README.md

---
license: mit
task_categories:
- audio-classification
language:
- en
- zh
tags:
- keyword-spotting
- tts-synthetic
- multilingual
- speech-commands
- tinyml
- edge-ai
---

# SynTTS-Commands: A Multilingual Synthetic Speech Command Dataset

<!-- Badges Row -->

![Python](https://img.shields.io/badge/python-3.8%2B-blue)
![Domain](https://img.shields.io/badge/domain-Speech--TinyML-orange)
![Size](https://img.shields.io/badge/size-35.76%20GB-purple)
![Utterances](https://img.shields.io/badge/utterances-384K-yellow)
![Speakers](https://img.shields.io/badge/speakers-8,100-lightblue)
[![arXiv](https://img.shields.io/badge/arXiv-2511.07821-b31b1b.svg)](https://arxiv.org/abs/2511.07821)
[![Benchmarks](https://img.shields.io/badge/🤗%20Hugging%20Face-Benchmarks-ffd21e)](https://huggingface.co/lugan/SynTTS-Commands-Media-Benchmarks)
[![Code](https://img.shields.io/badge/GitHub-Repository-181717)](https://github.com/lugan113/SynTTS-Commands-Official)
[![License](https://img.shields.io/badge/License-MIT-blue.svg)](https://opensource.org/licenses/MIT)


## 📖 Introduction

**SynTTS-Commands** is a large-scale, multilingual synthetic speech command dataset specifically designed for low-power Keyword Spotting (KWS) and speech command recognition tasks. As presented in the paper [SynTTS-Commands: A Public Dataset for On-Device KWS via TTS-Synthesized Multilingual Speech](https://huggingface.co/papers/2511.07821), this dataset is generated using advanced Text-to-Speech (TTS) technologies, aiming to address the scarcity of high-quality training data in the fields of TinyML and Edge AI.

### 🎯 Core Features

- **Multilingual Coverage**: Includes bilingual speech commands in both **Chinese** and **English**.
- **High-Quality Synthesis**: Generated via advanced TTS technology, ensuring high naturalness in speech.
- **Speaker Diversity**: Incorporates multiple acoustic feature sources to ensure a rich variety of speaker styles.
- **Real-World Scenarios**: Commands are designed for practical applications, including smart homes, in-car systems, and multimedia control.
- **Rigorous Quality Assurance**: All speech data has been screened via ASR models combined with manual human verification.

## 📊 Dataset Overview

### Statistics

The **SynTTS-Commands-Media-Dataset** contains a total of **384,621 speech samples**, covering **48 distinct multimedia control commands**. It is divided into four subsets with the following distribution:

| Subset | Speakers | Commands | Samples | Duration (hrs) | Size (GB) |
|------|----------|--------|----------|------------|----------|
| Free-ST-Chinese | 855 | 25 | 21,214 | 6.82 | 2.19 |
| Free-ST-English | 855 | 23 | 19,228 | 4.88 | 1.57 |
| VoxCeleb1&2-Chinese | 7,245 | 25 | 180,331 | 58.03 | 18.6 |
| VoxCeleb1&2-English | 7,245 | 23 | 163,848 | 41.6 | 13.4 |
| **Total** | **8,100** | **48** | **384,621** | **111.33** | **35.76** |

### Dataset Highlights

- **Massive Scale**: Totaling **111.33 hours** and **35.76 GB** of synthetic speech data, making it one of the largest synthetic speech command datasets for academic research.
- **Extensive Speaker Diversity**: Covers **8,100 unique speakers**, spanning various accent groups, age ranges, and recording conditions.
- **Multi-Dimensional Research Support**: The four-subset structure enables research into cross-lingual speaker adaptation, speaker diversity effects, and acoustic robustness in different recording environments.
- **Application-Oriented**: Specifically focused on multimedia playback control scenarios, providing high-quality training data for real-world deployment.

