CMU Kids: Children's Speech Dataset
- CMU Kids is a children’s speech dataset comprising 76 speakers (ages 6–11) with 5,180 utterances in American English used for age and gender classification.
- The corpus is analyzed with layer-wise extraction from Wav2Vec2 models paired with a CNN classifier, demonstrating that early layers capture speaker traits most effectively.
- Dimensionality reduction using PCA improves performance over MFCC baselines, providing compact, high-precision representations for speaker-specific information.
Searching arXiv for papers mentioning “CMU Kids” to ground the article in published sources. {"query": "\"CMU Kids\" dataset children speech", "max_results": 5, "sort_by": "submittedDate"} Searching arXiv for the Carnegie Mellon University formal-methods paper that uses CMU as a teaching site. {"query": "\"Engaging Millennials into Learning Formal Methods\" CMU", "max_results": 5, "sort_by": "relevance"} Searching arXiv for an original or foundational description of the CMU Kids children’s speech corpus. {"query": "\"CMU Kids\" speech corpus", "max_results": 10, "sort_by": "relevance"} “CMU Kids” denotes, in the supplied research literature, two distinct but Carnegie Mellon University–related referents. In speech processing, it is the name of a children’s speech corpus used for age and gender classification experiments, especially in a layer-wise analysis of self-supervised speech representations (Sinha et al., 14 Aug 2025). In formal-methods pedagogy, the phrase can be understood informally as referring to Carnegie Mellon University students in the Models of Software Systems course who participated in an Event-B survey during Fall 2015; the paper itself does not contain a literal section titled “CMU Kids,” but CMU functions as the initial empirical site for a later course redesign at Innopolis University (Cataño, 2018). The dominant technical usage in the provided sources is the speech corpus.
1. Term, scope, and disciplinary usage
Within the speech-processing context, CMU Kids is a children’s speech dataset used to test whether pretrained self-supervised learning representations encode speaker traits in children’s speech, specifically age and gender, without fine-tuning the SSL model (Sinha et al., 14 Aug 2025). The corpus is used alongside PFSTAR and is evaluated through layer-wise feature extraction from Wav2Vec2 variants followed by a CNN classifier.
Within the pedagogical context, “CMU Kids” is not a formal dataset or program name. The phrase is a plausible shorthand for MSS students at Carnegie Mellon University—mostly Software Engineering students in the MSIT-SE program—who had prior exposure to logic and software development and who answered an anonymous online survey about Event-B in Fall 2015 (Cataño, 2018). This suggests that the term is polysemous across the supplied sources: in one case it is a speech corpus; in the other it is an informal label for a CMU learner cohort.
A common source of confusion is to treat these two uses as equivalent. The supplied literature does not support that equivalence. One refers to child speech recordings; the other refers to graduate students in formal-methods instruction.
2. Corpus composition and supervised tasks
In the speech-study usage, CMU Kids contains 76 speakers, ages 6–11 years, with 5,180 utterances of American English (Sinha et al., 14 Aug 2025). The study uses a 70/30 train/test split, corresponding to 6.3 hours for training and 2.83 hours for testing.
| Item | Value |
|---|---|
| Speakers | 76 |
| Age range | 6–11 years |
| Total utterances | 5,180 |
| Language | American English |
| Split | 70/30 train/test |
| Training duration | 6.3 hours |
| Testing duration | 2.83 hours |
The utterance distribution reported for the split is 1,217 male and 2,349 female utterances in training, for a total of 3,566, and 529 male and 1,085 female utterances in testing, for a total of 1,614 (Sinha et al., 14 Aug 2025). The corpus supports two classification tasks. Age classification is formulated as a 6-class problem because the ages span 6 to 11 years. Gender classification is binary male/female classification.
The paper uses this setup to examine transfer from adult-pretrained SSL models to children’s speech. A plausible implication is that CMU Kids functions as a benchmark for child-specific speaker-trait probing rather than for ASR alone.
