Skip to main content
v2026.11,986 entries · CC-BY 4.0

LibriSpeech: Audio Corpus Guide and ASR Benchmarks

A comprehensive guide to the LibriSpeech 1,000-hour ASR corpus, covering dataset partitions, alignment pipelines, acoustic benchmarks, and lab protocols.

Written and maintained by CASRAI Editorial Board

Published

On this page

LibriSpeech is an open-access corpus of approximately 1,000 hours of 16 kHz read English speech derived from public-domain LibriVox audiobooks and Project Gutenberg texts, designed specifically for training and benchmarking automatic speech recognition systems under the Creative Commons Attribution 4.0 International license.

Definition and Scope of the LibriSpeech Corpus

Introduced by Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur at the 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), LibriSpeech was developed at Johns Hopkins University to resolve a critical bottleneck in speech recognition research: the scarcity of large-scale, standardized, and freely redistributable acoustic training corpora. Prior to its release, academic laboratories and independent researchers were largely dependent on proprietary, fee-gated datasets such as the Wall Street Journal (WSJ0 and WSJ1) corpora or smaller collections like TIMIT, which provided only a few hours or tens of hours of speech under restrictive licensing. Alternative community-driven collections such as VoxForge offered open audio but suffered from severe imbalances in speaker demographics, varying acoustic fidelity, and limited duration.

LibriSpeech established an open standard by sourcing thousands of volunteer-read audiobooks from the LibriVox project, matching the spoken audio against verified literary texts cataloged by Project Gutenberg. The corpus encompasses 2,484 unique speakers across 1,000 hours of 16 kHz, 16-bit audio, compressed losslessly in FLAC format. By operating under the Creative Commons Attribution 4.0 (CC BY 4.0) license and hosted permanently on the Open Speech and Language Resources repository under identifier SLR12, LibriSpeech aligns directly with modern FAIR data principles, ensuring that machine learning researchers can find, access, interoperate with, and reuse standardized acoustic data without commercial encumbrance. Unlike brief research deposits, its release established a landmark precedent in how speech data should be documented, contrasting with the structural divergence between dedicated data papers versus repository dataset records.

Within signal processing and electrical engineering, LibriSpeech spans both acoustic modeling and statistical language modeling. In addition to audio archives, the project released pre-built language models, pronunciation dictionaries based on CMUdict, and Kaldi-compatible recipe scripts hosted on OpenSLR under resource SLR11. This holistic infrastructure transformed LibriSpeech into the universal baseline benchmark against which novel acoustic architectures—ranging from early deep neural network hybrid models to end-to-end recurrent neural networks, Connectionist Temporal Classification (CTC) networks, Conformer architectures, and self-supervised foundation transformers—are evaluated.

Corpus Architecture and Partition Structure

A core architectural strength of LibriSpeech is its rigorous, stratified partitioning scheme. Sourcing 1,000 hours of volunteer audio introduces inherent heterogeneity: volunteers utilize diverse recording hardware, acoustic environments vary from treated home studios to untreated domestic rooms with noticeable reverberation, and speaker accents exhibit regional diversity. Rather than aggregating the data into an undifferentiated training pool, the creators systematically partitioned the corpus into distinct subsets calibrated by recording quality and speaker difficulty.

To establish an empirical boundary between high-fidelity and acoustically challenging recordings, Panayotov and colleagues trained an acoustic model on the Wall Street Journal si-84 dataset and used it to transcribe the candidate LibriSpeech recordings alongside a bigram language model estimated from the respective book texts. By computing the Word Error Rate (WER) of these automatic transcriptions relative to the ground-truth book text, the authors quantified the intrinsic acoustic recognition difficulty of each speaker. Speakers exhibiting lower WERs were categorized into the “clean” pool, characterized by lower ambient noise and accents closely aligned with General American English. Speakers exhibiting higher error rates were assigned to the “other” pool, capturing higher background noise, varied room acoustics, or non-standard accents.

