MyST Corpus: Children's ASR Benchmark

• MyST Corpus is a structured children’s conversational speech dataset collected during virtual science tutoring, offering real-world spontaneous child speech.
• It includes 393 hours of audio across 10,496 sessions from 1,371 students, with approximately 45% word-level transcriptions aiding ASR performance benchmarking.
• The corpus supports both supervised and self-supervised research, enabling ASR adaptation studies and multimodal AI tutoring advancements in elementary science.

Searching arXiv for papers on the MyST corpus and related child ASR work. MyST Corpus, or My Science Tutor Corpus, is a children’s conversational speech corpus collected during virtual science tutoring sessions and intended to support automatic speech recognition, conversational AI tutoring, and multimodal educational systems. It comprises 393 hours of student speech, 228,874 student utterances, 10,496 virtual tutor sessions, and 1,371 students in grades 3–5, with roughly 45% of utterances transcribed at the word level. The corpus is notable for combining scale with task-oriented spontaneous speech in an educational dialog setting rather than read-speech elicitation, and it has become a recurrent benchmark for child ASR and adaptation studies (Pradhan et al., 2023).

1. Origin and dialog setting

MyST was developed as part of the My Science Tutor project to support research and development in ASR for children and to enable conversational AI tutors and multimodal learning tools in elementary science education. Speech was collected during one-on-one virtual tutoring sessions with a lifelike character, “Marni,” after hands-on classroom science investigations based on FOSS modules. The interaction design is strict turn-taking: the tutor presents information, asks a question, and waits for the student response. Students speak using push-to-talk by pressing and holding the spacebar, and each student turn is recorded as a separate audio file (Pradhan et al., 2023).

The corpus is organized around science instruction rather than generic conversation. In Phase I, the modules were Mixtures and Solutions, Magnetism and Electricity, Variables, and Water. In Phase II, the modules were Energy and Electromagnetism, Living Systems, Mixtures, Soil/Rocks/Landforms, and Sun/Moon/Planets. Prompting targeted observations, causal explanations, procedural reflection, and scaffolded follow-ups tied to curriculum progression. Student responses range from short utterances to multi-sentence explanations and include domain vocabulary, disfluencies, false starts, and pronunciation variability; explicit mispronunciations were intentionally retained in transcripts where possible (Pradhan et al., 2023).

A common misconception is to treat MyST as a read-speech resource analogous to many older children’s corpora. The corpus paper instead describes it as conversational, explanation-oriented speech elicited in structured tutoring dialogs. This distinction matters because downstream ASR systems must handle spontaneous production phenomena, including fillers, topic restarts, and nonstandard child grammar, rather than only controlled reading performance (Pradhan et al., 2023).

2. Scale, phases, and corpus structure

The released corpus contains 393 hours of student speech from 1,371 students across 10,496 sessions and 228,874 student utterances. Sessions were designed to last 15–20 minutes, but analyses showed roughly 5 minutes of student speech per dialog because only net student speech during push-to-talk is recorded. The release includes predefined training, development, and test partitions created by stratified sampling to represent phases and modules, with each student appearing in exactly one partition (Pradhan et al., 2023).

Component	Value
Students	1,371
Sessions	10,496
Student utterances	228,874
Audio	393 h
Train / Dev / Test	315 h / 40 h / 39 h

The two-phase collection structure is explicit. Phase I involved grades 3–5, 421 students, 1,509 sessions, and 102 hours, with all sessions transcribed under “rich” guidelines. Phase II involved grades 4–5, 950 students, 8,987 sessions, and 291 hours; 1,426 sessions amounting to 95 hours were transcribed, while 3,711 sessions amounting to 196 hours remained untranscribed. The totals across phases are 393 hours, 10,496 sessions, and 1,371 students (Pradhan et al., 2023).

The corpus layout is utterance-centric. Audio is distributed as .flac files and transcripts as .trn files. Identifiers encode school, student, date, time, module, investigation, and utterance index. The release uses train, development, and test directories, and the student-disjoint partitioning is intended to support reproducible evaluation. Later ASR work preserves these official speaker-independent splits when constructing filtered subsets for model adaptation and testing (Attia et al., 2023).

3. Transcription, cleanup, and annotation properties

Approximately 45% of MyST utterances, or about 100K utterances, have been transcribed at the word level. Phase I used “rich” guidelines, whereas Phase II used “reduced” guidelines to scale annotation. The corpus paper does not report word-level timestamps, diarization, or auxiliary annotation layers such as correctness, intent, or emotion labels. It does report that the authors corrected reported test-set transcription errors and planned another quality-control pass for development and training data (Pradhan et al., 2023).

