cdminix/libritts-aligned

Dataset Overview

Name: LibriTTS Corpus with Forced Alignments

Description: This dataset contains forced‑alignment information for speech data, suitable for ASR and TTS tasks.

Dataset Details

Language: English (en)

Tags:

• speech
• audio
• automatic-speech-recognition
• text-to-speech

License: CC-BY-4.0

Task Categories:

• Automatic Speech Recognition
• Text‑to‑Speech

Dataset Contents:

• Each entry includes an audio file ID, speaker information, transcript, start and end times, phonemes with durations, and the audio file path.
• Phonemes are represented using the International Phonetic Alphabet (IPA) and durations are given in frames.

Dataset Splits:

• train: All training data except one sample per speaker reserved for validation.
• dev: One sample per speaker for validation.
• train.clean.100, train.clean.360, train.other.500: Training data extracted from different LibriSpeech subsets.
• dev.clean, dev.other: Validation data extracted from different LibriSpeech subsets.
• test.clean, test.other: Test data extracted from different LibriSpeech subsets.

Environment Variables:

• LIBRITTS_VERBOSE: Controls verbosity of the dataset creation process.
• LIBRITTS_MAX_WORKERS: Sets the maximum number of worker threads for alignment creation.
• LIBRITTS_PATH: Sets the download path for LibriTTS data.

Usage Requirements

Software Dependencies:

• pip install alignments phones (required)
• pip install speech-collator (optional)

Data Collator:

• A data collator is provided for creating training batches.
• Install via pip install speech-collator; supports custom speaker2idx and phone2idx mappings.

Citation Information

References:

• LibriTTS: A Corpus Derived from LibriSpeech for Text‑to‑Speech
• Montreal Forced Aligner: Trainable text‑speech alignment using Kaldi
• Evaluating and reducing the distance between synthetic and real speech distributions