Dataset
LibriSpeech
1000-hour English audiobook corpus, the standard speech-recognition benchmark.
Definition
LibriSpeech is a 1000-hour corpus derived from LibriVox audiobooks, with clean and noisy test sets (test-clean, test-other). Word error rate on LibriSpeech is the canonical metric for English ASR systems, including Whisper variants.
Common use cases
- ASR benchmarks
- Speech research
- Audio pre-training