Hugging Face's logo

Datasets:
OKHand
/
tw-daily-dialogue2Voice

Modalities:
Audio
Text
Formats:
parquet
Languages:
Chinese
Size:
10K - 100K
Tags:
text-to-speech
Synthetic
dialogue
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
Community
1

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.

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

TW Daily Dialogue Audio Dataset

TW 日常對話語音資料集

This dataset, tw-daily-dialogue-audio, contains synthetic Mandarin daily-life dialogues generated using CosyVoice, specifically FunAudioLLM/Fun-CosyVoice3-0.5B-2512.
本資料集 tw-daily-dialogue-audio 收錄由 CosyVoice 所生成的中文（日常）合成對話語音資料，實際使用的模型為 FunAudioLLM/Fun-CosyVoice3-0.5B-2512。

The data is organized in a multi-turn dialogue format and is suitable for research and applications such as Text-to-Speech (TTS), spoken dialogue systems, and speech understanding.
資料以多輪對話形式組織，適用於 語音合成（TTS）、語音對話系統 與 語音理解 等研究與應用。

Quality Control / Filtering

品質控管與資料過濾

All samples in this dataset are automatically filtered using ASR, and only utterances with Character Error Rate (CER) < 15% are retained.
本資料集所有樣本皆經由 自動語音辨識（ASR）進行過濾，僅保留 字元錯誤率（CER）小於 15% 的語音資料。

The ASR system used for transcription is MediaTek-Research/Breeze-ASR-25.
本資料集所使用的自動語音辨識模型為 MediaTek-Research/Breeze-ASR-25。

The CER is computed between the ground-truth text (text) and the ASR-transcribed text (pred_text) and is intended only as an automatic quality filter, not as human-level validation.
CER 的計算方式為標準文字（text）與 ASR 轉錄結果（pred_text）之間的差異，僅作為自動化品質篩選依據，不代表人工或主觀聽感驗證。

⚠️ This dataset is NOT a human-validated or manually verified TTS dataset.
⚠️ 本資料集並非經過人工驗證或主觀評分的 TTS 資料集。

Acknowledgements

致謝

We would like to thank the following contributors for their support and contributions to this project:
我們誠摯感謝以下貢獻者對本專案的支持與投入：

  • Timbre Seed: Contributed by Jaron
    音色種子（Timbre Seed）： 由 Jaron 提供

  • Dataset: Contributed by Liang Hsun
    資料集製作（Dataset）： 由 Liang Hsun 貢獻

  • This dataset is contributed to Twinkle AI
    本資料集貢獻給 Twinkle AI

Schema

資料結構說明

  • dialogue_id: Identifier of a dialogue session
    dialogue_id：對話工作階段的識別碼

  • turn_id: The order of the utterance within a dialogue
    turn_id：該句在對話中的順序

  • audio: Synthetic speech audio (24kHz) generated by Fun-CosyVoice3-0.5B-2512
    audio：由 Fun-CosyVoice3-0.5B-2512 生成的合成語音（24kHz）

  • text: Ground-truth transcript used for speech generation
    text：用於語音生成的標準文字內容

  • pred_text: ASR-transcribed text generated by Breeze-ASR-25
    pred_text：由 Breeze-ASR-25 產生的 ASR 轉錄文字

  • cer: Character Error Rate between text and pred_text
    cer：text 與 pred_text 之間的字元錯誤率（CER）

Usage

使用方式 

from datasets import load_dataset

ds = load_dataset("OKHand/tw-daily-dialogue2Voice")

# Inspect a sample (grouped by dialogue)
# 查看資料範例（以對話為單位）
print(ds[0])
Downloads last month
4