Dataset · charcot-ai
Marimba Bench multilingual audio benchmark
charcot-ai/marimba-bench
A collection of 136 hours of real and synthetic audio recordings in 6 languages for evaluating speech and sound recognition models.
What's inside
The research paper dataset contains 90,112 recordings with a total duration of 136.38 hours in mono 16 kHz format, packaged in parquet. The data is divided into 25 configurations in English, Spanish, French, Italian, Portuguese, and Romanian. It includes annotations for voice activity detection, speaker segmentation, language identification, keywords, sound classes, and synthetic markers.
What you can do with it
- Comprehensive evaluation and comparison of audio encoders across multiple tasks from VAD to sound classification
- Training or testing synthetic speech and deepfake detectors across six European languages
- Fine-tuning keyword spotting and language identification models in streaming audio
- Tasks
- Audio classification, Voice activity detection
- Languages
- English, Spanish, French, Italian, Portuguese, Romanian, Arabic, Chinese and 4 more
- Size
- 10K–100K rows
- Downloads this month
- 781
- Likes
- 2
How to load it
from datasets import load_dataset
ds = load_dataset("charcot-ai/marimba-bench")