XEUS: Speech Encoder & X-ray Telescope

• XEUS is a polysemous term denoting both a multilingual self-supervised speech encoder trained on 1.08 million hours of audio and a future X-ray telescope concept.
• In speech research, XEUS employs a HuBERT-style encoder with 19 E-Branchformer layers and is evaluated across multilingual ASR and child speech benchmarks.
• In X-ray optics, XEUS describes a telescope design analyzed via HEW simulations to assess mirror performance under surface microroughness-induced scattering.

XEUS denotes two distinct research objects in the cited arXiv literature. In speech processing, XEUS is a Cross-lingual Encoder for Universal Speech: a self-supervised encoder trained on 1.081 million hours of unlabeled audio spanning 4,057 ISO3 languages, with the explicit aim of robust multilingual speech representation learning (Chen et al., 2024). In an older X-ray-optics usage, XEUS refers to a future X-ray telescope whose mirror-module angular resolution was analyzed through Half-Energy Width (HEW) simulations under surface-microroughness-induced X-ray scattering (Spiga, 2015). The speech model is also used as a frozen front-end in child-ASR experiments conducted in ESPnet, where its multilingual pretraining is evaluated against child-speech-specific flat-start training strategies (Ying et al., 22 Aug 2025).

1. Terminology and scope

In the recent speech-technology literature, XEUS is introduced as a multilingual SSL model that extends the language coverage of speech pretraining to 4,057 languages and uses only publicly accessible pretraining data (Chen et al., 2024). The same acronym appears in earlier astronomy and optics literature as the name of a future X-ray telescope concept analyzed via HEW and X-ray scattering simulations (Spiga, 2015). The cited sources therefore treat “XEUS” as a polysemous term whose meaning is determined entirely by disciplinary context.

Within speech research, XEUS is positioned relative to HuBERT-style encoder-only SSL, WavLM-style denoising, and E-Branchformer acoustic modeling. Within child-ASR research, it appears not as a fully fine-tuned SSL system, but as a frozen feature extractor placed ahead of downstream E-Branchformer ASR encoders in controlled comparisons against WavLM and 80-dimensional Fbanks (Ying et al., 22 Aug 2025).

2. Speech-model architecture

XEUS follows a HuBERT-style encoder-only design but replaces standard Transformer blocks with E-Branchformer layers (Chen et al., 2024). Its feature extractor is a 7-layer convolutional front end identical to HuBERT, mapping the raw 16 kHz waveform into 512-dimensional frame-level features. The main encoder has 19 layers. Each layer contains two parallel branches: multi-head self-attention with 8 heads, model dimension $d=1024$, and key/value dimension $d_k=d_v=128$; and a gated MLP with convolution, using two feed-forward sublayers of hidden size 4,096 and convolution kernel size 31. The branch outputs are merged through residual connections and gated linear units (GLU) as in Branchformer by Peng et al.

The prediction head is a linear projection from the last encoder layer to a 2,048-way softmax that predicts discrete pseudo-labels. The model uses cross-entropy loss in place of the original HuBERT squared-error loss, following the “Fast-HuBERT” modification attributed in the source to Yang et al. The total model size is reported as 577 M parameters.

The child-ASR benchmark confirms the same high-level topology—a self-supervised encoder built on 19 E-Branchformer layers—but uses XEUS only as a fixed front-end rather than reoptimizing its internal parameters (Ying et al., 22 Aug 2025). In that setting, XEUS is paired with a 12-layer E-Branchformer ASR encoder and evaluated with several decoder families, which isolates the utility of its pretrained representations from end-to-end SSL adaptation.