Corpus preparation included several cleanup steps before release. The authors removed bad, empty, and corrupted sessions; inspected sessions shorter than 10 minutes or longer than 1 hour; deleted sessions missing significant audio; detected and removed or tagged files with significant clipping; trimmed leading and trailing silence while retaining a small fraction; and removed utterances with significant background noise or crosstalk when detected, especially in transcribed or manually verified samples. The release also includes an updated CMU-based pronunciation lexicon containing novel words from the corpus (Pradhan et al., 2023).

Later work using MyST for Whisper adaptation reports that transcript quality in the release it used was highly variable. That study documents outright mismatches between audio and transcript, evidence of lower-quality automatic transcriptions in parts of the data, special labels such as <DISCARD>, <NO_SIGNAL>, and <SILENCE>, and transcripts that were entirely uppercase in the version used by the authors. It also reports that no timestamps were provided with transcriptions in that version, that files were short on average at about 8 seconds, and that some files exceeded 30 seconds, complicating training for models with fixed context windows (Attia et al., 2023).

These two descriptions are not equivalent in scope: the corpus paper emphasizes release design and corpus-wide curation, while the later ASR study emphasizes model-oriented filtering and the practical consequences of transcript and segmentation noise. A plausible implication is that MyST should be understood not only as a dataset but also as a corpus whose effective research value depends strongly on the preprocessing regime applied to a given release.

4. Access, licensing, and ethical conditions

MyST is available for non-commercial use under a Creative Commons license at https://myst.cemantix.org, and it is also available for commercial use through Boulder Learning at https://boulderlearning.com/resources/myst-corpus/. The corpus is cataloged by LDC as “MyST Children’s Conversational Speech” (LDC2021S05). The corpus paper reports that ten organizations had licensed the corpus for commercial use and that approximately 40 university and other not-for-profit research groups had downloaded it (Pradhan et al., 2023).

The release is anonymized. No identifying information is stored beyond anonymized codes for schools and students, and all students and parents signed consent or assent for distribution of anonymous speech and transcriptions. The data collection was IRB approved. Later methodological work reiterates that use should honor academic-use terms where applicable, maintain privacy, and avoid re-identification; it also identifies demographic bias analysis as an open direction for future work (Pradhan et al., 2023, Attia et al., 2023).

A second common misconception is to assume that MyST is fully transcribed because it is widely used for supervised ASR. The corpus paper states that roughly 45% of utterances had been transcribed at the time of publication, and later ASR studies explicitly exploit the untranscribed portion for semi-supervised or self-supervised adaptation rather than treating MyST as a fully labeled resource (Pradhan et al., 2023, Wang et al., 9 Jun 2026).

5. Benchmarking and evaluation role

The original corpus paper positioned MyST as a benchmark resource with predefined train, development, and test partitions and a baseline ASR result. Using a SpeechBrain end-to-end transformer fine-tuned from a LibriSpeech model and restricting training to utterances shorter than 30 seconds due to memory limits, the authors reported 11.6% WER on the uncorrected MyST test set and 10.0% WER on the corrected test set. The paper gives the standard definition

$\mathrm{WER}=\frac{S+D+I}{N},$

where $S$ is substitutions, $D$ deletions, $I$ insertions, and $N$ the number of words in the reference (Pradhan et al., 2023).

The release structure was designed to support not only supervised evaluation but also semi-supervised augmentation. Untranscribed data is included in all partitions to support pseudo-labeling and “pseudo-unseen” conditions as more transcriptions become available. This design feature became consequential in later work, where MyST was used simultaneously as a source of labeled child speech and as a source of unlabeled in-domain speech for adaptation (Pradhan et al., 2023, Wang et al., 9 Jun 2026).

MyST is also frequently used after additional filtering rather than directly in its raw partition form. In Kid-Whisper, the authors start from the official speaker-independent splits, note about 197 hours of transcribed speech with uneven transcription quality, and derive a curated subset of 179.2 hours of well-transcribed speech: 132.5 hours for training, 20.9 for development, and 25.8 for test. Their reported effect is to remove 17.8 hours while substantially improving transcript reliability for Whisper fine-tuning (Attia et al., 2023).

6. Whisper adaptation and filtered MyST protocols

Kid-Whisper made MyST a central benchmark for Whisper-based child ASR. Its preprocessing pipeline transcribes all MyST audio with Whisper-Large-v1, computes WER against provided transcripts, and flags files with WER greater than 50%; removes files containing one or two words; excludes files labeled <DISCARD>, <NO_SIGNAL>, or <SILENCE>; removes files longer than 30 seconds from train and dev because timestamps are unavailable for safe truncation; concatenates short files from the same recording session up to but not exceeding 30 seconds; lowercases transcripts; and normalizes them with WhisperNormalizer, including contraction expansion and digit normalization. The official speaker-independent train, dev, and test splits are retained throughout (Attia et al., 2023).

