TimeGround-1M: 1 Million Synthetic Audio Examples for Time-Aware Speech Tasks

TimeGround-1M provides 1 million synthetic English audio examples annotated for time-aware speech understanding. The dataset was created by ai-sage from 14,000 hours of audio filtered from a 24,000-hour source pool, with annotations verified through LLM-based and deterministic checks. It was last updated on July 14, 2026.

Use Cases

• Train models for temporal localization based on synthetic time-grounded audio annotations.
• Develop systems for generating temporal descriptions from audio based on the described tasks.
• Benchmark timed audio summarization models using the dataset's structured summaries.
• Fine-tune speech understanding models on the three specific time-grounded tasks mentioned.

Strengths

• Derived from a filtered source of 14,000 hours of English audio.
• Annotations underwent LLM-based verification and deterministic validity checks.
• Specifically designed for three distinct time-grounded speech understanding tasks.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is synthetic, which may limit real-world applicability.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

YODAS2 English shards, filtered by ai-sage.

Collection Method

Synthetic annotations generated for three tasks, filtered via language- and silence-ratio, then verified.

Freshness

Last updated 2026-07-14 05:37:09; freshness should be verified.