MMAG: A Multi‑Control Mixed Audio Generation Benchmark

MMAG is a comprehensive benchmark for evaluating mixed audio generation under multiple control conditions. It assesses a model's ability to generate coherent acoustic scenes containing speech, music, and sound effects simultaneously, while supporting fine-grained control over speaker identity and temporal alignment.

Dataset Structure

The dataset is organized into three subsets, each targeting a specific evaluation dimension:

Config Name	Clips	Description
main_set	~4,000	Primary evaluation set for holistic generation quality
voice_cloning_set	~690	Speaker identity preservation from short voice prompts
timestamp_set	~1,800	Temporal alignment and event timing control

Each subset is provided as a separate Hugging Face configuration.

File Structure

MMAG/
├── README.md
├── dataset_infos.json
├── audio_files/
│   └── *.wav
├── prompt_audio/
│   └── *.wav
├── main_set/
│   └── metadata.jsonl
├── voice_cloning_set/
│   └── metadata.jsonl
└── timestamp_set/
    └── metadata.jsonl

Loading the Dataset

Option 1 Soundfile

Install the required dependencies:

pip install datasets soundfile huggingface_hub  

Load any subset via Hugging Face datasets:

from huggingface_hub import snapshot_download
import soundfile as sf
from datasets import load_dataset

# Step 1: Download the entire dataset repository
snapshot_download(
    repo_id="rookie9/MMAG",
    repo_type="dataset",
    local_dir="./mmag_data"
)

# Step 2: Load the metadata
dataset = load_dataset("rookie9/MMAG", "main_set", split="test")

# Step 3: Access audio files from local directory
example = dataset[0]
audio, sr = sf.read(f"./mmag_data/{example['file_name']}")
print(f"Audio shape: {audio.shape}, Sample rate: {sr}")

For the voice cloning subset, the corresponding voice prompt is loaded from the voice_prompt field:

prompt_audio, sr = sf.read(f"./mmag_data/{example['voice_prompt']}")

Option 2 TorchCodec

pip install datasets torchcodec

from datasets import load_dataset, Audio

dataset = load_dataset("rookie9/MMAG", "main_set", split="test")
dataset = dataset.cast_column("file_name", Audio())

example = dataset[0]
audio = example["file_name"]["array"]
sr = example["file_name"]["sampling_rate"]

Annotation Fields

Each subset's metadata file (metadata.jsonl) contains the following core fields:

Main Set (main_set/metadata.jsonl)

Field	Description
file_name	Audio file path (relative to root)
id	Unique identifier for the clip
caption	Overall caption describing the full acoustic scene (see Caption Annotation Coverage below)

Voice Cloning Set (voice_cloning_set/metadata.jsonl)

Field	Description
file_name	Target audio file path (relative to root)
voice_prompt	Voice prompt file path (relative to root)
caption	Overall caption
id	Unique identifier for the clip

Timestamp Set (timestamp_set/metadata.jsonl)

Field	Description
file_name	Audio file path (relative to root)
id	Unique identifier for the clip
caption	Overall caption
timestamped_caption	Caption with precise temporal boundaries

Caption Annotation Coverage

The caption field in each subset is not a simple description; it is a richly structured annotation that covers multiple acoustic and semantic dimensions. The following table summarizes the types of information encoded in the captions:

Category	Attributes Covered in Caption
Speech	Transcription, speaker gender, age, accent, emotion
Music	Instrumentation, genre, mood
Sound Events	Foreground and background sound description, ambient acoustic context
Temporal	Relative ordering of events (e.g., "before", "after", "throughout"), fine-grained timestamps

All captions are generated through a multi-expert annotation pipeline and verified by human quality control to ensure factual accuracy and consistency.

License

This dataset is released under the CC BY 4.0 license for non-commercial research use. All source audio materials are sourced from publicly available datasets and are used in accordance with their original licenses.