Skip to main content

> ML_BENCHMARK // LIBRISPEECH-WORD-ERROR-RATE_v1.0

LibriSpeech ASR Word Error Rate (WER)

Speech Recognition · task-speech-recognition · near-saturation

Speech Recognitionnear-saturationmoderate
Official LeaderboardBack to Benchmarks

Evaluation Protocol

Standard greedy or beam search decoding evaluated on test-clean and noisy test-other splits without language model rescoring.

Baseline & Metrics

Canonical Baseline:Kaldi (hybrid): 3.8% clean | wav2vec 2.0: 1.8% clean / 3.9% other | Whisper Large-v3: 1.5% clean / 2.7% other
Evaluated Metrics:
Word Error Rate (WER % on test-clean and test-other)

Contamination & Leakage Risks

Clean/other splits are strictly speaker-independent.

Reproducibility Concerns

Text normalization rules (capitalization, punctuation, number-to-words) alter WER by up to 2%.

CONNECTED DATASETLibriSpeech ASR Corpus (Panayotov et al. 2015)Speech Recognition & Acoustic Modeling · 1,000 hours of 16kHz read English audiobooks with aligned text transcripts
View Dataset Spec