The study quantifies the effect of this filtering with a dataset-level probe using Whisper-Large-v1. On the train split, WER changes from 29.5% over 145 hours with no filtering, to 26.8% over 140 hours after removing files with WER above 50%, and to 19.2% over 132.5 hours after removing files with WER above 50% or fewer than 3 words. On the dev split, the corresponding values are 26.2% over 25.5 hours, 22.3% over 25.5 hours, and 12.8% over 21 hours. On the test split, they are 26.2% over 28.1 hours, 22.3% over 26.7 hours, and 14.2% over 25.6 hours (Attia et al., 2023).

Fine-tuning Whisper on this curated MyST subset reduced MyST test WER from 13.93% to 9.11% for Whisper-Small (English-only) and from 13.23% to 8.61% for the best fine-tuned medium result, obtained by the multilingual medium variant trained on MyST. The same study also reports a clear zero-shot child-versus-adult gap, with MyST WERs around 12.9–14.1% versus 2.76–3.39% on LibriSpeech test-clean across Whisper medium and small configurations, and notes only small degradations on adult speech after child-speech fine-tuning (Attia et al., 2023).

Later work used the Kid-Whisper filtering protocol as an operational definition of a high-quality MyST subset. “Mind the Shift” applies the same filtering rules—removing samples with WER above 50%, fewer than 3 words, or duration above 30 seconds—and uses 133 hours for training, 21 for development, and 25 for test, with additional 1-hour, 5-hour, and 10-hour training subsets. In that setting, the best reported SSL result is a WER of 9.64 using WavLM fused with delta W2V2 embeddings by concatenation, which the authors describe as a new state of the art among SSL models on MyST (Wang et al., 28 Jan 2026).

7. Self-supervised adaptation, methodological uses, and open issues

MyST has also become a testbed for low-resource and self-supervised adaptation. Gumbel-BEARD describes MyST as a corpus of 448 hours of conversational child speech with 240 hours transcribed, and, after filtering following Attia et al. (2023), retains 133/21/25 hours for labeled train/dev/test while using the remaining 208 hours of untranscribed audio for stage-one self-supervised adaptation. It defines random 1-hour and 10-hour labeled subsets from the training partition for limited-supervision experiments (Wang et al., 9 Jun 2026).

In Gumbel-BEARD, the stage-one objective is

$$\mathcal{L} = \mathcal{L}_q^L + \lambda \mathcal{L}_d^L + \beta \lambda \mathcal{L}_d^N,$$

where $\mathcal{L}_q^L$ is the BEST-RQ quantization loss, and $\mathcal{L}_d^L$ and $\mathcal{L}_d^N$ are inner and output distillation losses. The method uses a hard Gumbel-Softmax selector over all Whisper encoder layers, with temperature $\tau$ annealed linearly from $S$0 to $S$1. On MyST, it reports for Whisper-small: 13.40% zero-shot, 10.64%/9.94%/9.34% for supervised fine-tuning with 1 hour, 10 hours, and the full 133 hours, and 10.18%/9.35%/8.51% for Gumbel-BEARD under the same label budgets. For Whisper-medium, it reports 13.10% zero-shot, 9.56%/9.19%/8.86% for supervised fine-tuning, and 9.15%/8.88%/8.21% for Gumbel-BEARD, with the full-set 8.21% identified as a new state of the art on MyST for Whisper-medium (Wang et al., 9 Jun 2026).

These results show that MyST is used in at least three distinct ways in recent work: as a supervised benchmark with standardized partitions, as a corpus requiring aggressive transcription-aware filtering for Whisper adaptation, and as an in-domain unlabeled source for self-supervised adaptation. The corpus has also supported analyses of child-specific error sources. Kid-Whisper highlights disfluencies and filler words, unusual grammatical constructions, topic variability, prosodic variability, distortion from children speaking too close to microphones, and diverse recording conditions as persistent challenges. Gumbel-BEARD adds a more general account of child-speech domain mismatch, citing shorter vocal tracts, higher $S$2, increased acoustic variability, and disfluencies (Attia et al., 2023, Wang et al., 9 Jun 2026).

Several limitations recur across studies. The corpus paper notes that transcription coverage is incomplete and that accent or dialect distribution, socioeconomic diversity, and other demographics are not specified. Kid-Whisper reports that individual age and gender were unavailable in the release it used and that no timestamps were provided with transcriptions. Gumbel-BEARD explicitly states that MyST speaker counts, exact age ranges in years, channel or microphone types, sampling rate, text normalization policies, decoding setup, and several preprocessing details are unspecified in that paper’s MyST experiments (Pradhan et al., 2023, Attia et al., 2023, Wang et al., 9 Jun 2026).

Taken together, these properties position MyST as a large, structured, educationally grounded child-speech corpus whose scientific value lies not only in its size but also in its combination of spontaneous explanatory speech, student-disjoint partitions, partial transcription, and compatibility with both supervised and self-supervised research workflows. This suggests that MyST is especially useful for studying the interaction between corpus curation, domain mismatch, and model adaptation in child ASR.