Hugging Face's logo

Datasets:
Databoost
/
TTS_Multilingual_Data

Modalities:
Audio
Tabular
Text
Formats:
parquet
Size:
1K - 10K
Libraries:
Datasets
pandas
Dataset card Data Studio Files Community
2
TTS_Multilingual_Data / README.md
Databoost's picture
Databoost
Update README.md
27aad4e verified
preview code
|
raw
history blame contribute delete
3.15 kB

Documentation Dataset: TTS_Multilingual_Data

Dataset Summary

This large-scale multilingual corpus is designed for linguistic analysis and the development of speech processing models. It supports tasks such as Text-to-Speech (TTS), Automatic Speech Recognition (ASR), and speaker identification. Structured in Parquet format, it serves as a key resource for training and evaluating models, using metrics tailored to ASR and speech technologies.

Thematic Categories

Our dataset is organized into the following thematic categories. Please note that all audio files have a maximum duration of 20 seconds.

Speeches & Conferences

  • "TED talk"
  • "political speech"
  • "interview"
  • "podcast"

Conversations & Dialogues

  • "phone conversation"
  • "spontaneous dialogue"
  • "group discussion"
  • "audio interview"

Media Content & Entertainment

  • "radio clip"
  • "radio commentary"
  • "audio narration"

Instructions & Voice Assistants

  • "voice commands"
  • "voice assistant"
  • "audio notification"
  • "automated message"

Informal Language & Common Expressions

  • "slang"
  • "French expressions"
  • "colloquial language"
  • "youth speech"
  • "emotions in speech"

Accessibility & Inclusion

  • "speech with an accent"
  • "elderly voices"
  • "children speaking"

Literature & Culture

  • "literature"
  • "tale"
  • "fable"
  • "poetry"
  • "novel excerpt"

Supported Tasks

  • Text-to-Speech (TTS): The dataset can be used to train models for generating speech from text.
  • Automatic Speech Recognition (ASR): The dataset can be used to train models for transcribing speech to text. The most common evaluation metric is the Word Error Rate (WER).
  • Speaker Identification: The dataset supports tasks related to identifying speakers based on their voice.

Dataset Structure

Organisation of the Project

The dataset, TTS_Multilingual_Data, is organized as follows: containing one subfolder. The train subfolder includes data files in Parquet format (e.g., data1.parquet), while the audio subfolder contains audio files in WAV format (e.g., audio1.wav). Additionally, a readme.md file at the root level provides detailed information about the dataset's content and usage.

Columns

  • audio_path (string): Path to the audio file.
  • text (string): Ground truth transcription.
  • duration (float64): Duration of the audio file in seconds.
  • speaker_id (string or int): Identifier for the speaker.
  • audio_format (string): Format of the audio file (e.g., WAV, MP3).
  • sampling_rate (int): Sampling rate of the audio file.
  • language (string): Language of the transcription.
  • gender (string): Gender of the speaker (if available).

File Format

The dataset is delivered in Parquet format, optimized for efficient storage and processing.

8. Contact

For inquiries, please contact:

Citations Information

If you use this dataset, please cite it as follows:

@article{
  title={TTS_Multilingual_Data},
  author={Databoost},
  year={2025}
}