3. Pretraining corpus, objectives, and optimization

XEUS is pretrained on 1.081 million hours of unlabeled audio collected from 37 public datasets plus newly crawled corpora, with a total coverage of 4,057 ISO3 languages (Chen et al., 2024). The data summary reported for the model is:

Source	Languages	Hours
Public corpora (37 datasets)	150+	1,074,000
Newly crawled corpora	4,057	7,413
Total	4,057	1,081,413

The public portion includes YODAS, VoxPopuli, LibriLight, MLS, Common Voice 13, MP3 podcasts, Googlei18n, BABEL, FLEURS, and smaller datasets covering code-switching, singing, accented speech, and indigenous languages such as Quechua, Mixtec, and Totonac. The newly crawled component includes MMS-unlab v2 with 6,700 h from 4,023 languages, WikiTongues with 70 h from approximately 700 languages, and Jesus Dramas with 643 h from 430 languages. All audio is unified to 16 kHz and segmented by energy-based VAD into utterances of at most 40 s.

Pseudo-labels are constructed by sampling 20 K h for feature extraction with a pretrained WavLabLM MS encoder and then applying K-means with $K=2{,}048$. XEUS does no explicit down-/upsampling of languages, so the pretraining distribution remains strongly long-tailed: the top 50 languages account for 99.5% of data, while approximately 2,000 languages have at least 1 h of speech each.

The objective combines three components in a single training regime: masked prediction, implicit acoustic denoising, and a dereverberation task (Chen et al., 2024). The core loss is cross-entropy on masked frames,

$$\mathcal{L}_{\mathrm{mask}} = - \sum_{t\in \mathcal{M}} \log p_{\theta}\bigl(z_t \mid \tilde{x}\bigr),$$

where $\tilde{x}$ is produced by masking and audio augmentation. Noise augmentation is applied with probability $p_{\text{noise}}=0.2$, split evenly between random noise and utterance mixing. Dereverberation is enabled with probability $p_r=0.3$ by convolving the waveform with a random room impulse response,

$r = x \,\circledast\, h,$

while still predicting pseudo-labels produced from the clean teacher signal. During the first 3,000 steps, XEUS also uses an intermediate cross-entropy loss on layer 10, weighted by 0.3, with masking ratio 0.65 and no augmentations; from steps 3,001–670 K, the masking ratio increases to 0.8, augmentations are enabled, and the intermediate loss is removed.

Training uses 64× A100 40 GB GPUs, bfloat16, and FlashAttention v2. The batch size is 100 s per GPU, for 106 min total per step. Optimization uses Adam with 32 K warmup steps to a peak learning rate of $3\times 10^{-4}$, followed by linear decay. Total training length is 670 K steps, described as approximately 1 M hours × 2 passes. The reported compute cost is 100 K CPU h for data preparation, 15 K GPU h for pseudo-labeling, 2.3 K GPU h for ablations and tuning, and 63 K GPU h for pretraining.

4. Benchmark performance and robustness profile

XEUS is evaluated against XLS-R 128, MMS, and w2v-BERT 2.0 v2 across multilingual ASR, language identification, speech translation, English SUPERB, and speech resynthesis (Chen et al., 2024). On ML-SUPERB, using lightweight fine-tuning on 143 languages under 10 min and 1 h supervision budgets, XEUS attains SUPERB$_s$: 956 / 956. Its reported task-level results are monolingual ASR CER 33.3 / 25.7, multilingual ASR CER 21.1 / 20.1, LID ACC 81.5 / 87.3, and ASR+LID CER/ACC 22.9 / 32.7. The source states that XEUS outperforms MMS 1B by 0.8% SUPERB$d_k=d_v=128$0 and w2v-BERT 2.0 v2 by 4.4%, while using less pretraining data than w2v-BERT 2.0 v2.

On FLEURS, XEUS reaches CER 8.9 and LID ACC 93.0, while on JesusFilm low-resource speech translation it achieves chrF 22.1. On the English-only SUPERB benchmark, the excerpted leaderboard reports XEUS as first on Recognition WER 3.34, Keyword Spotting 98.32, Speaker SD 3.11, and Paralinguistics ER 71.08. For speech resynthesis on VCTK, XEUS obtains MOS 3.23, WER 10.0, log-F0 0.25, and MCD 3.80.

