AnimeVox: Character TTS Corpus

🗣️ Dataset Overview

AnimeVox is an English Text-to-Speech (TTS) dataset featuring 11,020 audio clips from 19 distinct anime characters across popular series. Each clip includes a high-quality transcription, character name, and anime title, making it ideal for voice cloning, custom TTS model fine-tuning, and character voice synthesis research.

The dataset was created and processed using TTSizer, an open-source tool that automates creating high-quality TTS datasets from raw media.

Watch the Demo Video:

📊 Dataset Statistics

• Total samples: 11,020
• Characters: 19
• Anime series: 15
• Audio format: 44.1kHz mono WAV
• Storage size: ~3.5GB

🎧 Dataset Structure

• Instances: Each sample is a dictionary with the following structure:

  {
    "audio": {"path": "...", "array": ..., "sampling_rate": 44100},
    "transcription": "English text spoken by the character.",
    "character_name": "Character Name",
    "anime": "Anime Series Title"
  }
  

• Fields:
  • audio: Audio object (44.1kHz).
  • transcription: (str) English transcription.
  • character_name: (str) Name of the speaking character.
  • anime: (str) Anime series title.
• Splits: A single train split with all 11,020 samples from 19 characters.

🛠️ Dataset Creation

Source

Audio clips were sourced from official English-dubbed versions of popular anime series. The clips were selected to capture diverse emotional tones and vocal characteristics unique to each character.

Processing with TTSizer

This dataset was generated using TTSizer, which offers an end-to-end automated pipeline for creating TTS-ready datasets. Key features utilized include:

• Advanced Multi-Speaker Diarization: To accurately identify and segment speech for each of the characters, even in complex audio environments.
• State-of-the-Art Model Integration: Leveraging models such as MelBandRoformer (for vocal separation), Gemini (for diarization), CTC-Aligner (for precise audio-text alignment), and WeSpeaker (for speaker embedding/verification).
• Quality Control: Implementing automatic outlier detection to flag and help refine potentially problematic audio-text pairs, ensuring higher dataset quality.

The tool's configurable nature allowed for fine-tuning the entire process to suit the specific needs of this anime voice dataset.

📜 Licensing & Usage

• License: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

🚀 How to Use

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("taresh18/AnimeVox")

# Access the training split
train_data = dataset["train"]

# Print dataset information
print(f"Dataset contains {len(train_data)} samples")

# Access a specific sample
sample = train_data[0]
print(f"Character: {sample['character_name']}")
print(f"From anime: {sample['anime']}")
print(f"Transcription: {sample['transcription']}")