Chinese Multi-Speaker Spontaneous Conversations with Separate Audio Tracks

MagicData's Multi-stream conversation dataset captures each speaker's audio track separately, preserving natural conversational phenomena. The dataset is an ASR corpus containing Chinese audio recorded on mobile devices at 16 kHz and 16 bits. It was created by MagicHub and last updated on the platform in June 2026.

Use Cases

• Train speaker diarization models based on the separate, labeled audio tracks for each speaker.
• Develop automatic speech recognition systems for spontaneous Chinese conversation.
• Research conversational speech phenomena like overlaps and disfluencies using the multi-stream format.
• Fine-tune speech models for mobile device audio quality based on the specified recording parameters.

Strengths

• Audio is recorded at a standard quality of 16 kHz and 16 bits.
• Data captures spontaneous conversation, which may include natural speech phenomena.
• Each speaker's audio is captured and labeled separately, enabling multi-speaker analysis.

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 size, row count, and specific recording conditions are unknown.

Provenance

Source

MagicHub on Hugging Face.

Collection Method

Recorded using mobile devices.

Freshness

Last updated 2026-06-09 06:08:46; freshness should be verified.