Voicesdatasets/Character_Voices

Dataset Summary

The Voices.com Expressive & Emotional Voice Character Dataset is a unique, custom-curated collection of professional recordings designed for conversational AI, expressive TTS, ASR, and emotion recognition. Unlike web-scraped corpora, every recording was produced with full participant consent, licensed for AI training, and verified for quality.

Expressive & Emotional Voice Character Dataset

---

Summary

A unique, custom-curated dataset combining expert voice actors with multi-character performances — delivering high-quality training data for conversational AI, character-driven assistants, and next-gen voice technologies.

---

What’s Inside

• 600+ hours ready for immediate delivery
• 450+ unique character styles
• Rich expressive variety with 19 tones, 35 emotions, and 21 paralinguistic cues
• Strong creative direction and scripting ensure authentic performances
• JSON-formatted metadata for easy integration

---

License Terms

• Licensed for model training only
• Non-exclusive license — Voices.com retains ownership
• See the included Voices_Sample_LICENSE.txt for evaluation-only terms

---

How It Helps Your Team

• Diverse vocal performances → more expressive AI responses
• Emotion, tone, and character tags → nuanced generation capabilities
• High-quality audio → cleaner model training input
• Faster time-to-value → ingest immediately, minimal prep needed
• Multiple English variants → broader coverage & adaptability

---

Contact

• Email: [email protected]
• Phone: +1-647-458-7602
• Website: Voices.com

What’s Inside

This repository contains a limited sample of the full dataset. The complete collection includes:

• 600+ hours of audio recorded in studio-grade conditions
• 450+ unique character styles across a wide range of roles
• Rich expressive variety with 19 tones, 35 emotions, and 21 paralinguistic cues
• 537 utterances per character, enabling variety in training and testing
• Strong creative direction and scripting to ensure authentic, consistent performances

Dataset Schema

Field	Type	Description
file_name	string	Audio filename with extension (e.g., Hamlet_001.wav).
audio_path	string	Relative or absolute path to the audio file.
audio	wav/flac	Studio-quality audio recording.
transcript	string	Human-verified transcript of the audio (or auto-aligned if not available).
speaker_id	string	Unique anonymized speaker identifier (e.g., spk_045).
gender	string	Optional speaker gender (male, female, nonbinary, unspecified).
age_range	string	Optional speaker age bucket (e.g., 25-34, 35-44, child, teen).
language	string	Language or BCP-47 code (e.g., en, en-US).
accent	string	Optional accent/locale tag (e.g., North American, RP English).
character_style	string	Character role/style performed (e.g., Hamlet, Narrator).
emotion	string	Emotion label (from 35 classes, e.g., contemplative, neutral).
tone	string	Expressive tone label (from 19 classes, e.g., measured, storytelling).
paralinguistics	string/arr	Non-verbal cues (21 labels, e.g., laugh, pause, sigh).
duration_sec	float	Length of audio in seconds.
sampling_rate	int	Sampling frequency in Hz (e.g., 48000).
split	string	Dataset partition (train, validation, test).
script_id	string	Stable identifier for the script (helps prevent text leakage).
consent_id	string	Internal reference to participant consent/session (kept private).

Dataset Splits

The full dataset is structured into standard partitions for training, validation, and testing.
Splits are defined at the speaker/session level to prevent overlap between training and evaluation voices, ensuring no speaker or identical script utterance appears in more than one split.

Split	Hours	Percentage	Notes
Train	~400h	80%	Used for model training
Validation	~50h	10%	Used for hyperparameter tuning & QA
Test	~50h	10%	Used only for final model evaluation

Note: This preview repository does not include full splits. Commercial license holders receive the complete dataset with predefined train/validation/test partitions to ensure reproducibility and benchmark integrity.

Data Instances

The examples below illustrate how metadata records are structured.
They use public domain texts (Shakespeare, Aesop’s Fables) for demonstration only.
👉 Note: Voices.com’s commercial datasets use proprietary scripts with richer annotation. These examples are placeholders for documentation purposes.

---

Human-Verified Example

(Full metadata, including human-verified transcript segments)

{
  "file_name": "Fable_TortoiseHare_014.wav",
  "audio_path": "data/audio/Fable_TortoiseHare_014.wav",
  "sampling_rate": 48000,
  "duration_sec": 22.85,
  "speaker_id": "spk_132",
  "gender": "male",
  "age_range": "35-44",
  "language": "en-US",
  "accent": "North American",
  "script_id": "public_domain_fables",
  "split": "validation",
  "character_style": "Narrator",
  "emotion": "neutral",
  "tone": "storytelling",
  "paralinguistics": ["pause"],
  "script": {
    "title": "The Tortoise and the Hare",
    "content": "The hare laughed at the slow tortoise, but the race had only just begun.",
    "tone": "storytelling",
    "emotion": "neutral",
    "character_type": "Narrator",
    "direction": "Classic narration"
  },
  "transcript": {
    "human_verified_content": [
      {
        "start": 0.0,
        "end": 8.1,
        "speaker": "Narrator",
        "text": "The hare laughed at the slow tortoise,",
        "manner": "storytelling",
        "emotion": "neutral"
      },
      {
        "start": 8.1,
        "end": 15.6,
        "speaker": "Narrator",
        "text": "but the race had only just begun.",
        "manner": "storytelling",
        "emotion": "neutral",
        "paralinguistics": "pause"
      }
    ],
    "normalized": true
  }
}


