PFSTAR: Child Speech Corpus Benchmark

• PFSTAR is a British children’s speech corpus covering ages 4–14, widely used to benchmark ASR, keyword spotting, and paralinguistic classification.
• The corpus comprises both read and spontaneous child speech with detailed transcriptions and varied data splits for training, evaluation, and testing.
• Studies using PFSTAR highlight significant adult-child acoustic mismatches and demonstrate that self-supervised learning features outperform traditional MFCC approaches.

Searching arXiv for PFSTAR-related papers to ground the article in published work. PFSTAR is a British English children’s speech corpus that has become a recurrent benchmark in child-speech research, especially for automatic speech recognition, keyword spotting, and paralinguistic classification. Across the studies summarized here, it is described as covering children aged 4–14 years, containing transcriptions, and including both read and spontaneous child speech; later work additionally treats it as a benchmark with gender labels, 11 age groups, and substantial class imbalance. Its research significance derives from the fact that children’s speech differs substantially from adult speech in pitch, spectral shape, pronunciation, articulation, speaking rate, vocal-tract configuration, and developmental variability, so PFSTAR is repeatedly used to test whether methods trained or pretrained on adult speech remain robust under child-specific acoustic mismatch (Jain et al., 2022, Sinha et al., 20 Jun 2026).

1. Corpus profile and reported configurations

In one widely cited description, PFSTAR is the “PF-STAR Corpus of British English child speech.” It contains speech from 158 children aged 4 to 14 years, includes both read and spontaneous child speech, has transcriptions, and was collected at a university laboratory and two primary schools. That paper reports an original partition of 7.5 hours for training, 1 hour for evaluation, and 5.6 hours for test, and notes that PFSTAR is widely used in child speech ASR research (Jain et al., 2022).

Subsequent studies operationalize PFSTAR through task-specific subsets rather than a single immutable protocol. For age and gender classification, PFSTAR is reported with 8.3 hours from 122 speakers for training and 1.1 hours from 60 speakers for testing, corresponding to 856 training utterances and 129 test utterances, with 11 age classes and binary gender labels. A later multi-architecture study repeats the same 8.3-hour/1.1-hour split and adds that the corpus has an average utterance duration of 41.32 s and a naturally imbalanced age distribution, especially at younger and older extremes (Sinha et al., 14 Aug 2025, Sinha et al., 20 Jun 2026).

For zero-shot transfer experiments, PFSTAR is typically used only as a child-speech test set. In both zero-shot ASR and zero-shot KWS, the reported evaluation subset is 1.1 hours from 60 speakers. The ASR paper describes the tested subset as ages 4–13, while the KWS paper describes the test condition as ages 4–14 and notes that 28 of the 60 speakers are female (Sinha et al., 28 Aug 2025, Kutum et al., 28 Aug 2025).

Study	PFSTAR role	Reported configuration
(Jain et al., 2022)	Child ASR corpus	158 children, ages 4–14; 7.5 h train, 1 h eval, 5.6 h test
(Sinha et al., 14 Aug 2025)	Age/gender classification	8.3 h train from 122 speakers; 1.1 h test from 60 speakers
(Sinha et al., 20 Jun 2026)	Multi-architecture layer analysis	Same split; 11 age groups; average utterance duration 41.32 s
(Sinha et al., 28 Aug 2025)	Zero-shot ASR test set	1.1 h test subset; 60 speakers; ages 4–13
(Kutum et al., 28 Aug 2025)	Zero-shot KWS test set	1.1 h test subset; 60 speakers; ages 4–14

This variation suggests that PFSTAR functions both as a corpus and as a family of benchmark protocols. The underlying resource is stable, but the effective evaluation unit depends on the task: full-corpus fine-tuning, train/test classification splits, or zero-shot adult-to-child transfer.

2. Why PFSTAR is a difficult benchmark

The central difficulty associated with PFSTAR is adult–child mismatch. The zero-shot ASR study explicitly motivates the corpus by noting that children’s speech differs substantially from adult speech in pitch and spectral shape, pronunciation and articulation, speaking rate, vocal-tract configuration and developmental variability, and strong age-dependent differences, especially for younger children. The KWS study uses nearly the same rationale, emphasizing higher pitch, less stable pronunciation, and strong age-related variation (Sinha et al., 28 Aug 2025, Kutum et al., 28 Aug 2025).

