Models / Qwen
Audio

Qwen3 TTS 12Hz 1.7B

Multilingual, low-latency speech with instruction-controlled tone and emotion

About model

Qwen3 TTS 12Hz 1.7B has 1.7B parameters. On Together AI it costs $20.00 per 1M characters.

Qwen3 TTS 12Hz 1.7B is a text-to-speech model from Qwen covering 10 major languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian) along with dialect voice profiles. It understands context, adjusting tone, speaking rate, and emotional expression from natural-language instructions and the meaning of the text itself, and holds up well on noisy input text. It ships with 9 preset voices spanning different genders, ages, languages, and dialects, each able to speak any supported language. Built on Qwen's own 12Hz speech tokenizer and a single end-to-end language model architecture, it supports streaming output, with first audio in as little as 97ms. Released under Apache 2.0. Available on Together AI.

Aiden
0:00

"Thanks for calling Northside Dental. I have you down for a cleaning with Doctor Patel on Thursday, March twelfth, at two thirty in the afternoon. We'll text a reminder to the number ending in four four one seven the day before. If you need to reschedule, just reply to that message. Is there anything else I can help you with today?"

Ryan
0:00

"Okay, you are not going to believe what just happened. I was halfway out the door, keys in one hand, coffee in the other, when my phone buzzed. It was the bakery. They said the cake was ready. Great, except I never ordered a cake. So now I'm standing in my hallway, late for work, wondering who is getting married, and whether I'm supposed to be there."

Uncle Fu (Mandarin)
0:00

"小时候,爷爷常常在院子里给我们讲故事。夏天的晚上,风很凉,星星很多。他总是说,慢慢来,好故事要一句一句地听。"

Languages and Voices

10

Languages across 9 preset voices, plus Beijing and Sichuan dialects

Voice Control

Plain language

Steer delivery with an instruction, like 'use a compassionate voice'

First-Audio Latency

97ms

Time to first audio, streaming from the first character of input

Model key capabilities
  • Instruction-Controlled Delivery: Tell it how to speak in plain language, for example "use a compassionate voice", and it adjusts tone, speaking rate, and emotion to match
  • Multilingual Voices: 10 major languages with dialect voice profiles, and every preset voice can speak any supported language
  • Low-Latency Streaming: One model handles both streaming and non-streaming output, built for real-time conversational use
  • Production-Ready Infrastructure: 99.9% SLA, available on serverless and dedicated infrastructure
Quickstart guides
  • API usage

    • cURL
    • Python
    • Typescript

    Endpoint:

    Qwen/Qwen3-TTS-12Hz-1.7B

    curl -X POST "https://api.together.ai/v1/audio/speech" \
         -H "Authorization: Bearer $TOGETHER_API_KEY" \
         -H "Content-Type: application/json" \
         -d '{
           "model": "Qwen/Qwen3-TTS-12Hz-1.7B",
           "input": "Today is a wonderful day to build something people love!",
           "voice": "aiden",
           "language": "en",
           "response_format": "mp3"
         }' \
         --output speech.mp3
    from together import Together
    
    client = Together()
    
    speech_file_path = "speech.mp3"
    
    with client.audio.speech.with_streaming_response.create(
        model="Qwen/Qwen3-TTS-12Hz-1.7B",
        input="Today is a wonderful day to build something people love!",
        voice="aiden",
        response_format="mp3",
    ) as response:
        response.stream_to_file(speech_file_path)
    import Together from 'together-ai';
    import { Readable } from 'stream';
    import { createWriteStream } from 'fs';
    
    const together = new Together();
    
    async function generateAudio() {
      const res = await together.audio.speech.create({
        input: 'Today is a wonderful day to build something people love!',
        voice: 'aiden',
        response_format: 'mp3',
        sample_rate: 24000,
        stream: false,
        model: 'Qwen/Qwen3-TTS-12Hz-1.7B',
      });
    
      if (res.body) {
        const nodeStream = Readable.from(res.body as ReadableStream);
        const fileStream = createWriteStream('./speech.mp3');
    
        nodeStream.pipe(fileStream);
      }
    }
    
    generateAudio();
  • Model card

    Architecture Overview:
    • 1.7B parameter text-to-speech model in the Qwen3-TTS family
    • Built on Qwen3-TTS-Tokenizer-12Hz, Qwen's own speech tokenizer, which compresses audio efficiently while keeping paralinguistic detail and acoustic environment features
    • End-to-end discrete multi-codebook language model architecture, avoiding the cascading errors of separate language-model and diffusion stages
    • Dual-Track hybrid streaming design: a single model supports both streaming and non-streaming generation

    Training Methodology:
    • Trained for contextual understanding, adapting delivery to the semantics of the input text
    • Supports generation driven by natural-language instructions controlling timbre, emotion, and prosody
    • Qwen reports markedly improved robustness to noisy input text

    Performance Characteristics:
    • First audio in as little as 97ms, with the first audio packet available after a single input character
    • Preset voices: Vivian, Serena, and Uncle_Fu (Chinese); Dylan (Beijing dialect); Eric (Sichuan dialect); Ryan and Aiden (English); Ono_Anna (Japanese); Sohee (Korean)
    • Each voice performs best in its native language and can speak any of the 10 supported languages

    ‍

  • Prompting

    Together AI API Access:
    • Access Qwen3 TTS 12Hz 1.7B via the Together AI speech API using the endpoint Qwen/Qwen3-TTS-12Hz-1.7B
    • Authenticate using your Together AI API key in request headers
    • Choose a preset voice with the voice parameter, for example voice="aiden"
    • For best quality, pair each voice with its native language
    • Available on Together AI serverless and dedicated infrastructure

    ‍

  • Applications & use cases

    Voice Agents & Assistants:
    • Give real-time assistants natural, low-latency speech
    • Match tone and emotion to the conversation through instructions
    • Serve users across 10 languages from one model

    Content Narration:
    • Narrate articles, audiobooks, and learning material
    • Voice video content with expressive delivery
    • Produce localized voiceover across languages and dialects

    Customer Experience:
    • Power multilingual support lines and IVR prompts
    • Keep a consistent brand voice across languages
    • Generate dynamic spoken responses at scale

    ‍

  • Released
    September 28, 2026
  • Quantization level
    BF16
  • External link
  • Category
    Audio