lugan/Syntts-Commands-Media-Dataset

SynTTS-Commands: A Multilingual Synthetic Speech Command Dataset

📖 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, 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

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

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 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} }