The corpus is partitioned into seven standardized subsets designed to prevent speaker leakage between training, development, and testing phases:

  • train-clean-100: Contains 100.6 hours of speech from 251 speakers (125 female, 126 male). Audio duration per speaker is capped at 25 minutes to prevent over-representation of prolific volunteers. This subset serves as the standard entry-level training benchmark for rapid prototyping and low-resource acoustic modeling.
  • train-clean-360: Encompasses 363.6 hours of speech across 921 speakers (439 female, 482 male), with per-speaker audio similarly constrained to 25 minutes. When combined with train-clean-100, it forms the canonical 460-hour clean training set.
  • train-other-500: Comprises 496.7 hours of speech from 1,166 speakers (564 female, 602 male), with per-speaker duration capped at 30 minutes. This partition introduces acoustic variability, room reverberation, and diverse articulation styles, enabling models to build robustness against real-world transmission channels.
  • dev-clean: A development validation subset of 5.4 hours drawn from 40 speakers (20 female, 20 male), providing approximately 8 minutes of audio per speaker. It is used for hyperparameter tuning, language model weight optimization, and checkpoint validation under clean conditions.
  • test-clean: The primary evaluation benchmark consisting of 5.4 hours from 40 distinct speakers (20 female, 20 male), providing approximately 8 minutes per speaker with zero speaker overlap with training or development sets.
  • dev-other: A development subset of 5.3 hours from 33 speakers (16 female, 17 male), approximately 10 minutes per speaker, drawn randomly from the third quartile of the sorted speaker difficulty list.
  • test-other: The standard robustness test set comprising 5.1 hours across 33 speakers (17 female, 16 male), sampled from the third quartile of difficulty, used to evaluate generalization against challenging acoustic conditions.

Gender balance was enforced across every individual partition, maintaining near-perfect parity between female and male speakers. Total unique speaker count reaches 2,484 individuals (1,201 female and 1,283 male). In total, the full training corpus (train-clean-100, train-clean-360, and train-other-500) totals 960.9 hours of audio, universally referenced as the LibriSpeech 960h benchmark.

Audio Processing and Text-Audio Alignment Pipeline

Transforming raw, unstructured LibriVox audiobook chapters into sentence-level acoustic utterances required an automated, noise-tolerant alignment and segmentation pipeline. The creators implemented a two-stage alignment protocol in Kaldi that remains an exemplar for speech corpus engineering.

In the text preprocessing phase, reference book texts from Project Gutenberg were normalized by converting all characters to uppercase, stripping punctuation marks, and expanding common abbreviations and non-standard numeric tokens into written words. The pronunciation lexicon was derived from the Carnegie Mellon University Pronouncing Dictionary (CMUdict), stripped of numeric lexical stress markers. For out-of-vocabulary (OOV) words absent from CMUdict, phonetic transcriptions were generated automatically using the Sequitur Grapheme-to-Phoneme (G2P) toolkit. Book-specific bigram language models smoothed with Witten-Bell discounting were trained using SRILM to guide initial acoustic decoding.

Long audio chapters, often lasting up to 30 minutes, were aligned through the following multi-stage protocol:

  1. First Decoding and Region Alignment: The audio was decoded using Kaldi’s gmm-decode-faster driven by triphone Hidden Markov Models with Gaussian Mixture Model emissions (HMM-GMM) trained on the VoxForge dataset using Boosted Maximum Mutual Information (BMMI). Features comprised 13 Mel-Frequency Cepstral Coefficients (MFCC) spliced across 7 context frames, projected via Linear Discriminant Analysis (LDA) and rotated using a global Semi-Tied Covariance (STC) transform. The resulting hypothesis was aligned against the normalized book text using the Smith-Waterman local alignment algorithm with tunable error weights. This isolated the maximal contiguous region of similarity and established islands of confidence where phone matches reached 12 or more consecutive phonemes. Audio was then initially split into chunks of 35 seconds or less.
  2. Second Stage Validation Graph: To filter out disfluencies, reader ad-libs, omissions, or G2P pronunciation failures, the authors engineered a custom decoding graph. The graph combined the linear word transcript with a phone-level bigram acceptor configured as a single-state pushdown transducer. The decoder allowed transitions into the phone bigram to represent arbitrary insertions or word substitutions, but constrained re-entry strictly to the adjacent word index. Any utterance whose decoded path deviated from the ground-truth transcript was rejected. This stage employed speaker-adapted acoustic models with feature-space Maximum Likelihood Linear Regression (fMLLR) transforms computed per speaker.
  3. Utterance Segmentation: Validated audio segments were split into final training and evaluation utterances. For the training subsets, splits were placed at any silence interval exceeding 0.3 seconds. For development and test sets, splits were strictly restricted to silence intervals coinciding with sentence boundaries in the reference text. This deliberate constraint ensured that evaluation utterances retained complete grammatical clauses, preventing truncated sentences that would unfairly penalize language model scoring.

