---
language:
- en
task_categories:
- automatic-speech-recognition
- text-generation
- audio-classification
tags:
- spoken-dialogue
- full-duplex
- turn-taking
- interruption
- backchannel
pretty_name: InteractSpeech English Text and Timeline
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*.parquet
---

# InteractSpeech: English Text-and-Timeline Release

This is the English text-and-timeline release associated with **InteractSpeech: A Speech Dialogue Interaction Corpus for Spoken Dialogue Model** (Findings of EMNLP 2025).

InteractSpeech is designed for real-time spoken-dialogue interaction, including interruptions, backchannels, pauses, gaps, overlaps, and turn transitions. The paper describes an approximately 150-hour English corpus containing 90,000 text utterances.

## Release contents

| Item | Value |
| --- | ---: |
| Training records | 59,020 |
| Parquet shards | 3 |
| Audio waveforms | Not included |
| Video assets | Not included |

The release preserves dialogue text, speaker timing, interaction events, overlap regions, complete interrupted utterances, quality metadata, and per-utterance TTS reconstruction specifications. Audio can be regenerated with any TTS system while preserving each published segment duration and the global dialogue timeline.

## Data format

Each Parquet row contains:

| Column | Description |
| --- | --- |
| `id` | Public record identifier |
| `view` | Dataset view |
| `category` | Dataset category |
| `record_json` | Complete record serialized as JSON |

```python
import json
import pyarrow.parquet as pq

table = pq.read_table("data/train-00000.parquet", columns=["record_json"])
sample = json.loads(table["record_json"][0].as_py())
```

The decoded record contains:

- `turns`: the training dialogue with timestamps and input-speech reconstruction recipes;
- `meta`: dialogue-level language, topic, event counts, and truncation information;
- `source_record`: the preserved InteractionSpeech dialogue, event annotations, full utterances, overlap regions, and source-to-training turn mapping.

For each input-speech segment, synthesize the text in `audio_in.source.transcript` using a consistent voice for the same `voice_key`, then fit it to `reference_duration_ms`. Do not alter the published global `start_ms` and `end_ms` values.

## Citation

```bibtex
@inproceedings{chen-etal-2025-interactspeech,
    title = {{I}nteract{S}peech: A Speech Dialogue Interaction Corpus for Spoken Dialogue Model},
    author = {Chen, Yifu and Ji, Shengpeng and Wang, Ziqing and Wang, Hanting and Zhao, Zhou},
    booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2025},
    month = nov,
    year = {2025},
    address = {Suzhou, China},
    publisher = {Association for Computational Linguistics},
    pages = {8024--8033},
    doi = {10.18653/v1/2025.findings-emnlp.424},
    url = {https://aclanthology.org/2025.findings-emnlp.424/}
}
```

- Paper: https://aclanthology.org/2025.findings-emnlp.424/
- Project page and audio examples: https://interactspeech.github.io/