LibriSpeech

Panayotov, Chen, Povey and Khudanpur (Johns Hopkins CLSP/HLTCOE), ICASSP 2015. T1 — the corpus paper itself, read in full from the PDF. The benchmark almost every WER quoted in this spoke is measured on, and the first of the three datasets the wiki was citing without holding.

What it is

1000 hours of read English at 16 kHz, cut from LibriVox audiobooks — a volunteer project with roughly 8000 public-domain recordings at the time, mostly read from Project Gutenberg texts. Released CC BY 4.0, with language-model training data, pre-built LMs and Kaldi recipes alongside. Its claimed advance over VoxForge (~100 hours, badly imbalanced by gender and per-speaker duration) is scale plus balance.

The build is an alignment problem, not a recording session: audio is aligned against the book text in two decoding passes, the second speaker-adapted with fMLLR, and mismatches are detected and dropped — “since the original amount of audio in the audiobooks is so large, we can afford to lose a certain percentage of it.” About 1200 hours of aligned audio came out of roughly 65 hours of compute on two EC2 cc2.8xlarge instances. Segmentation differs by purpose: training splits on any silence longer than 0.3 s, test data only where a silence coincides with a sentence break.

The subsets, and what “clean” actually means

“Clean” is not an acoustic measurement. An acoustic model trained on WSJ’s si-84 subset transcribed the corpus; speakers were ranked by the WER of that model’s output against the book text and “divided roughly in the middle, with the lower-WER speakers designated as ‘clean’ and the higher-WER speakers designated as ‘other’.” The dev/test sets for “other” were then chosen deliberately to be more challenging, not at random.

That definition should change how this spoke reads a test-clean number. The easy half of the corpus is the half a 2015 WSJ-trained model already found easy — a mix of recording quality, accent proximity to US English, and speaking style, bundled into one label by the errors of a specific old system. A model reporting 2% on test-clean is reporting on speakers selected for being recognizable.

Why it matters here

Every comparison the spoke holds — open-asr-models-2026-comparison, stt-apis-comparison, tts-benchmarks — leans on LibriSpeech numbers. Two limits are now on the record: it is read speech from audiobooks, so it says nothing about conversational overlap, disfluency or far-field capture; and it is English only, which is what common-voice and fleurs exist to address.