Researchers processing raw speech signals in modern computational acoustics often mirror these pre-processing conventions, utilizing specialized signal analysis suites such as the Python audio signal processing library librosa for spectral transformations or Praat phonetic analysis software to inspect formant tracks and pitch contours. When annotating multi-speaker audio or clinical speech, researchers frequently coordinate these segmentations within the ELAN linguistic annotation tool to maintain time-aligned tier structures.

Language Modeling and Decontamination Protocols

Evaluating an automatic speech recognition system requires decoupling acoustic modeling performance from language model memorization. Because Project Gutenberg contains thousands of overlapping literary editions, anthologies, and multi-volume collections, training a statistical language model on raw Gutenberg texts posed a severe risk of data contamination: if test sentences appeared verbatim in language model training corpora, language model perplexity would plummet, masking acoustic recognition failures.

To eliminate evaluation leakage, Panayotov and colleagues developed an inverted index decontamination protocol. The team downloaded approximately 22,000 candidate public-domain books from Project Gutenberg. They first excluded all books containing the development and test texts, as well as any candidate book whose title shared a letter-3-gram cosine similarity greater than 0.7 with an evaluation title. Next, an inverted index of all 5-grams (with stop words deleted) was constructed from the dev and test books. Every candidate book was evaluated against this index; any text where more than 1.0% of its 5-grams matched the evaluation index was discarded. This decontaminated common anthologies, recurring poetry quotations, and shared fairy tales.

Additional heuristic filters removed non-narrative texts, including genomic sequences from the Human Genome Project, astronomical tables, and numeric indexes. The remaining corpus comprised approximately 14,500 books containing 803 million word tokens and 900,000 unique vocabulary types. From this vocabulary, the top 200,000 words were selected as the official lexicon, covering 97.5% of evaluation tokens with an out-of-vocabulary rate of approximately 0.4%. Using SRILM, the authors trained standard 3-gram and 4-gram language models smoothed with Modified Kneser-Ney discounting, achieving test set perplexities of approximately 170 for the 3-gram model and 150 for the 4-gram model (distributed via OpenSLR 11).

Benchmark Performance Evolution: From GMM-HMM to Deep Learning

LibriSpeech has served as the definitive proving ground across three consecutive eras of speech recognition architecture: traditional statistical hybrid modeling, supervised end-to-end deep learning, and self-supervised foundation modeling.

In the original 2015 ICASSP publication, baseline evaluations were established using Kaldi. Acoustic models evaluated included Speaker-Adapted Training (SAT) GMMs and Deep Neural Networks (DNN) utilizing p-norm non-linearities trained on top of fMLLR features. Rescored with the official 4-gram language model, the baseline Kaldi systems recorded the following benchmark error rates:

  • SAT GMM (100h clean): dev-clean 8.19% WER; test-clean 9.32% WER; dev-other 29.31% WER; test-other 31.52% WER.
  • SAT GMM (960h full): dev-clean 7.08% WER; test-clean 8.04% WER; dev-other 21.14% WER; test-other 22.65% WER.
  • p-norm DNN (100h clean): dev-clean 5.93% WER; test-clean 6.59% WER; dev-other 20.42% WER; test-other 22.52% WER.
  • p-norm DNN (960h full): dev-clean 4.90% WER; test-clean 5.51% WER; dev-other 12.98% WER; test-other 13.97% WER.

