DuplexConv: 2,000-Hour Chinese Conversational Speech with LLM Annotations

DuplexConv is a 2,000-hour Chinese multi-channel conversational speech dataset with annotations assisted by large language models. It was released by ASLP@NPU and QualiaLabs as part of the SmoothConv & DuplexConv project. The dataset is intended for large-scale training, with its companion dataset, SmoothConv, serving for fine-grained evaluation.

Use Cases

• Train automatic speech recognition (ASR) models based on multi-channel conversational audio.
• Develop speech enhancement or diarization systems based on multi-channel conversational data.
• Fine-tune or evaluate large language models on speech-related tasks using the LLM-assisted annotations.
• Benchmark conversational AI systems using the companion SmoothConv dataset for evaluation.

Strengths

• 2,000 hours of audio data, which is a substantial scale for training.
• Annotations are created with assistance from large language models, potentially increasing consistency.
• Part of a project with a companion expert-annotated dataset (SmoothConv) for evaluation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

ASLP@NPU and QualiaLabs

Collection Method

LLM-assisted annotations on a multi-channel conversational speech corpus.

Freshness

Last updated 2026-06-10 09:55:53; freshness should be verified.

Geography

Likely China, based on the language focus.