Mdf En Emilia Yodas: 616 Hours of English Audio Events

616 hours of English audio extracted from the Emilia-Dataset, licensed under CC BY 4.0. The audio events are classified using Scribe v1, an STT/ASR system from ElevenLabs, and filtered using Facebook audio aesthetics metrics. The dataset is described as a 'v1' version, with further collaboration invited via Discord.

Use Cases

• Training speech-to-text models based on the English audio content.
• Developing audio event classifiers based on the event classification via Scribe v1.
• Pre-filtering audio datasets based on aesthetic metrics derived from Facebook's system.
• Benchmarking ASR systems using a large-scale, English-only audio corpus.

Strengths

• Contains 616 hours of English audio, providing a substantial volume for training.
• Audio events are classified using a named STT/ASR system (Scribe v1 from ElevenLabs).
• Utilizes a named prefilter (Facebook audio aesthetics) for quality selection.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is described as a 'v1', suggesting it may be an early, potentially unpolished version.

Provenance

Source

Derived from the Emilia-Dataset on Hugging Face (https://huggingface.co/datasets/amphion/Emilia-Dataset).

Collection Method

Audio events classified via Scribe v1 (ElevenLabs STT/ASR) and filtered using Facebook audio aesthetics.

Freshness

Last updated 2025-12-18 03:51:18; freshness should be verified.