ASR-KCSC: 5.22 Hours of Korean Conversational Speech

ASR-KCSC is an open-source Korean conversational speech corpus containing 5.22 hours of transcribed audio. The data consists of 22 conversations between seven pairs of speakers recorded on mobile devices in indoor environments. Author MagicHub released the dataset on Hugging Face, with a last recorded update in June 2026.

Use Cases

• Train Korean speech recognition models based on conversational audio data.
• Fine-tune language models for Korean dialogue generation based on transcribed speech.
• Benchmark ASR system performance on mobile-recorded, indoor Korean conversations.
• Study conversational dynamics and turn-taking in Korean based on speaker-pair recordings.

Strengths

• Contains 5.22 hours of transcribed conversational speech, providing a focused resource.
• Audio is recorded at a consistent 16 kHz, 16-bit quality in WAV (PCM) format.
• Includes 22 conversations between seven distinct speaker pairs, offering some speaker diversity.
• Recording equipment and environment (mobile device, indoor) are clearly specified.

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's scale (5.22 hours) is relatively small for training large-scale ASR models from scratch.

Provenance

Source

MagicHub on Hugging Face.

Collection Method

Recorded on mobile devices in indoor environments.

Freshness

Last updated 2026-06-09 06:07:42; freshness should be verified.