Together AI has added Qwen3 TTS to its serverless inference platform and dedicated endpoints. The open text-to-speech model separates voice identity from delivery, so developers can keep one preset speaker while changing tone or emotion with a plain-language instruction.
The service provides nine preset voices, and each can speak Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. Qwen recommends choosing a voice whose native language matches the main use case, however, and developers must supply separately translated text rather than relying on the speech model to translate it.
Together says the first audio can arrive in as little as 97 milliseconds. HTTP streaming is available when the complete text is known, while a WebSocket interface accepts text in pieces as a language model generates it. Output options include WAV, MP3, raw PCM and mu-law audio; word-level timestamps are not supported.
Pricing is $20 per million characters on both serverless and dedicated deployments. The release is useful for multilingual assistants and notifications that need a consistent voice, but teams should test pronunciation and delivery in every target language. The latency figure is also a best-case vendor measurement, not a guarantee for every request or region.