Vibeleaderboard
← All Intel
Intel / article

Qwen3 TTS is now available on Together AI

Source
together.ai
Date

Context

Together AI, a cloud service that runs open models for developers, now hosts Qwen3 TTS, an open text-to-speech model from Qwen, Alibaba's model team (announcement). Qwen published the weights on Hugging Face under the Apache 2.0 license, so the news here is a hosted way to call the model rather than a new model. Together charges $20 per million characters on its serverless service, and the model is also available on dedicated endpoints.

The model's main feature is that it keeps who is speaking separate from how they speak. A developer picks one of nine preset voices, such as `aiden` or `serena`, and can then change the delivery with a short plain-language instruction like "excited and upbeat" or "sad and disappointed" without changing the voice. Each voice speaks all 10 supported languages, among them Chinese, English, Japanese, Spanish and German, so a product can keep one voice across markets. The text has to arrive already translated, though, and each voice has a native language that Qwen recommends for the best quality.

Together aims the hosted model at voice agents, where speech must start before the full reply exists. It says its tuned serving stack returns the first audio in as little as 97 milliseconds, the same best-case figure Qwen's model card gives for the model, and neither source reports typical latency. Audio streams in two ways. When the whole text is known, an HTTP request streams the audio back in chunks. When the text is still coming from a by token, a WebSocket connection lets the app send it in pieces, and the server speaks each sentence once that sentence is complete, so the can start talking before the model has finished writing.

Some limits matter when choosing it. Together's version does not return word timestamps, so an app that highlights words as they are spoken cannot get that timing from this model. Output comes as WAV, MP3, raw PCM or mu-law audio, not Opus, AAC or FLAC. In Together's Python SDK the delivery instruction travels in an `instructions` field passed through `extra_body`, and Together warns that a misspelled field is dropped without an error, so a voice that ignores its instruction is worth checking for a typo first.

Key takeaways · AI-distilled
  • Qwen3 TTS separates who speaks from how they speak: you pick one of nine preset voices, then change emotion and tone with a plain-language instruction such as "excited and upbeat" while the voice identity stays the same.
  • Every preset voice can speak all 10 supported languages, so a product can keep one voice across markets, but Qwen recommends using each voice in its native language for best quality and you must supply text already translated.
  • Together offers two paths: HTTP server-sent events when the full text is known, and a WebSocket that accepts text in pieces as an LLM writes it, synthesizing each sentence once complete. Together cites first audio in as little as 97ms.
  • Practical limits from Together's docs: no word timestamps, no opus/aac/flac output, and in the Python SDK the instructions field goes through extra_body, where a misspelled field is silently ignored.
Terms in this piece · Glossary
  • LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
  • token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
Read the source www.together.ai
Recommended reads
Comments

Checking sign-in…

Loading comments…