🗣️ LibriConvo-Segmented

LibriConvo-Segmented is a segmented version of the LibriConvo corpus — a simulated two-speaker conversational dataset built using Speaker-Aware Conversation Simulation (SASC).
It is designed for training and evaluation of multi-speaker speech processing systems, including speaker diarization, automatic speech recognition (ASR), and overlapping speech modeling.

This segmented version provides ≤30-second conversational fragments derived from full LibriConvo dialogues, with 40% of them having room impulse responses applied on them.

The full paper, detailing the creation of the corpus, as well as baseline ASR and diarization results can be found here: 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 segmented release provides shorter, self-contained audio clips suitable for fine-tuning ASR and diarization models.

---

📦 Dataset Summary

Split	# Segments
Train	30,313
Validation	3,595
Test	3674

Sampling rate: 16 kHz
Audio format: WAV (mono)
Split criterion: Speaker-disjoint

---

📂 Data Structure

Each row represents a single speech segment belonging to a simulated conversation between two speakers.

Field	Type	Description
conversation_id	string	Conversation identifier
utterance_idx	int64	Utterance index within the conversation
abstract_symbol	string	High-level symbolic utterance ID ('A' or 'B')
transcript	string	Text transcription of the utterance
duration_sec	float64	Segment duration (seconds)
rir_file	string	Room impulse response file used
delay_sec	float64	Delay applied for realistic speaker overlap
start_time_sec, end_time_sec	float64	Start and end times within the conversation
abs_start_time_sec, abs_end_time_sec	float64	Global (absolute) start and end times
segment_id	int64	Local segment index
segment_conversation_id	string	Unique segment identifier
split	string	One of train, validation, or test
audio	Audio (16 kHz)	Decoded audio data

---

🚀 Loading the Dataset

from datasets import load_dataset

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

print(ds)
# DatasetDict({
#     train: Dataset(...),
#     validation: Dataset(...),
#     test: Dataset(...)
# })

---

📚 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}, 
}