Qwen3 TTS Now Available on Together AI Serverless

Qwen3 TTS Now Available on Together AI Serverless

At a Glance

Item Value
Publisher Together AI
Published 2026-11-20
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Qwen’s open text-to-speech (TTS) model, “Qwen3 TTS" (Qwen/Qwen3-TTS-12Hz-1.7B), is now available on Together AI’s serverless inference and dedicated endpoints. This model generates speech from text inputs and is characterized by its ability to separately specify the speaker’s voice quality and instructions for speaking style (tone and emotion).

Through optimization of the serving stack on Together AI, it achieves fast responses, returning the first audio in as fast as 97ms during streaming generation.

Specifications

According to the materials, the main specifications are as follows:

  • Model Name: Qwen/Qwen3-TTS-12Hz-1.7B
  • Parameters: 1.7B (12Hz)
  • Input/Output Specifications: Outputs audio from text input (Text-to-Speech)
  • Preset Voices: 9 types in total (aiden, ryan, vivian, serena, uncle_fu, dylan, eric, ono_anna, sohee)
  • Supported Languages: 10 languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian)
  • Output Audio Formats: wav, mp3, raw (PCM), mulaw (Note: opus, aac, and flac are not supported)
  • Sampling Rate: 24000 Hz (default) or 8000 Hz
  • Unsupported Features: Word timestamps (alignment=word)
  • Pricing (Together AI): $20 per 1 million characters (Together Serverless Inference)
  • License Conditions: Apache 2.0 license (commercial and modification use conditions comply with Apache 2.0)

Performance and Quality

Qwen3 TTS can generate audio that separates the control of the speaker (who is speaking) and the speaking style (how they are speaking). Even with the same preset voice selected, by providing natural language instructions (instructions) such as “excited and upbeat", “angry and impatient", or “sad and disappointed", it is possible to change the emotion and expression tone without altering the speaker’s voice.

In addition, all 9 preset voices support reading in all 10 supported languages. This makes it possible to deploy speech across multiple languages while maintaining the same voice quality. On the other hand, each voice has a designated native language (for example, aiden is English, ono_anna is Japanese, etc.). Qwen, the releasing organization, recommends using each voice in its native language to achieve the best quality.

In terms of performance, tuning of the serving stack by Together AI’s inference team enables the first audio to start in as fast as 97ms during response streaming. This achieves high speed suitable for applications requiring immediate responsiveness, such as real-time voice agents.

Strengths and Use Cases

Qwen3 TTS is particularly suited for applications such as flexible expression control of emotions and tone, building consistent multilingual brand voices, and developing real-time voice agents.

Specifically, because speaking styles can be specified using short natural language texts, it is possible to differentiate expressions according to scenes and situations while using the same voice. For example, by giving instructions like “excited and upbeat", “angry and impatient", or “sad and disappointed", nuances can be flexibly changed while keeping the speaker’s voice fixed.

Furthermore, since all 9 preset voices support all 10 languages, applications expanding globally can generate audio guidance in various national languages while maintaining a single voice. Each preset voice has its own native language and characteristics set.

  • aiden (English): Bright and clear mid-range American English male voice
  • ryan (English): Dynamic male voice with a strong sense of rhythm
  • vivian (Chinese): Bright and slightly edgy young female voice
  • serena (Chinese): Warm and gentle young female voice
  • uncle_fu (Chinese): Experienced male voice with a low, mellow tone
  • dylan (Chinese – Beijing Dialect): Youthful Beijing male voice with a clear and natural tone
  • eric (Chinese – Sichuan Dialect): Energetic Chengdu male voice with a slightly husky brightness
  • ono_anna (Japanese): Charming Japanese female voice with a brisk and light tone
  • sohee (Korean): Warm Korean female voice with rich emotions

Qwen recommends using each voice in its native language to generate the highest quality audio. To make it easier to select voices tailored to specific use cases, “VoiceFinder" (voicefinder.together.ai) is available to check and compare voice qualities in a web browser. In VoiceFinder, users can select the model (Qwen/Qwen3-TTS-12Hz-1.7B), sample each voice, and test custom text inputs of up to 500 characters.

Moreover, with a very low latency of as fast as 97ms until the first audio is returned, it is extremely well-suited for real-time processing applications that require immediate responsiveness, such as conversational AI and voice agents.

How to Get It

Qwen3 TTS (Qwen/Qwen3-TTS-12Hz-1.7B) weights are available as an open-source model on Hugging Face under the Apache 2.0 license, and it can also be accessed via Together AI’s cloud service. On Together AI, it is provided as Together Serverless Inference and dedicated endpoints at $20 per 1 million characters.

When using Together AI’s cloud API, install the Together Python SDK from a Python environment using the pip install together command, and call it after setting the TOGETHER_API_KEY environment variable.

For standard audio generation, the API is called by specifying parameters such as the following:

  • model: "Qwen/Qwen3-TTS-12Hz-1.7B"
  • input: Text to be read aloud
  • voice: One of the 9 voice names (specified in lowercase)
  • language: ISO 639-1 language code for the input text (en, zh, ja, ko, etc.)
  • extra_body: Specify speaking style instructions such as {"instructions": "excited and upbeat"} (optional)
  • response_format: One of wav, mp3, raw (PCM), or mulaw (Note: opus, aac, and flac are not supported)
  • sample_rate: 24000 (default) or 8000

When using the Python SDK, speaking style instructions (instructions) are passed via the extra_body parameter. If the spelling differs, it will be ignored without throwing an error, so the field name instructions must be specified accurately.

For real-time voice agents, the following two streaming methods are supported:

  1. HTTP Streaming (When the complete text is available)
    Send a request specifying stream=True and response_format="raw". The server sequentially returns the generated audio data in Server-Sent Events format. Each conversation.item.audio_output.delta event contains Base64-encoded 16-bit PCM data (24kHz), which is sequentially decoded to play or save.

  2. WebSocket API (When text is generated sequentially token-by-token from an LLM)
    Connect to the endpoint wss://api.together.ai/v1/audio/speech/websocket using the websockets package (version 14 or later, pip install websockets). Send text using the input_text_buffer.append event each time text arrives, and send input_text_buffer.commit once input is complete. The server performs synthesis sentence by sentence and delivers Base64-encoded PCM audio data via events, similar to HTTP streaming.

Related Articles

Sources