This repository is a Private Research Fork of the MinSpeech corpus. It is maintained strictly for individual research purposes, specifically for fine-tuning Automatic Speech Recognition (ASR) and Speech-to-Text Translation (S2TT) models.

1. Ownership & Licensing

Annotations & Metadata: The transcriptions and segment metadata are derived from the MinSpeech project (Interspeech 2024). These elements are used under the CC BY-NC-SA 4.0 license.
Audio Content: We do not claim ownership of the audio data. All audio files are sourced from public YouTube content. The copyright for the raw audio remains entirely with the original content creators (YouTube uploaders).
Usage Intent: This data is processed and stored here solely for non-commercial, non-expressive machine learning training. No audio data is intended for redistribution or public performance.

2. Fair Use & Compliance Statement

Following the copyright regulations of Mainland China, Taiwan, the US, and Australia regarding AI research:

Non-Infringement: This repository does not seek to infringe upon the rights of the original creators. By keeping this repository Private, we ensure that no unauthorized public distribution of copyrighted audio occurs.
Transformation: The audio is utilized to extract statistical linguistic features for S2TT/ASR model weights, which is considered a transformative, non-consumptive use under research exemptions.

Dataset Description

This private version contains cleaned and pre-processed samples from the MinSpeech corpus to optimize training efficiency.

Original Paper: MinSpeech: A Corpus of Southern Min Dialect for Automatic Speech Recognition (Lin et al., 2024).
Modifications: * Audio standardized to 16kHz Mono WAV.
- VAD-based silence removal and noise reduction.
- Alignment verified between YouTube segments and transcriptions.

License and Usage Restrictions

This private dataset repository contains a compilation of data from multiple sources with different licensing terms.

1. Annotations & Metadata (CC BY-NC-SA 4.0)

The text transcriptions, timestamps, and translation labels are derived from the MinSpeech project.

These elements are strictly governed by the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license.
Any redistribution of these specific metadata files must adhere to this license.

2. Audio Data (Original Creator Copyright)

The audio files in this repository are processed derivatives of public content from YouTube.

No License Granted: We do not hold the copyright to these audio recordings and cannot grant any license for their use.
Third-Party Rights: Copyright remains with the original YouTube content creators.
Research Exception: This data is stored here in a Private capacity for non-commercial machine learning research (Fine-tuning ASR/S2TT models) under "Fair Use" or "Research/Study" exemptions as defined in relevant jurisdictions (Mainland China, Taiwan, US, Australia).

3. Usage Policy

By accessing this private repository, you agree that:

You will not redistribute the audio files.
You will use this data solely for non-commercial academic research.
You acknowledge that the model trained on this data (ASR/S2TT) is a statistical representation and does not contain the original audio.

Downloads last month: 294

Total file size:

285 GB