3. Experimental pipeline for layer-wise SSL analysis
The layer-wise study evaluates four Wav2Vec2 variants: base-100h, base-960h, large-960h-lv60, and large-960h-lv60-self (Sinha et al., 14 Aug 2025). The reported model characteristics are explicit. Base models have 13 hidden layers total, indexed 0–12, produce 768-dimensional features, and have about 95M parameters. Large models have 25 hidden layers total, indexed 0–24, produce 1024-dimensional features, and have about 317M parameters.
Representations are extracted without fine-tuning on children’s speech. This is central to the experimental design, because the objective is to measure direct transfer of pretrained SSL features to age and gender classification in children.
The downstream classifier is a simple CNN with 3 convolutional layers, filter sizes 64, 128, and 256, kernel size , ReLU activations, and batch normalization (Sinha et al., 14 Aug 2025). The same classifier is used across all layers and models to isolate the effect of the SSL representations themselves.
The reported methodological emphasis is layer-wise probing rather than end-to-end adaptation. This suggests that CMU Kids is being used as an analysis substrate for representational stratification across network depth.
4. Layer-wise behavior and predictive performance
The principal empirical finding is that early layers are the most informative for children’s age and gender classification, while deeper layers become less useful for these speaker traits (Sinha et al., 14 Aug 2025). The paper characterizes early layers, roughly 0–7 and especially 1–2, as better for speaker-specific cues, whereas deeper layers increasingly encode linguistic information.
For age classification on CMU Kids, the best-performing layers are: MFCC baseline with 26 features at 89.97%, base-100h at layer 1 with 92.13%, base-960h at layer 0 with 91.63%, large-960h-lv60 at layer 1 with 96.84%, and large-960h-lv60-self at layer 1 with 92.37% (Sinha et al., 14 Aug 2025). For gender classification, the best-performing layers are: MFCC baseline with 26 features at 88.41%, base-100h at layer 2 with 93.78%, base-960h at layer 1 with 94.96%, large-960h-lv60 at layer 2 with 96.68%, and large-960h-lv60-self at layer 2 with 96.53%.
| Task | Best raw result |
|---|---|
| Age classification | 96.84% with large-960h-lv60, layer 1 |
| Gender classification | 96.68% with large-960h-lv60, layer 2 |
For the best raw CMU Kids results, the paper reports full classification metrics. Age classification with large-960h-lv60 yields accuracy 96.84%, precision 0.97, recall 0.97, and F1 0.97. Gender classification with the same model yields accuracy 96.68%, precision 0.97, recall 0.97, and F1 0.97 (Sinha et al., 14 Aug 2025).
The paper interprets the early-layer advantage through preservation of low-level acoustic information relevant to pitch, formant structure, timbre, vocal tract characteristics, and pronunciation/prosodic patterns. As depth increases, the representation becomes more abstract and more oriented toward linguistic or phonetic content. In the paper’s terms, this helps tasks like ASR but can hurt speaker-trait classification.
5. PCA, compactness, and statistical interpretation
The study explicitly applies Principal Component Analysis to the best-performing layer of each model and reduces dimensions from 512 down to 32 in steps of 64 (Sinha et al., 14 Aug 2025). PCA is used to reduce feature dimensionality while retaining the most informative variance.
For CMU Kids, PCA improves or preserves performance across models. The reported results are 93.18% age and 96.22% gender for base-100h at 192 dimensions, 93.43% age and 96.71% gender for base-960h at 192 dimensions, 97.14% age and 98.20% gender for large-960h-lv60 at 256 dimensions, and 96.84% age and 97.95% gender for large-960h-lv60-self at 128 dimensions (Sinha et al., 14 Aug 2025).
| Model | Best PCA result |
|---|---|
| large-960h-lv60 | 97.14% age, 98.20% gender |
| large-960h-lv60-self | 96.84% age, 97.95% gender |
The paper’s interpretation is that speaker-related information is concentrated in a small subset of principal components. PCA helps by removing redundant dimensions, reducing noise, improving generalization, and lowering computation and storage cost. The paper further states that, for CMU Kids especially, gender cues are even more compactly represented than age cues, since gender reaches 98.20% with only 64 dimensions for the best large model (Sinha et al., 14 Aug 2025).