*(Basic metadata only — no transcript segments included)*

```json
{
  "file_name": "Hamlet_001.wav",
  "audio_path": "data/audio/Hamlet_001.wav",
  "sampling_rate": 48000,
  "duration_sec": 15.20,
  "speaker_id": "spk_045",
  "language": "en",
  "accent": "British",
  "script_id": "public_domain_hamlet",
  "split": "train",
  "character_style": "Hamlet",
  "emotion": "contemplative",
  "tone": "measured",
  "script": {
    "title": "Hamlet, Act III, Scene I",
    "content": "To be, or not to be, that is the question...",
    "tone": "measured",
    "emotion": "contemplative",
    "character_type": "Hamlet",
    "direction": "Delivered as a soliloquy"
  }
}

## Collection Process & Consent

All recordings in this dataset were sourced directly through Voices.com’s global network of professional voice actors.  

- **Participant Consent**  
  Every contributor provided explicit, written consent for their recordings to be used in AI training and research. Consent includes approval for derivative works and guarantees that contributors were fully informed of the purpose of the project.  

- **Licensing for AI**  
  All recordings are licensed under commercial agreements that grant rights for model training, fine-tuning, evaluation, and related research. Voices.com retains ownership of the underlying recordings, and redistribution is restricted without a valid license.  

- **Ethical Sourcing**  
  No content was scraped from the internet, extracted from unauthorized sources, or generated synthetically. Every voice sample was recorded in controlled, high-quality conditions by consenting professionals.  

- **Quality Assurance**  
  Our team oversees every step of the process — from script development and talent casting to studio-grade recording and human-verified annotation. This ensures not only technical quality, but also compliance with ethical and legal standards.  

👉 This transparent process guarantees that the dataset is **safe, ethical, and commercially viable** for building next-generation AI applications.

# The World's Best Multi-Character Voice Dataset

A unique, custom-curated dataset combining expert voice actors with multi-character performances, delivering high-quality training data for conversational AI, character-driven assistants, and next-gen voice technologies.

## What's Inside

This repository contains a limited sample of our full dataset. The full dataset is a meticulously crafted collection of high-quality English-language recordings from professional voice talent.

* **Content:**
    * Over 600 hours of audio ready for immediate delivery.
    * More than 450 unique character styles.
    * Rich expressive variety with 19 tones, 35 emotions, and 21 paralinguistic cues.
    * Extensive coverage with 537 unique utterances per character, providing ample variety for training and testing.
    * Strong creative direction and scripting ensure consistent and authentic performances.

* **Data Tags:**
    * Character, text, tone, and emotion tags are included.
    * Delivered in JSON format for easy integration.

## How to Get Started

This gated repository is a preview of our premium offering. To access the files in this repository, please click the "Access repository" button and agree to the license terms.

For commercial inquiries and full dataset access, please contact us directly at [email protected]

## Intended Uses and Limitations

**Intended Uses**  
This dataset is designed for training and evaluating models in:  
- Automatic Speech Recognition (ASR)  
- Text-to-Speech (TTS) and expressive voice synthesis  
- Emotion and paralinguistic classification  
- Conversational AI and character-driven assistants  


## Licensing and Terms of Use

This dataset is provided under a custom, non-exclusive license for model training only.  

**This sample dataset is provided for non-commercial evaluation purposes only. Any other use, including redistribution, model training, or commercial exploitation, requires a separate license agreement with Voices.com.**

Any commercial use, reproduction, or redistribution of this data without a valid commercial license is strictly prohibited.

See the included `Voices_Sample_LICENSE.txt` file for the full terms.


**Limitations / Not Intended For**  
- Voice impersonation or cloning of real individuals  
- Biometric identification or surveillance  
- Deepfakes, misinformation, or deceptive content  
- Any use outside the scope of licensed AI training and research

## How It Helps Your Team

* **Diverse Vocal Performances:** Train models for better expressiveness and more nuanced AI responses.
* **High-Quality Audio:** Start with a cleaner training input for superior model performance.
* **Faster Time to Value:** Instantly ingest the dataset to begin training without extensive data preparation.
* **Broad Coverage:** Multiple English variants provide adaptability for a wider range of applications.

## Contact Us

Ready to secure the first 600 hours now and scale effortlessly as the dataset expands?

* **Email:** [email protected]
* **Phone:** 1-647-458-7602
* **Website:** www.voices.com

@dataset{voices2025character,
  title   = {Expressive & Emotional Voice Character Dataset},
  author  = {Voices.com Inc.},
  year    = {2025},
  url     = {https://huggingface.co/datasets/Voicesdatasets/CharacterDataset}
}


---
© 2005-2025 Voices.com Inc. All Rights Reserved