### Directory Structure
```text
SynTTS-Commands-Media-Dataset/
├── Free_ST_Chinese/        # 21,214 Chinese media control samples (855 speakers)
├── Free_ST_English/        # 19,228 English media control samples (855 speakers)
├── VoxCeleb1&2_Chinese/    # 180,331 Chinese media control samples (7,245 speakers)
├── VoxCeleb1&2_English/    # 163,848 English media control samples (7,245 speakers)
├── reviewed_bad/           # Rejected speech samples (failed quality audit)
├── splits_by_language/     # Dataset splits organized by language
│   ├── train/              # Training set
│   ├── val/                # Validation set
│   └── test/               # Test set
└── comprehensive_metadata.csv # Complete metadata file
```


🎯 Media Command Categories

English Media Control Commands (23 Classes)

Playback Control: "Play", "Pause", "Resume", "Play from start", "Repeat song"
Navigation: "Previous track", "Next track", "Last song", "Skip song", "Jump to first track"
Volume Control: "Volume up", "Volume down", "Mute", "Set volume to 50%", "Max volume"
Communication: "Answer call", "Hang up", "Decline call"
Wake Words: "Hey Siri", "OK Google", "Hey Google", "Alexa", "Hi Bixby"

Chinese Media Control Commands (25 Classes)

Playback Control: "播放", "暂停", "继续播放", "从头播放", "单曲循环"
Navigation: "上一首", "下一首", "上一曲", "下一曲", "跳到第一首", "播放上一张专辑"
Volume Control: "增大音量", "减小音量", "静音", "音量调到50%", "音量最大"
Communication: "接听电话", "挂断电话", "拒接来电"
Wake Words: "小爱同学", "Hello 小智", "小艺小艺", "嗨 三星小贝", "小度小度", "天猫精灵"

## 📈 Benchmark Results and Analysis

We present a comprehensive benchmark of **six representative acoustic models** on the SynTTS-Commands-Media Dataset across both English (EN) and Chinese (ZH) subsets. All models are evaluated in terms of **classification accuracy**, **cross-entropy loss**, and **parameter count**, providing insights into the trade-offs between performance and model complexity in multilingual voice command recognition.

### Performance Summary

| Model | EN Loss | EN Accuracy | EN Params | ZH Loss | ZH Accuracy | ZH Params |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| **MicroCNN** | 0.2304 | 93.22% | 4,189 | 0.5579 | 80.14% | 4,255 |
| **DS-CNN** | 0.0166 | 99.46% | 30,103 | 0.0677 | 97.18% | 30,361 |
| **TC-ResNet** | 0.0347 | 98.87% | 68,431 | 0.0884 | 96.56% | 68,561 |
| **CRNN** | **0.0163** | **99.50%** | 1.08M | 0.0636 | **97.42%** | 1.08M |
| **MobileNet-V1** | 0.0167 | **99.50%** | 2.65M | **0.0552** | 97.92% | 2.65M |
| **EfficientNet** | 0.0182 | 99.41% | 4.72M | 0.0701 | 97.93% | 4.72M |

## 🗺️ Roadmap & Future Expansion

We are expanding SynTTS-Commands beyond multimedia to support broader Edge AI applications. 

👉 **[Click here to view our detailed Future Work Plan & Command List](https://github.com/lugan113/SynTTS-Commands-Official/blob/main/Future_Work_Plan.md)**

Our upcoming domains include:
*   🏠 **Smart Home:** Far-field commands for lighting and appliances.
*   🚗 **In-Vehicle:** Robust commands optimized for high-noise driving environments.
*   🚑 **Urgent Assistance:** Safety-critical keywords (e.g., "Call 911", "Help me") focusing on high recall.

We invite the community to review our [Command Roadmap](https://github.com/lugan113/SynTTS-Commands-Official/blob/main/Future_Work_Plan.md) and suggest additional keywords!

📜 Citation

If you use this dataset in your research, please cite our paper:

@misc{gan2025synttscommands,
      title={SynTTS-Commands: A Public Dataset for On-Device KWS via TTS-Synthesized Multilingual Speech}, 
      author={Lu Gan and Xi Li},
      year={2025},
      eprint={2511.07821},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2511.07821}, 
      doi={10.48550/arXiv.2511.07821}
}