Between 2016 and 2019, end-to-end sequence-to-sequence paradigms—such as Listen, Attend and Spell (LAS), Connectionist Temporal Classification (CTC), and Recurrent Neural Network Transducers (RNN-T)—displaced hybrid HMM systems. Models such as Jasper and QuartzNet demonstrated that deep 1D convolutional networks trained end-to-end on LibriSpeech 960h could match hybrid pipelines without separate pronunciation lexicons or phonetic alignment graphs.

In 2020, self-supervised representation learning fundamentally altered speech benchmarks. Architectures such as wav2vec 2.0 (Baevski et al.), HuBERT (Hsu et al.), and WavLM (Chen et al.) introduced contrastive and masked prediction objectives over raw speech waveforms. When fine-tuned on LibriSpeech 960h, these models drove test-clean WER below 2.0% and test-other WER below 4.0%. In low-resource benchmarks utilizing only train-clean-100 or the official 10-hour and 1-hour subsets defined by Libri-Light, self-supervised models achieved accuracy levels previously unattainable without hundreds of hours of manual transcription.

Subsequently, Conformer architectures combining multi-head self-attention with depthwise separable convolutions (Gulati et al.), alongside massive weakly supervised models such as OpenAI Whisper (Radford et al.), demonstrated human-level transcription fidelity on LibriSpeech test-clean (often reaching 1.4% to 1.8% WER). Because test-clean is now considered largely saturated by top-tier models, speech laboratories prioritize test-other performance, cross-dataset transfer, and robustness under adverse noise to gauge architectural superiority.

The success of LibriSpeech also catalyzed major derivative corpora that expanded its methodology across scale and language:

  • Libri-Light (Kahn et al., ICASSP 2020): Scaled unannotated LibriVox audio to 60,000 hours, establishing the premier open benchmark for semi-supervised and self-supervised speech representation learning with standardized 10-minute, 1-hour, and 10-hour labeled fine-tuning splits.
  • Multilingual LibriSpeech / MLS (Pratap et al., Interspeech 2020): Extended the LibriSpeech alignment pipeline to 8 European languages (English, German, Dutch, French, Spanish, Italian, Portuguese, and Polish), releasing 50,000 hours of speech under OpenSLR identifier SLR87.
  • LibriCSS (Chen et al., ASRU 2021): Re-recorded overlapping LibriSpeech utterances in physical conference rooms to benchmark continuous speech separation, multi-talker diarization, and meeting transcription.

Applications in Research Laboratories and Speech Technology

In academic laboratories and commercial research divisions, LibriSpeech functions as both a development testbed and an operational standard. Its primary utility lies in providing a reproducible, zero-cost acoustic baseline. When researchers introduce a novel attention mechanism, acoustic feature extraction technique, or loss formulation, reporting evaluation metrics on LibriSpeech dev-clean, test-clean, dev-other, and test-other is mandatory for peer review in leading conferences such as IEEE ICASSP and Interspeech.

For research computing environments, loading LibriSpeech has been standardized across modern machine learning frameworks. In PyTorch, torchaudio.datasets.LIBRISPEECH provides automated archive downloading, integrity checking, and streaming extraction, returning audio waveforms alongside standardized metadata tuples containing sample rate, transcript, speaker ID, chapter ID, and utterance ID. Similarly, Hugging Face Datasets offers the librispeech_asr loader, enabling distributed streaming directly to training workers without requiring local pre-extraction of 60 gigabytes of compressed archives.

To establish rigorous, reproducible experimental protocols when evaluating ASR pipelines on LibriSpeech, research facilities should implement the following operational verification checklist:

  • Text Normalization Consistency: Standardize reference and hypothesis text by removing punctuation, uppercasing all alphabetic characters, and verifying the treatment of contractions (e.g., whether words like “CAN’T” are expanded or tokenized as single words) prior to running Levenshtein WER scoring.
  • Zero Data Leakage Auditing: Ensure that language model training corpora do not incorporate Project Gutenberg books present in the dev-clean, test-clean, dev-other, or test-other evaluation partitions.
  • Stratified Reporting: Always report evaluation metrics separately for test-clean and test-other. Combining both partitions into an unweighted aggregate obscures acoustic robustness regressions.
  • Acoustic Model Checkpointing: Validate training convergence against dev-clean and dev-other separately; avoid selecting checkpoints based solely on clean validation sets if deployment models require noise resistance.
  • Reproducible Rescoring: When employing external language models or rescoring lattices, explicitly document whether evaluations use the official 3-gram, 4-gram, or neural language models (e.g., Transformer-LM), as language model capacity substantially influences final word error rates.

