Japanese Audio Events with Speech Recognition and Aesthetic Scores

Japanese audio data contains 266 hours of speech processed by Scribe v1 for automatic speech recognition and classified using Facebook's audio aesthetics model as a prefilter. The dataset is derived from the Japanese portion of the Emilia Yodas collection and is licensed under CC BY 4.0. It includes text transcriptions and aesthetic scores for audio events.

Use Cases

• Train or fine-tune automatic speech recognition (ASR) models using the Japanese speech data and its Scribe v1 transcriptions.
• Develop audio quality classifiers by leveraging the Facebook audio aesthetics scores as a prefilter or training signal.
• Analyze the relationship between audio aesthetic scores and transcription accuracy across different audio events.
• Build Japanese language models or text-to-speech systems using the transcribed audio corpus.

Strengths

• Contains 266 hours of Japanese audio data, providing a substantial volume for model training.
• Includes both automatic speech recognition outputs from Scribe v1 and audio aesthetic scores from a Facebook model.
• Derived from the established Emilia Yodas dataset, indicating a structured source.

Limitations

• The dataset is described as a 'v1' version, suggesting it is an initial release that may contain inconsistencies or require refinement.
• Specific details on audio quality, speaker demographics, or transcription accuracy are not provided in the input.
• The sample data and exact column structure are unavailable, limiting pre-download assessment.

Provenance

Source

Japanese portion of the Emilia Yodas dataset from Hugging Face.

Collection Method

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

Freshness

Last updated on 2025-04-14.

Geography

Japan (Language: Japanese).