OpenETD: Metadata for Speech Turn Detection Training Splits

HJOK released this metadata for the OpenETD dataset in April 2026. It provides train, development, and test splits for a dataset of real and synthetic audio files used to train end-of-turn detectors for speech chatbots. The splits contain 6,290 real and 96,773 synthetic files for training, with corresponding development and test sets.

Use Cases

• Train end-of-turn detection models based on the described split of real and synthetic audio files.
• Benchmark speculative turn-taking algorithms based on the provided train/dev/test partitions.
• Analyze the impact of synthetic data on model performance based on the described file and hour counts per split.

Strengths

• Provides explicit split sizes: 6,290 real and 96,773 synthetic files for training.
• Includes duration information in hours for each split and data type (real/synthetic).
• Associated with a peer-reviewed ACL 2026 Findings paper and public GitHub repository.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

HJOK, associated with the ACL 2026 Findings paper 'Speculative End-Turn Detector for Efficient Speech Chatbot Assistant'.

Collection Method

Likely created to partition the OpenETD dataset for machine learning experiments.

Freshness

Last updated 2026-04-20 16:15:03; freshness should be verified.