RealTalk-CN: A Realistic Chinese Speech-Text Dialogue Benchmark With Cross-Modal Interaction Analysis

📌 Resources:

• GitHub Repository
• Arxiv Paper

RealTalk-CN is the first large-scale, multi-domain, bimodal (speech-text) Chinese Task-Oriented Dialogue (TOD) dataset. All data come from real human-to-human conversations, specifically constructed to advance research on speech-based large language models (Speech LLMs). Existing TOD datasets are mostly text-based, lacking real speech, spontaneous disfluencies, and cross-modal interaction scenarios. RealTalk-CN achieves breakthroughs in these aspects, fully supporting Chinese speech dialogue modeling and evaluation. The dataset is released under the CC BY-NC-SA 4.0 license, and can be freely used for non-commercial research.

---

Dataset Composition

• Total Duration: ~150 hours of verified real human-to-human dialogue audio
• Dialogue Scale: 5,400 multi-turn dialogues, over 60,000 utterances
• Speakers: 113 individuals, balanced gender ratio, ages 18–50, covering major dialect regions across China
• Dialogue Domains: 58 task-oriented domains (e.g., dining, transportation, shopping, healthcare, finance), including 55 intents and 115 slots
• Audio Specifications: 16kHz sampling rate, WAV format, recorded via both professional and mobile devices
• Transcription & Annotation:
• Manually transcribed at the character level, preserving spoken language features
• Annotated with 4 categories of disfluencies (elongation, repetition, self-correction, hesitation)
• Includes transcriptions, slot values, intents, and speaker metadata (gender, age, region, etc.)

---

Dataset Features

1. Natural and Colloquial: Contains spoken features and disfluencies in real task-oriented dialogues, overcoming the limitation of “read speech” corpora.
2. Bimodal and Real Interaction: Provides paired speech-text annotations and introduces a cross-modal chat task, supporting dynamic switching between speech and text—closer to real-world human-computer interaction.
3. Complete Dialogues and Multi-Domain Coverage: Average of 12 turns per dialogue, covering 58 real-world domains, supporting both single-domain and cross-domain dialogue modeling.
4. Diverse Speakers: Covers major regions in China, balanced across gender and age, enabling research on the impact of accents, dialects, and demographic differences.
5. High-Quality Annotation and Strict Quality Control: Multiple rounds of manual verification, detailed timestamps, and slot annotations ensure reliability and research value.

---

Advantages

• The first large-scale Chinese speech-text TOD corpus, filling the gap in benchmark datasets for Chinese spoken dialogue.
• Provides disfluency annotations, supporting robustness evaluation and error correction research in speech-based TOD systems.
• Enables research in speech recognition, speech synthesis, intent recognition, slot filling, dialogue management, and cross-modal studies.
• Serves as a benchmark for Speech LLMs in Chinese TOD tasks, driving the development of advanced speech interaction systems.

---

Conclusion

The release of RealTalk-CN lays the foundation for research in Chinese speech-text bimodal dialogue. With its large scale, multi-domain coverage, natural spoken language, diverse speakers, and cross-modal interaction, it not only advances the development of Speech LLMs in task-oriented dialogue but also provides a key resource for future cross-modal and multimodal intelligent systems.

Data format notes

In subset_json, text is the target intent label; the utterance transcript is original_data["文本content"]. context contains all earlier effective, nonempty original turns from both speakers, including opening turns. next is the immediately following effective turn and is {} at the original dialogue end. It can be another fragment from the same speaker, so adjacency alone does not establish a question–answer pair.

Gender follows the original role annotation. Unknown age and region are null; known ages remain numeric. choices retains its existing serialized Python-list representation. Intent and character-level slot IDs are recorded in label_mappings.json. Existing intent IDs are unchanged; 提供天气信息 uses ID 55 for a corrected weather-information example.

The audio archive has the prefix media/storage/wangenzhi_space/direct_test/data_submit/Spoken3MC/download/wavs/. To use audio_file as a relative path, place that extracted wavs directory at Spoken3MC/wavs/.

data_quality_review.json lists source annotations that still need review and remaining original timestamp gaps. A timestamp gap alone does not establish whether the audio contains silence or unannotated speech.