Qlean Dataset Launches Japanese 3-Speaker Business Dialogue Audio and Transcript Dataset

Visual Bank Co., Ltd. (Minato-ku, Tokyo; CEO Saneyuki Nagai) announces the launch of the "Japanese 3-Speaker Business Theme Dialogue Audio & Transcript Dataset" through amana images Inc., a subsidiary under its AI training data solution, Qlean Dataset.

This dataset consists of 13 sets of Japanese three-speaker business dialogues recorded in a web conferencing format, paired with high-quality, human-created transcripts. It can be used for developing and fine-tuning speech-based AI models, including ASR domain adaptation and improving LLMs' understanding of business conversations.

What Is a Business Theme Dialogue Audio Dataset?

A business theme dialogue audio dataset is a speech corpus capturing natural, multi-speaker conversations in business contexts such as investment, insurance, and negotiations. It serves as machine learning data for ASR domain adaptation, multi-speaker speech recognition, and enhancing LLMs' comprehension of business dialogues.

Overview of the "Japanese 3-Speaker Business Theme Dialogue Audio & Transcript Dataset"

This dataset records business dialogues among 13 diverse groups of Japanese speakers (varying by gender and age), each comprising three participants in a web conferencing setting. The dialogues feature spontaneous speech on business-related topics, including specialized terminology, speaker turn-taking, and fillers, closely reflecting real-world working environments. Transcripts are manually created—without using automated transcription—ensuring high accuracy by eliminating mis-transcribed technical terms, missing fillers, and speaker boundary errors. This guarantees reliable WER and CER evaluation and high-quality training data.

Data Type

Audio (3-speaker dialogue format)

Speakers

Japanese speakers with diverse gender and age profiles (13 groups)

Duration / Volume

Approx. 25 hours (63 files) / Approx. 55GB

Data Format

mp3

Sampling / Bitrate

48kHz / 192kbps, stereo recording

Content Features

Three-speaker business dialogues (e.g., investment, insurance) simulating web conferencing environments; approx. 90 minutes per session

Usage

Commercial use permitted · Research use permitted · Eligible for free academic use

Transcripts

High quality ensured through manual transcription

Sample data available here: https://qleandataset.visual-bank.co.jp/lineup/ds-050

Frequently Asked Questions (FAQ)

Q. How does a 3-speaker setup differ from 2-speaker data? A. The speaker turn-taking patterns become more complex, closely replicating real-world business meetings. It is also effective for evaluating the generalization performance of ASR models trained on 2-speaker data.

Q. Can this dataset be used for ASR or LLM development in finance and insurance? A. Yes. The human-created transcript and audio pairs allow direct use in domain adaptation fine-tuning (e.g., Whisper LoRA) and as SFT or evaluation data for finance-specific LLMs, without noise from automated transcription errors.

Q. Is this suitable for developing AI for meeting minutes or meeting summarization? A. Yes. With long-form recordings of approximately 90 minutes per session and human-created transcripts, it enables high-precision use as SFT data for summarization, meeting minutes generation, and action item extraction tasks.

Q. Do you support custom recording requests? A. Yes. We support custom data collection with specified industry sectors, job roles, age groups, and dialogue scenarios.

Use Case Examples

Fine-tuning for Business-Domain ASR Using human-created transcripts paired with audio, this dataset supports LoRA or full fine-tuning of models like Whisper and ESPnet. Free from noise caused by automated transcription, it enables highly reliable WER and CER evaluation.

Multi-Speaker ASR Performance Evaluation The 3-speaker audio, including speaker turn-taking, overlapping speech, and fillers, allows testing ASR model robustness against complex speech patterns that cannot be replicated with 2-speaker datasets.

LLM-Based Business Dialogue Summarization and Meeting Minutes Generation Human-created transcripts support the construction of SFT data for summarization, meeting minutes generation, and action item extraction. The long-form nature (approx. 90 minutes/session) also enables evaluation and enhancement of long-context processing capabilities.

About Qlean Dataset

Qlean Dataset is an AI training data solution provided by amana images Inc., a wholly owned subsidiary of Visual Bank. With over 40 years of experience in legally acquiring, managing, and distributing photos, videos, and audio content, amana images has built its business on protecting data rights—forming the foundation of Qlean Dataset.

Each data item comes with clearly documented sources and rights clearance, delivered ready for direct use in model training. The platform has established a track record of large-scale deliveries to domestic and international foundation model developers. Currently supporting audio, image, video, 3D, and text modalities, Qlean Dataset continuously expands its offerings through collaborations with data holders and media organizations worldwide. Custom recording and data collection services are also available.

Qlean Dataset website: https://qleandataset.visual-bank.co.jp/

AI Data Recipe: https://qleandataset.visual-bank.co.jp/lineup

Contact

Visual Bank Co., Ltd.

As a startup building and providing next-generation data infrastructure to maximize AI development capabilities, Visual Bank operates under the mission of "unlocking the potential of all data." It wholly owns amana images Inc., which provides AI-assisted tools supporting manga creators' creativity through "THE PEN" and offers AI training dataset development services via "Qlean Dataset (Qlean Dataset)."

Additionally, Visual Bank has been selected for the national R&D program "GENIAC," accelerating its initiatives toward social implementation.

CEO: Saneyuki Nagai

Address: C-Cube Minami-Aoyama Building 6F, 7-1-7 Minami-Aoyama, Minato-ku, Tokyo 107-0062

Visual Bank corporate URL: https://visual-bank.co.jp/

amana images corporate URL: https://amanaimages.com/about/