Spark-TTS-0.5B

Spark-TTS-0.5B

Speaks Chinese and English from text with a Qwen2.5-based model, including zero-shot voice cloning and virtual voices set by gender, pitch and speaking rate. For developers and researchers working on speech.

Made by SparkAudio on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Spark-TTS-0.5B is a text-to-speech model built on Qwen2.5. It needs no separate generation model such as flow matching, because it rebuilds audio directly from the codes the language model predicts. It speaks Chinese and English, clones a voice without training on it, and can create virtual speakers by adjusting gender, pitch and speaking rate.

Who it suits: It suits developers and researchers who want a speech model for Chinese and English that runs on their own machine.

Industry | Who uses it | How they use it | Example
Software | Voice app developers | Add Chinese and English speech to an application | Reading app text aloud from a script
Media | Narration makers | Create a virtual speaker by setting gender, pitch and rate | Voicing a personal, non-commercial video
Localization | Bilingual teams | Switch between Chinese and English in cloned speech | Speaking a sentence that mixes both languages
Research | Speech researchers | Study speech generation driven by a language model | Running the demo script and listening to the output

If you need commercial use, or want to clone a voice you have no right to use, you can skip it. Use only voices you have permission for.

Where to use it:

• A terminal with a Conda environment on Python 3.12
• The web interface started by webui.py, with Voice Cloning and Voice Creation
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Voice cloning | Researchers | Speech in the voice of a short reference recording
Voice creation | Content makers | Speech from a virtual speaker with the gender, pitch and rate you set
Demo run | Newcomers | An audio file from the example script

How to set it up:

• Git and Conda
• Python 3.12 inside a Conda environment
• The Spark-TTS code from its GitHub page
• The model files, downloaded with Python or with Git LFS

1. Clone the Spark-TTS repository from GitHub and move into it. The card gives the clone command, which holds a web address, so copy it from there.
2. Create and activate the Conda environment.
3. Install the requirements. The card also lists a mirror option for users in mainland China.
4. Download the model into the pretrained_models folder.

conda create -n sparktts -y python=3.12
conda activate sparktts
pip install -r requirements.txt

from huggingface_hub import snapshot_download

snapshot_download("SparkAudio/Spark-TTS-0.5B", local_dir="pretrained_models/Spark-TTS-0.5B")

How to use it: Run the demo from the example folder, or call the command line interface yourself. You can also start the web interface, which lets you clone a voice from uploaded or recorded audio, or create a voice.

cd example
bash infer.sh

python -m cli.inference \
    --text "text to synthesis." \
    --device 0 \
    --save_dir "path/to/save/audio" \
    --model_dir pretrained_models/Spark-TTS-0.5B \
    --prompt_text "transcript of the prompt audio" \
    --prompt_speech_path "path/to/prompt_audio"

python webui.py

Tool | What it does
infer.sh | Runs the demo from the example folder
cli.inference | Synthesizes speech from text, with an optional prompt recording and its transcript
webui.py | Starts the web interface for Voice Cloning and Voice Creation
--save_dir | Sets the folder where the audio is saved

Example: You run the command line tool with the text to speak, a short prompt recording and a transcript of that recording. The model saves speech in the voice of the prompt into the folder you named. Clone only voices you have the right to use.

For newcomers and experts: A language model normally predicts words. Here it predicts codes that stand for sounds, and a decoder turns those codes into audio. A good first try is the demo script in the example folder.

Under the hood, the paper title in the card describes single-stream decoupled speech tokens. The language model predicts those tokens and the audio is rebuilt from them directly. The Qwen2.5 base keeps the system simple. The hub lists 6 quantized and 25 fine-tuned versions made by other people.

Text, and optional prompt audio
Language model predicts speech codes
Codes are rebuilt directly into audio
Audio file is saved

Good to know:

• The training code and the VoxBox training dataset are not released yet, according to the card's to-do list.
• License: CC BY-NC-SA 4.0, which does not allow commercial use
• Cost to run: none named
• Data it reaches: the text and prompt audio you give it, and the model files downloaded from the Hugging Face hub
• Not for: commercial use, and cloning voices without the right to use them

On Hugging Face, 13 November 2026

• Task: text to speech
• Formats: Safetensors
• Downloads in the last 30 days: 2,124
• Likes: 750
• License: cc-by-nc-sa-4.0
• Languages: en, zh
• Quantized versions: 6
• Fine-tunes: 25
• Tags: safetensors, text-to-speech

Links

• Model: https://huggingface.co/SparkAudio/Spark-TTS-0.5B
• Files: https://huggingface.co/SparkAudio/Spark-TTS-0.5B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download SparkAudio/Spark-TTS-0.5B

Then follow the usage in its model card. Nothing in this kit runs by itself.