Because of these properties, PFSTAR is repeatedly used in deliberately mismatched training/testing protocols. In zero-shot ASR, the training data are WSJCAM0 adult British English speech, comprising 15.5 hours from 92 speakers, and the test data are PFSTAR child speech, with no child-specific fine-tuning. The system is implemented in Kaldi with a DNN acoustic model. The MFCC baseline uses 40-dimensional MFCCs with a 20 ms frame and 10 ms shift, applies fMLLR, and trains a DNN with 5 hidden layers, 1024 nodes per layer, 30 epochs, and a learning rate of 0.005 followed by 0.0005. Decoding uses a bigram LLM trained on PFSTAR transcripts excluding test utterances; by contrast, direct SSL decoding uses no external LLM (Sinha et al., 28 Aug 2025).

The zero-shot KWS study adopts an analogous WSJCAM0-to-PFSTAR mismatch. WSJCAM0 contributes 15.5 hours from 92 of 140 adult speakers, while PFSTAR provides 1.1 hours from 60 children for evaluation. The goal is not merely recognition accuracy but adult-trained keyword-detection robustness on child speech without child-speech training or adaptation (Kutum et al., 28 Aug 2025).

In both studies, frozen self-supervised learning models are used as feature extractors. The ASR paper evaluates Wav2Vec2-Large-960h-lv60-self, HuBERT-Large-LS960-ft, Data2Vec-Audio-Large-960h, and WavLM-Large; the KWS paper evaluates Wav2Vec2, HuBERT, and Data2Vec. For these large SSL models, each architecture is described as producing 25 hidden-layer feature sets, with layer 0 corresponding to the CNN feature encoder and layers 1–24 to transformer layers, each yielding 1024-dimensional representations (Sinha et al., 28 Aug 2025, Kutum et al., 28 Aug 2025).

3. PFSTAR in zero-shot automatic speech recognition

PFSTAR serves as the main evidence for the claim that layer-wise SSL features can improve zero-shot ASR for children’s speech without child-speech fine-tuning. The ASR study compares direct zero-shot decoding against a conventional Kaldi back end trained on adult WSJCAM0 speech but fed with frozen, layer-specific SSL features instead of MFCCs. The experiment is explicitly a layer-selection and feature-substitution study rather than a feature-combination method (Sinha et al., 28 Aug 2025).

The direct zero-shot decoding results on PFSTAR are reported as 10.65% WER for Wav2Vec2, 10.67% for HuBERT, and 9.82% for Data2Vec. WavLM is described as not competitive, and the paper suggests that it may overfit due to extra pretraining objectives. The layer-wise analysis then shows a marked depth effect: early layers 0–5 are worse and retain more low-level acoustic information, middle layers 6–15 improve, and later layers 16–24 are best overall, capturing more abstract and phonemic information. The best PFSTAR results are 5.15% WER for Wav2Vec2 at layer 22, 5.69% for HuBERT at layer 24, and 5.43% for Data2Vec at layer 22. For Wav2Vec2, the reduction from 10.65% to 5.15% corresponds to a reported 51.64% relative improvement (Sinha et al., 28 Aug 2025).

The paper also stresses that deeper is not universally better. For Wav2Vec2, layer 23 is reported to spike to 13.62% WER, so the effective region is late but not necessarily final. The interpretation offered is that the best late transformer layers appear to separate phonemic content from age-specific variation (Sinha et al., 28 Aug 2025).

Age-group analysis on PFSTAR further illustrates the developmental difficulty of the corpus. Using Wav2Vec2 layer 22, the reported WERs are 13.51% for ages 4–6, 3.75% for ages 7–9, and 4.09% for ages 10–13, compared with baselines of 27.35%, 8.39%, and 7.09%, respectively. Relative improvements are 50.61%, 55.32%, and 42.30%. Older children therefore have lower absolute WERs, younger children remain hardest to recognize, and the largest absolute gain occurs for the youngest group, which the paper characterizes as the most acoustically variable (Sinha et al., 28 Aug 2025).

Within that study, PFSTAR is also the basis for comparison with earlier zero-shot child-ASR approaches. The reported prior systems achieve 9.5% with pitch robust BS-MFCC + TDNN, 8.69% with formant modification + TDNN, and 7.1% with jitter/strength-of-excitation + MFCC + TDNN, whereas the proposed layer-wise SSL + DNN system reaches 5.15% (Sinha et al., 28 Aug 2025).

