🗣️ LibriConvo-Raw

LibriConvo-Raw is the full-length, unsegmented version of the LibriConvo corpus — a simulated two-speaker conversational dataset created using Speaker-Aware Conversation Simulation (SASC).
It is designed for training and evaluation of conversational speech systems, particularly for multi-speaker ASR, speaker diarization, and overlap detection.

Unlike the segmented release, this version contains complete simulated dialogues with natural temporal structure, pauses, and overlaps preserved exactly as modeled by SASC.

The full paper describing the dataset generation, simulation pipeline, and baseline results is available at:
🔗 https://arxiv.org/abs/2510.23320

---

🧠 Overview

LibriConvo ensures natural conversational flow and contextual coherence by:

• Organizing LibriTTS utterances by book to maintain narrative continuity.
• Using statistics from CallHome for pause modeling.
• Applying compression to remove excessively long silences while preserving turn dynamics.
• Enhancing acoustic realism via a novel Room Impulse Response (RIR) selection procedure, ranking configurations by spatial plausibility.
• Producing speaker-disjoint splits for robust evaluation and generalization.

In total, the full LibriConvo corpus comprises 240.1 hours across 1,496 dialogues with 830 unique speakers.

This version is particularly suited for end-to-end conversational modeling, long-form ASR, diarization pretraining, and speaker interaction analysis.

---

🎧 Dataset Summary

Split	# Conversations	Duration (approx.)
Train	1,199	~193.7 hours
Validation	137	~23.1 hours
Test	160	~23.4 hours

Total duration: ~240.1 hours
Unique speakers: 830
Sampling rate: 16 kHz
Audio format: WAV (mono)
Split criterion: Speaker-disjoint
RIR coverage: ~40% of conversations include room impulse response convolution

---

📂 Data Structure

Each row in the dataset represents a complete two-speaker conversation with full dialogue audio and time-aligned utterances.

Field	Type	Description
conversation_id	string	Unique conversation identifier
split	string	One of train, validation, or test
utterance_idx	sequence(int64)	Ordered list of utterance indices
abstract_symbol	sequence(string)	Speaker label sequence (A or B)
start_time	sequence(float64)	Start time of each utterance (seconds)
end_time	sequence(float64)	End time of each utterance (seconds)
text	sequence(string)	Transcription of each utterance
duration_sec	sequence(float64)	Duration of each utterance (seconds)
rir	bool	Indicates if a Room Impulse Response was applied
audio	Audio (16 kHz)	Full conversation waveform

---

🗂️ Example

from datasets import load_dataset

ds = load_dataset("gedeonmate/LibriConvo-raw")

sample = ds["train"][0]

print(sample["conversation_id"])
print(sample["text"][:5])   # First few utterances

---

📚 Citation

If you use the LibriConvo dataset or the associated Speaker-Aware Conversation Simulation (SASC) methodology in your research, please cite the following papers:

@misc{gedeon2025libriconvo,
  title         = {LibriConvo: Simulating Conversations from Read Literature for ASR and Diarization},
  author        = {Máté Gedeon and Péter Mihajlik},
  year          = {2025},
  eprint        = {2510.23320},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS},
  url           = {https://arxiv.org/abs/2510.23320}
}

@misc{gedeon2025sasc,
      title={From Independence to Interaction: Speaker-Aware Simulation of Multi-Speaker Conversational Timing}, 
      author={Máté Gedeon and Péter Mihajlik},
      year={2025},
      eprint={2509.15808},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2509.15808}, 
}