with a 60-second maximum and a file size no greater than 10 MB. The supported reference formats are 16-bit WAV, MP3 and M4A, at a sample rate of at least 16 kHz. The recording needs at least five seconds of continuous clear speech; other pauses should be no longer than two seconds. Avoid background sound, other voices and singing. Stereo is accepted, but only the first channel is processed. The cloning workflow returns a voice identifier bound to the exact target model, and synthesis uses that identifier once the voice is ready. The same recording requires separate cloned voices for Flash and Plus.

Natural-language instructions control delivery for both system and cloned voices. Describe audible qualities such as a gentle tone, clear articulation or a slightly slower pace. Inline tags provide a different control: [empathetic] or [serious] sets a style for subsequent text, while [laughing] or [sighing] inserts a vocal effect at a particular position. A control tag lasts until another control tag or automatic sentence segmentation. These emotion and rich-language tags are documented only for unidirectional streaming, rather than the bidirectional text-input mode. Prepare an expressive line in the mode appropriate to those controls.

The producer’s speech interface supports PCM, WAV, MP3 and Opus. Its listed sample rates are 8,000, 16,000, 22,050, 24,000, 44,100 and 48,000 Hz, with MP3 and 22,050 Hz as defaults. Volume ranges from 0 to 100; speech rate and pitch each range from 0.5 to 2.0. MP3 and Opus also expose a bitrate setting from 6 to 510 kbps. Select a format and sample rate to match the intended player or processing workflow. For an interactive application, the producer documents streaming audio frames and sentence metadata, with word-level timestamps available in streaming output for eligible voices, including cloned voices.

The documented WebSocket text protocol accepts up to 20,000 characters in one text message and 200,000 characters cumulatively in a task, with no more than 23 seconds between submissions. These are producer protocol limits, separate from a universal model output-duration limit. It supports cancellation of the current synthesis turn and a subsequent task on the same connection. Keep text segments in their intended spoken order and end the input when the script is complete. For pronunciation preparation, the producer documents target-language hints, pinyin annotations and text replacements, useful when digits, abbreviations or symbols are spoken unexpectedly.

SSML is a separate markup option with voice and interaction-mode conditions: Qwen documents it for cloned voices, with non-streaming or unidirectional-streaming calls, and a single text submission when SSML is enabled. It should not be assumed for every preset voice or bidirectional stream. For ordinary dialogue, begin with a compatible voice, a clean spoken script and a specific delivery instruction; add targeted pronunciation or expressive controls where the task requires them. This keeps the Flash profile centred on responsive speech from prepared text rather than on capabilities of other Qwen audio models.