Skip to content
AI.info

The Pulse

Together AI adds Qwen3 TTS, citing first audio in as little as 97ms

Qwen3 TTS is available on Together AI's serverless inference service and dedicated endpoints. Together says streamed responses can produce first audio in as little as 97 milliseconds.

Together AI adds Qwen3 TTS, citing first audio in as little as 97ms

Developers can now run Qwen3 TTS on Together AI's serverless inference service or dedicated endpoints. Together says streamed responses can produce first audio in as little as 97 milliseconds.

The open text-to-speech model offers nine preset voices whose delivery can be changed with a plain-language instruction. The serverless service costs $20 per 1 million characters.

Voices and languages

Each preset supports Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. A developer can keep the same voice while changing its tone or the language it speaks.

Users must supply the text in the language selected; the multilingual examples in Together's post use separately translated text. Each voice also has a native language, and Qwen recommends using it in that language for the best quality.

The presets include two English voices, five Chinese voices, a Japanese voice and a Korean voice. Two of the Chinese voices are identified with particular dialects: dylan with Beijing and eric with Sichuan.

The voice, language and delivery instruction can be set in the same speech request. In the Python SDK, instructions are passed through extra_body; Together notes that a misspelled instructions field is skipped without an error.

Previewing the presets

VoiceFinder lists the Qwen3 TTS voices and lets users hear a sample of each or generate speech from their own text. Custom text can be up to 500 characters, and users can repeat the process with another voice to compare the same passage.

The tool also provides the request for a selected voice in Python, TypeScript or cURL. Its model menu identifies the offering as Qwen/Qwen3-TTS-12Hz-1.7B.

Streaming and output

For requests where the complete text is available, the speech API can stream audio over HTTP as server-sent events. Each audio-output event carries a base64-encoded audio chunk as it is generated.

The WebSocket endpoint accepts text in pieces, including text arriving from a language model. The server synthesizes each sentence once it has the whole sentence, and the input is committed when complete.

The connection carries the model, voice, language and audio settings in its URL. Audio comes back in the same base64 events used by the HTTP stream.

Qwen3 TTS supports WAV, MP3, raw PCM and mu-law output; Opus, AAC and FLAC are not supported. The default sample rate is 24000, with 8000 also available. Word timestamps are not supported for the model.

Sources

  1. Original source · Blog post Together AI together.ai

Companies in this story

Related stories

Explore

More articles