The structured data table below compares the technical parameters, acoustic conditions, and licensing frameworks of LibriSpeech alongside other canonical speech recognition datasets.

Common Misunderstandings and Methodological Pitfalls

Despite its ubiquitous adoption, several persistent misunderstandings compromise experimental validity when researchers utilize LibriSpeech:

  • Assuming Clean Means Studio Grade: The “clean” designation in LibriSpeech does not imply professional anechoic studio recording. LibriVox audio is recorded by volunteers in home environments using consumer-grade USB microphones or entry-level audio interfaces. Clean merely designates recordings that achieved lower WER when evaluated against a baseline Wall Street Journal acoustic model. Mild room reverberation, subtle background hiss, and varying microphone proximity effects remain present even in train-clean-100.
  • Treating LibriSpeech as Spontaneous Conversational Speech: LibriSpeech consists exclusively of read literary prose. It exhibits formal syntax, complete sentences, elevated vocabulary, and deliberate oratorical pacing. It does not contain conversational phenomena such as lexical overlap, interruptions, hesitation markers (e.g., “um”, “uh”), spontaneous self-corrections, or colloquial slang. Models trained exclusively on LibriSpeech suffer substantial degradation when deployed on spontaneous conversational speech or telephone dialogues.
  • Overfitting and Benchmark Saturation on test-clean: With modern deep neural networks achieving WERs below 2.0% on test-clean, performance differences of 0.1% or 0.2% often fall within the margin of label noise, minor phonetic ambiguities, or reader pronunciation quirks. Treating marginal gains on test-clean as evidence of architectural breakthroughs is statistically unsound; researchers must cross-validate improvements on test-other or independent out-of-domain benchmarks.
  • Neglecting MP3 Compression Artifacts: Raw LibriVox recordings were originally compressed as lossy MP3 files prior to their ingestion into the LibriSpeech alignment pipeline, which decoded them to 16 kHz FLAC. While high-bitrate MP3 compression preserves intelligible speech formants, it introduces subtle psychoacoustic quantization artifacts and high-frequency band-limiting. Acoustic researchers developing ultra-high-fidelity neural vocoders or super-resolution audio front-ends should remain mindful of this heritage.
  • Uncontrolled Vocabulary Filtering: The official vocabulary omits numbers and punctuation. Researchers who train sequence-to-sequence models that output punctuated, mixed-case text often penalize their own systems during evaluation if they calculate WER against the unpunctuated, uppercase LibriSpeech reference text without stripping predicted formatting.

Frequently asked questions

What is LibriSpeech used for?

LibriSpeech is primarily used as a standardized benchmark for training and evaluating automatic speech recognition (ASR) systems. It provides 1,000 hours of segmented, transcription-aligned audio that allows researchers to benchmark acoustic architectures, self-supervised representations, and language models on an openly redistributable corpus.

How many hours and speakers are in LibriSpeech?

LibriSpeech contains approximately 1,000 total hours of audio (960.9 training hours and approximately 21 hours of development and test data) recorded by 2,484 unique speakers, maintaining balanced representation between female and male contributors across all subsets.

What is the difference between train-clean and train-other?

The subsets are stratified based on acoustic recognition difficulty measured by a baseline Wall Street Journal acoustic model. The train-clean subsets (train-clean-100 and train-clean-360) feature lower ambient noise, higher recording fidelity, and predominantly General American accents. The train-other-500 subset contains recordings with higher background noise, varied room acoustics, or more diverse accents, serving as a test of acoustic robustness.

What audio format and sampling rate does LibriSpeech use?

LibriSpeech audio files are formatted as 16-bit mono FLAC files sampled at a uniform rate of 16 kHz, which represents the standard bandwidth requirement for wideband speech recognition models.

What license applies to the LibriSpeech dataset?

