Source: www.together.ai/blog/qwen3-tts
Qwen3 TTS separates who is speaking from how they speak. Developers choose one of nine preset voices and then shape its tone with a short instruction, such as "excited and upbeat." Every preset voice speaks Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian.
Together AI says its inference team tuned the serving stack for real-time conversation, and first audio can arrive in as little as 97ms. Audio can be streamed over HTTP by setting stream to true and response_format to raw, or sent over a WebSocket that accepts text in pieces as a language model produces it.
Output is available as wav, mp3, raw or mulaw at a sample rate of 24000 or 8000 Hz. Opus, aac and flac are not supported, and word timestamps are not available. VoiceFinder lets developers preview each voice and generate up to 500 characters of their own text before choosing one.
Why it matters if you build agents
Voice agents can keep one consistent voice while changing tone and language for new markets, without picking a different voice per locale. Check that the instructions field is spelled exactly, because a misspelled field is skipped without an error. Plan playback around the supported formats, since opus, aac and flac are not returned.
Get stories like this every weekday
One short email with the day's agent tooling news.