Relative to the MFCC baseline, the gains are substantial. The baseline yields 89.97% for age and 88.41% for gender, whereas the best raw SSL results reach 96.84% and 96.68%, and the best PCA results reach 97.14% and 98.20%. The paper also reports Wilcoxon signed-rank tests showing statistically significant improvements over MFCC for age classification across all models, and for gender only in the larger models.
These results support the paper’s broader conclusion that self-supervised representations transfer well to children’s speech and that acoustic and linguistic information are structured differently across SSL model depth.
6. Secondary usage in formal-methods pedagogy at Carnegie Mellon
A distinct use of the phrase arises in “Engaging Millennials into Learning Formal Methods,” where Carnegie Mellon University is one of the universities used as a reference point for teaching formal methods and is the site of the first survey reported in the paper (Cataño, 2018). The first author visited CMU in Pittsburgh in Fall 2015, trained in the Models of Software Systems master course, gave guest lectures on predicate and first-order logic, UML and OCL, and the introduction of Event-B, and later used that experience to reshape teaching at Innopolis University.
The CMU learners in that survey were MSS students at CMU, mostly Software Engineering students in the MSIT-SE program, with prior exposure to logic and software development, typically covered by courses such as Discrete Mathematics and Software Engineering (Cataño, 2018). The paper characterizes them as a millennial graduate-student population that was more comfortable with practical software concerns than with abstract mathematical formalism.
The anonymous online survey was designed to evaluate whether Event-B was a good addition to the course. Twenty-nine students answered. For the overall rating of Event-B sessions, 4 rated them Excellent, 18 Good, 6 Neutral, 1 Poor, and 0 Terrible; the paper summarizes this as about 76% Good or Excellent, 21% Neutral, and 3% Poor or Terrible (Cataño, 2018). Responses to questions about favorite or attractive aspects emphasized real examples, the close link between Event-B and code generation and programming languages, ease of syntax, tool support, the approach to software development, and ease of use.
The educational setting at CMU is described as 16 weekly classes and 16 weekly recitation sessions, each session lasting 2 hours 45 minutes, with weekly homework and a project with multiple deliverables (Cataño, 2018). Topics included propositional and predicate logic, proof techniques, sets, relations, functions, sequences and induction, state machines, Z, concurrency, and linear temporal logic, with the last sessions devoted to “FM in Practice,” including FM tools and industrial examples. The CMU project focused on modeling and verification of an Infusion Pump and had 3 deliverables on FSP, Z, and FSP with temporal logic.
The paper does not claim that Event-B is better or more expressive than Z. Rather, it states that Event-B is more suitable for millennials because it has a more direct practical link to code refinement and code generation, whereas Z is more oriented toward formal specification (Cataño, 2018). That CMU experience later informed concrete changes at Innopolis University, including expanding Event-B sessions from 2 to 4, replacing the final “FM in Practice” sessions, moving from an infusion pump project to Android app development, and making code generation and interface implementation part of the final project.
In this pedagogical sense, “CMU Kids” refers not to children but to CMU master’s students whose feedback helped motivate a more tool-supported, example-driven, and implementation-oriented approach to formal-methods teaching.
7. Significance and boundaries of the term
Taken together, the supplied literature gives “CMU Kids” a sharply domain-dependent meaning. In speech technology, it is a children’s American English corpus with 76 speakers and 5,180 utterances, used to show that early Wav2Vec2 layers encode child age and gender cues effectively and that PCA can improve compactness and accuracy (Sinha et al., 14 Aug 2025). In formal-methods education, it is an informal reading of a CMU student cohort whose survey responses favored realistic examples, code generation, and tool support, thereby influencing later curriculum design (Cataño, 2018).
The first usage is corpus-centric and quantitative; the second is cohort-centric and pedagogical. A plausible implication is that the expression “CMU Kids” should be interpreted only in relation to its disciplinary context. In current arXiv usage within the supplied sources, the technically dominant meaning is the children’s speech dataset.