4. PFSTAR in zero-shot keyword spotting

PFSTAR is likewise the central test set in zero-shot KWS for children’s speech, again under WSJCAM0-to-PFSTAR mismatch. Here the key evaluation measures are the NIST spoken-term-detection metrics ATWV, MTWV, probability of false alarm, and probability of miss. The paper states that TWV values closer to 1 indicate better KWS performance, while lower false alarm and miss rates are preferred (Kutum et al., 28 Aug 2025).

The study reports best-layer PFSTAR results separately for 10-, 20-, and 30-keyword sets. For 30 keywords, the strongest result comes from Wav2Vec2 layer 22, with ATWV $= 0.691$, MTWV $= 0.700$, $P_{\text{fa}} = 0.016$, and $P_{\text{miss}} = 0.054$. For the same 30-keyword condition, the MFCC baseline has ATWV $-1.907$, MTWV $-1.764$, $P_{\text{fa}} = 0.176$, and $P_{\text{miss}} = 0.256$; HuBERT layer 21 yields ATWV $0.457$, and Data2Vec layer 22 yields ATWV $0.471$ (Kutum et al., 28 Aug 2025).

The same broad depth trend observed in zero-shot ASR reappears in KWS. Early layers are described as weak, middle layers capture more phonetic structure, and later layers—especially around layers 13–22—provide the strongest keyword-detection performance. The headline system is again Wav2Vec2 layer 22 (Kutum et al., 28 Aug 2025).

PFSTAR also enables age-specific KWS analysis. Using Wav2Vec2 layer 22 on the 30-keyword set, ages 4–6 obtain ATWV $= 0.700$0, ages 7–9 obtain $= 0.700$1, and ages 10–13 obtain $= 0.700$2, with all three groups outperforming the MFCC baseline by large margins. The paper’s main takeaway is that performance improves with age and that the youngest children are hardest, reinforcing the view that PFSTAR exposes developmental acoustic variability that makes child KWS particularly difficult (Kutum et al., 28 Aug 2025).

A further role of PFSTAR in this study is noise-robustness testing. The authors corrupt the PFSTAR test set with babble, factory, Volvo, white, ambulance siren, crowd, thunderstorm, and birds chirping noise at 5 dB, 10 dB, and 15 dB SNR, and compare the MFCC baseline with Wav2Vec2 layer 22. The reported examples are uniformly favorable to SSL features: under babble noise with 30 keywords at 10 dB, MFCC yields $= 0.700$3 while layer 22 yields $= 0.700$4; under factory noise with 30 keywords at 15 dB, the values are $= 0.700$5 and $= 0.700$6; under Volvo at 15 dB, $= 0.700$7 and $= 0.700$8; and under thunderstorm at 15 dB, $= 0.700$9 and $P_{\text{fa}} = 0.016$0. The paper concludes that SSL embeddings are substantially more noise-robust than MFCCs (Kutum et al., 28 Aug 2025).

5. PFSTAR in age and gender representation analysis

PFSTAR is also used to probe how SSL models encode child-specific paralinguistic cues. In the Wav2Vec2-only study, PFSTAR is one of two corpora used to assess age and gender classification across layer depth. The best PFSTAR age layers are early: layer 6 for base-100h, layer 5 for base-960h, and layer 7 for both large models. Gender layers are earlier still: layer 4 for base-100h, layer 2 for base-960h, layer 1 for large-960h-lv60, and layer 2 for large-960h-lv60-self. The reported best PFSTAR accuracies are 84.25% for age from base-100h layer 6 and 94.57% for gender from both base-960h layer 2 and large-960h-lv60-self layer 2, against MFCC baselines of 80.92% for age and 87.63% for gender. Statistical significance is assessed with Wilcoxon signed-rank tests, and age-classification improvements are reported as significant for all models with $P_{\text{fa}} = 0.016$1 (Sinha et al., 14 Aug 2025).

That paper further applies PCA only after selecting the best-performing layer. The dimensionality is reduced from 512 to 32 in steps of 64, and PFSTAR results improve in selected cases: base-100h reaches 86.05% age accuracy at 320 dimensions, while large-960h-lv60-self reaches 95.00% gender accuracy at 384 dimensions. The interpretation is that early Wav2Vec2 layers preserve pitch, spectral shape, and other speaker-dependent properties, whereas deeper layers become increasingly linguistically oriented (Sinha et al., 14 Aug 2025).

A broader multi-architecture study extends this analysis from Wav2Vec2 to HuBERT, Data2Vec, and WavLM. On PFSTAR, the best exact-age result is 87.40% from HuBERT large layer 3, while the best gender accuracy is 94.57%, achieved by Wav2Vec2 large-960h-lv60-self layer 2, HuBERT layer 6, and Wav2Vec2 base-960h layer 2. The paper concludes that PFSTAR’s strongest age and gender cues are concentrated in early to mid-level layers rather than the deepest ones, and explicitly contrasts this with the deeper-layer advantage seen in phonemic or recognition-oriented tasks (Sinha et al., 20 Jun 2026).

The same study uses PFSTAR to analyze compactness, aggregation, and robustness. PCA improves PFSTAR age classification for HuBERT from 87.40% to 89.15% at 320 dimensions and yields 95.00% gender accuracy for Wav2Vec2 large-960h-lv60-self at 384 dimensions; HuBERT large gender reaches 94.57% with only 64 PCA dimensions. When exact ages are merged into broader groups—4–6, 7–9, and 10–13/14—the best PCA-reduced HuBERT representation achieves 93.80% overall age-group accuracy. Under 7-fold speaker-wise cross-validation, PFSTAR becomes much harder: MFCC baselines are $P_{\text{fa}} = 0.016$2 for age and $P_{\text{fa}} = 0.016$3 for gender, while HuBERT embeddings yield $P_{\text{fa}} = 0.016$4 and $P_{\text{fa}} = 0.016$5. Layer aggregation helps age more than gender, with PFSTAR age reaching 92.91% by concatenating the best two HuBERT layers. Short-segment experiments after VAD also show that PFSTAR supports usable classification from 1–3 s chunks, though with reduced performance, especially for age (Sinha et al., 20 Jun 2026).

A common misconception would be to treat “best SSL layer” as task-invariant. PFSTAR results argue against that view: late layers are optimal for zero-shot ASR and KWS, but early to mid layers are optimal for age and gender classification. The corpus is therefore informative not only about model quality, but also about the task-dependent organization of information across SSL depth.

6. PFSTAR in supervised child ASR and benchmark significance

PFSTAR is not only a zero-shot benchmark; it is also used for child-specific fine-tuning. In a wav2vec2-based child ASR study, the original corpus was cleaned and segmented into 12 hours of usable data with utterance lengths between 5 and 20 seconds, using .trs files, time-aligned timestamps, the sp tag, FFmpeg, and Python. The resulting subsets were PFS_10m, PFS_1h, PFS_10h for fine-tuning and PFS_test for inference, with the final split being 10 hours for training/fine-tuning and 2 hours for inference. Audio was prepared as 16 kHz mono .wav, with transcripts in .txt (Jain et al., 2022).

Within that supervised setting, PFSTAR provides evidence about data quantity, domain match, and model scale. With adult pretraining and PFSTAR fine-tuning, WER improves from 16.43 to 7.36 to 3.48 for BASE as the PFSTAR fine-tuning data increase from 10 minutes to 1 hour to 10 hours; the corresponding LARGE results are 16.78, 14.19, and 3.50. The best PFSTAR WER reported in the paper is 2.91 on PFS_test, achieved with BASE pretraining and mixed fine-tuning on MyST_55h + PFS_10h. The paper interprets this as evidence that combining broader child-speech exposure with PFSTAR-specific data yields the strongest PFSTAR recognition, while also noting that the LARGE configuration is not clearly better for PFSTAR in low-data settings (Jain et al., 2022).

Taken together, these studies establish PFSTAR as a high-value benchmark for several distinct but related questions. It is used to test zero-shot transfer from adult British English to child British English, to examine age-related variability and its impact on WER and KWS metrics, to measure robustness under additive noise, to study how speaker traits are distributed across SSL layers, and to evaluate how much compactness can be introduced through PCA without sacrificing performance. A plausible implication is that PFSTAR’s importance lies not only in absolute scores, but in the way it exposes mismatches among linguistic content, speaker traits, developmental stage, and acoustic domain.