Hugging Face's logo

Datasets:
olewave
/
OleSpeech-IV-2025-EN-AR-100

Modalities:
Audio
Languages:
English
Size:
n<1K
ArXiv:
Tags:
audio
speech
conversation
dialog
dialogue
speaker diariazation
License:
Dataset card Files
xet
Community
2

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

By accessing or using this dataset, you agree to the following terms:

  • Non-commercial research use only
  • No redistribution or sublicensing
  • Citation required for publications:
    Olewave, “OleSpeech-IV: A Large-Scale Multispeaker and Multilingual Conversational Speech Dataset with Diverse Topics.” arXiv preprint, arXiv:2509.04702, 2025. https://arxiv.org/abs/2509.04702
  • All rights remain with Olewave
  • For the full license terms, see LICENSE or contact [email protected]

Log in or Sign Up to review the conditions and access this dataset content.

OleSpeech-IV-2025-EN-AR-100

OleSpeech-IV-2025-EN-AR-100 is a publicly released subset of OleSpeech-IV, a large-scale multilingual and multispeaker conversational speech dataset. This release is intended for non-commercial research use and provides high-quality speech data aligned with transcripts, speaker turns, and additional metadata. A sample of the dataset can be found in the sample/ folder.

Overview

OleSpeech is a tiered dataset series designed to provide speech resources with varying levels of annotation. OleSpeech-IV-2025-EN-AR-100 represents a curated subset of Tier IV data, focusing on English content, featuring:

  • Multispeaker conversational recordings
  • Diverse topics from English podcasts, talk shows, teleconferences, and other conversations
  • Human-refined transcripts, including speaker labels and turn information
  • Fine-grained timestamps and confidence scores generated using the proprietary Olign alignment pipeline

This subset aims to support research in:

  • Speech recognition (ASR)
  • Speaker diarization
  • Speech-to-text alignment
  • Multilingual conversational AI

Contents

The dataset includes:

  1. Audio files (FLAC format, mono, 16 kHz)
  2. Transcripts with spoken content and audio events
  3. Metadata: timestamps, speaker IDs, and confidence scores

The folder structure follows a conventional train/dev/test split:

├── dev
│ ├── pub_air_program
│ ├── pub_indi_podcast
│ └── pub_talk_shows
├── dev_id2dur.txt
├── test
│ ├── pub_air_program
│ ├── pub_indi_podcast
│ └── pub_talk_shows
├── test_id2dur.txt
├── train
│ ├── pub_air_program
│ ├── pub_indi_podcast
│ └── pub_talk_shows
└── train_id2dur.txt

dev/pub_air_program

├── 3SDOr0_BtF67d5b
│ ├── 3SDOr0_BtF67d5b.en.spker.block.json
│ └── 3SDOr0_BtF67d5b.flac

The format of *.en.spker.block.json

{
  "speaker": 0,
  "start_time": "0.05",
  "end_time": "37.80",
  "confidence": "0.88",
  "block": [
    {
      "start_time": "0.05",
      "end_time": "2.08",
      "transcript": " that he got there,",
      "confidence": "0.84",
      "overlap": false
    },
    {
      "start_time": "2.14",
      "end_time": "4.43",
      "transcript": "that he got to the place to act like a fighter,",
      "confidence": "0.93",
      "overlap": false
    },
    // ...
    {
      "start_time": "34.23",
      "end_time": "37.80",
      "transcript": "this is Cus, with punches with bad intentions.",
      "confidence": "0.81",
      "overlap": false
    }
  ]
},
// ...
{
        "speaker": 1,
        "start_time": "264.38",
        "end_time": "264.62",
        "confidence": "0.80",
        "block": [
            {
                "start_time": "264.38",
                "end_time": "264.62",
                "transcript": "Yeah.",
                "confidence": "0.80",
                "overlap": true
            }
        ]
    },
    {
        "speaker": 0,
        "start_time": "264.38",
        "end_time": "270.78",
        "confidence": "0.74",
        "block": [
            {
                "start_time": "264.38",
                "end_time": "265.10",
                "transcript": "You know what I mean?",
                "confidence": "0.54",
                "overlap": true
            },
            // ...
    }
// ...
}

License and Data Use

This dataset is provided under the Olewave Data Use Agreement (DUA). By accessing or using this dataset, you agree to the following terms:

  • Non-commercial research use only
  • No redistribution or sublicensing
  • Citation required for publications Olewave, “OleSpeech-IV: A Large-Scale Multispeaker and Multilingual Conversational Speech Dataset with Diverse Topics.” arXiv preprint, arXiv:2509.04702, 2025. https://arxiv.org/abs/2509.04702
  • All rights remain with Olewave

For the full license terms, see LICENSE or contact [email protected].

Downloads last month
1,075