LibriSpeech is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. It is free for both academic research and commercial application, provided appropriate attribution is given to the original authors and OpenSLR.

How does LibriSpeech differ from Common Voice and Wall Street Journal?

Unlike the Wall Street Journal corpus, which is a proprietary dataset distributed under fee-gated licenses by the Linguistic Data Consortium, LibriSpeech is open access under CC BY 4.0. Unlike Mozilla Common Voice, which crowdsources short, varied sentences across global web users in many languages, LibriSpeech consists of long-form English audiobook chapters partitioned into calibrated clean and challenging subsets with pre-computed language modeling resources.

Sources

  • Panayotov, V., Chen, G., Povey, D., and Khudanpur, S. (2015). LibriSpeech: An ASR corpus based on public domain audio books. 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 5206–5210. https://danielpovey.com/files/2015_icassp_librispeech.pdf (accessed 2026-11-16)
  • Open Speech and Language Resources. (2015). Resource SLR12: LibriSpeech ASR corpus. http://www.openslr.org/12/ (accessed 2026-11-16)
  • Open Speech and Language Resources. (2015). Resource SLR11: LibriSpeech language models and LM training data. http://www.openslr.org/11/ (accessed 2026-11-16)
  • Kahn, J., Riviere, M., Zheng, W., Kharitonov, E., Xu, Q., Mazaré, P. E., Karadayi, J., Liptchinsky, V., Collobert, R., Fuegen, C., Lavechin, G., Synnaeve, G., and Dupoux, E. (2020). Libri-Light: A benchmark for ASR with limited or no supervision. IEEE ICASSP 2020. https://arxiv.org/abs/1912.07875 (accessed 2026-11-16)
  • Pratap, V., Xu, Q., Sriram, A., Synnaeve, G., and Collobert, R. (2020). MLS: A large-scale multilingual dataset for speech research. Interspeech 2020. https://arxiv.org/abs/2012.03411 (accessed 2026-11-16)
  • PyTorch Audio Team. (2026). Torchaudio LIBRISPEECH dataset documentation. https://pytorch.org/audio/stable/generated/torchaudio.datasets.LIBRISPEECH.html (accessed 2026-11-16)
  • Hugging Face. (2026). librispeech_asr dataset card and viewer. https://huggingface.co/datasets/librispeech_asr (accessed 2026-11-16)

Data: Canonical Speech Recognition Benchmark Datasets Comparison

Comparative specifications of canonical automatic speech recognition corpora, detailing audio duration, sampling rates, speaker counts, acoustic domains, licensing terms, and primary research utility, compiled from official repository documentations and peer-reviewed conference publications.

Period covered: 1993/2026. Coverage: Global.

Canonical Speech Recognition Benchmark Datasets Comparison
Corpus NameTotal Audio Hours (hours)Sampling Rate (kHz)Speaker CountAcoustic DomainPrimary LicenseKey Research Utility
LibriSpeech (SLR12)1000162484Read audiobooks (LibriVox)CC BY 4.0Supervised ASR benchmarking across clean and other acoustic conditions
TIMIT (LDC93S1)5.416630Phonetically balanced read promptsLDC User Agreement (Proprietary)Phoneme recognition and acoustic-phonetic modeling
Wall Street Journal (WSJ0/WSJ1)8216284Read newspaper articlesLDC User Agreement (Proprietary)Traditional hybrid HMM-GMM acoustic modeling baseline
Common Voice (Mozilla)3000048100000Crowdsourced read sentencesCC0 1.0 UniversalMultilingual and diverse accent robustness evaluation
Libri-Light600001610000Unannotated read audiobooksCC BY 4.0Self-supervised representation learning (wav2vec 2.0, HuBERT)
Multilingual LibriSpeech (MLS)500001615000Multilingual read audiobooks (8 European languages)CC BY 4.0Cross-lingual transfer learning and multilingual speech recognition
VoxCeleb1350161251Extracted YouTube interview audioCC BY 4.0Speaker identification and speaker verification

Follow CASRAI

Research-administration guidance, standards updates and independent tool reviews.

Ask CASRAI · free to try

Ask about LibriSpeech: Audio Corpus Guide and ASR Benchmarks

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →