Qwen /

Qwen3-TTS-12Hz-1.7B-CustomVoice

A 1.7B text-to-speech model with 9 preset voices and style control, in 10 languages.

2.2k likes
1.9M downloads a month
1.9B parameters
Some experience
Speech

Open on Hugging Face (opens in a new tab)

This is the CustomVoice version of Qwen3-TTS, a text-to-speech family from the Qwen team. It ships with 9 premium voices that differ by gender, age, language and dialect (for example Vivian, Ryan, Sohee and Ono_Anna). You pick a voice, give it text, and optionally add an instruction such as 'Very happy.' to change how it sounds. The card says each voice works best in its native language but can speak any of the 10 supported languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian.

What sets it apart, according to the card, is a streaming design that can output the first audio packet after a single character and reaches end-to-end latency as low as 97ms. Honest limits: you are limited to the 9 preset voices. If you want to invent a voice from a description, you need the VoiceDesign sibling, and if you want to copy a voice from a short clip, you need the Base model. The card's examples all use a CUDA GPU with FlashAttention 2, which needs compatible hardware, so a laptop without a GPU may be a rough ride. The card text I was given is cut off before the benchmark results, so I can't tell you how it scores.

What you need to run it

1.9B parameters. The weights take about 4 GB at full 16-bit precision (fits a 8 GB Mac or a GPU with 8 GB), or about 2 GB as a 4-bit download, which runs on a 8 GB Mac or a GPU with 8 GB. Long inputs and images need more memory on top.

What it does well

9 ready-made voices
Voices cover different genders, ages, languages and dialects, including Beijing and Sichuan Chinese, English, Japanese and Korean speakers.
Talk to it in plain language
An optional instruct field lets you ask for things like an angry tone or a very happy delivery, and the model also adapts tone and rhythm to the meaning of the text.
10 languages
Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian are supported, and you can pass Auto for language or omit it.
Low-latency streaming
One model handles both streaming and non-streaming generation, with latency as low as 97ms per the card, which suits real-time apps.
Easy to try
A pip package, a local web demo command, batch generation and vLLM-Omni offline examples are all documented. The Apache 2.0 licence keeps things simple.

What you could build

Build a voice reader that narrates articles or study notes in a chosen voice and mood.
Give a game character or chatbot a spoken voice that changes emotion with the scene.
Make a multilingual demo that reads the same sentence in several languages with different speakers.
Prototype a real-time voice assistant that uses streaming output to start talking quickly.
Generate audio for short videos or language-learning clips with native-sounding voices.

Get started

Create a fresh Python 3.12 conda environment, install the qwen-tts package, then load Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice and call generate_custom_voice with a language, a speaker and optional instruct text. If you'd rather click than code, the package also includes a local web demo.

conda create -n qwen3-tts python=3.12 -y

conda activate qwen3-tts

pip install -U qwen-tts

qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --ip 0.0.0.0 --port 8000

Commands are copied from the model card. Read it on Hugging Face for the full setup.