Librispeech: An ASR corpus based on public domain audio books

Key Points

• To introduce and evaluate LibriSpeech, a large-scale public domain read English speech corpus designed for training and benchmarking automatic speech recognition systems.
• Extracted and prepared 1,000 hours of read English speech sampled at 16 kHz derived from LibriVox public domain audiobooks.
• Constructed complementary language-model training data, pre-built language models, and Kaldi training scripts for open distribution.
• Trained acoustic models on LibriSpeech and evaluated their performance against standard Wall Street Journal (WSJ) test sets.
• Acoustic models trained on LibriSpeech achieved lower word error rates on WSJ test sets than models trained directly on the WSJ corpus.
• Released a publicly accessible 1,000-hour speech recognition benchmark with accompanying language models and baseline recipes.