The ablations are as important as the headline numbers. On a 7,000 h subset evaluated on 1 h multi-ASR+LID, adding denoising reduces CER from 36.9 to 30.5, and then replacing the backbone with E-Branchformer further reduces it to 29.0, with LID ACC increasing from 78.5 to 84.3 and then 85.2. A separate dereverberation experiment on English ASR reports HuBERT base at 13.1 / 22.6 WER on test-clean / test-other, improved to 12.7 / 21.1 with dereverberation. These results indicate that XEUS is not defined solely by scale or language coverage; the denoising and dereverberation objectives are reported as materially contributing to robustness.

5. Use as a frozen front-end in child ASR

In the ESPnet child-ASR benchmark, XEUS is evaluated only as a frozen SSL feature extractor; fine-tuning adult-pretrained checkpoints is not done with XEUS in this work (Ying et al., 22 Aug 2025). The downstream ASR stack consists of a 12-layer E-Branchformer encoder over XEUS features. The decoders tested are AED with a 6-layer Transformer, 4 heads, and 2,048-unit feed-forward layers; RNN-T; and CTC with or without a 2-layer LSTM LM. No additional data augmentation beyond SpecAugment is used, and the optimizer is Adam with warm-up learning-rate schedule, held identical across front-end choices.

The benchmark reports the following single-corpus flat-start WER dev/test results for XEUS: MyST 11.6 / 13.0, OGI Script 3.7 / 3.7, OGI Spontaneous 49.3 / 48.2, and CMU Kids 40.5 / 39.6. In the multi-corpus flat-start setting, the corresponding XEUS results are MyST 12.2 / 13.2, OGI Script 4.8 / 6.4, OGI Spontaneous 24.2 / 25.4, and CMU Kids 16.0 / 16.0. Age-stratified WER for E-Branchformer + XEUS AED is reported as 23.4% for ages 4–7, 11.3% for 8–10, and 5.0% for 11–15.

The comparative findings are narrower than XEUS’s general multilingual benchmark profile. In single-corpus flat-start child ASR, XEUS improves on Fbanks for several conditions, for example 13.0% vs. 14.8% on MyST and 3.7% vs. 9.7% on OGI Script, though it does not uniformly surpass WavLM (Ying et al., 22 Aug 2025). In multi-corpus flat-start, however, the benchmark observes that both WavLM and XEUS degrade relative to Fbanks, which it interprets as evidence that existing SSL front-ends carry adult-speech biases. The source further states that flat-start child-only training can only partially overcome these biases, and that when trained exclusively on multi-corpus child data, Fbank features actually outperform the SSL front-ends, in part because child data “retune” the model’s acoustic mapping to children’s formant and prosodic ranges.

This result is methodologically significant because it constrains a common extrapolation from multilingual SSL benchmarks: strong large-scale pretraining performance does not automatically imply front-end optimality for child speech. The same benchmark also separates ASR and speaker verification: XEUS is not used for SV, which is instead performed with RawNet3 + AM-Softmax (Ying et al., 22 Aug 2025).

6. Reproducibility, engineering, and stated limitations

XEUS is accompanied by unusually explicit reproducibility assets. The model paper states that all data, checkpoints, training configs, intermediate logs, and code are publicly released at the WavLab project site (Chen et al., 2024). It also documents engineering optimizations in ESPnet: removing unnecessary GPU synchronizations yields +120% throughput, while length-aware batching yields +60% throughput and −113% memory usage. In the child-ASR benchmark, the entire evaluation stack is likewise public through the ESPnet recipe directory, including data preparation, model configs, training scripts, hyperparameter settings, decoding scripts, and WER/CER scoring pipelines; XEUS front-end extraction is handled by feats extract_xeus.sh (Ying et al., 22 Aug 2025).

The limitations reported for XEUS are specific. First, the long tail remains difficult: many languages still have less than 1 h of pretraining audio, and downstream performance remains lower than for high-resource languages (Chen et al., 2024). Second, the evaluation suite emphasizes lightweight fine-tuning, which the source notes only approximates large-scale downstream SOTA. Third, privacy and ethics are explicitly addressed through CC licenses only, speaker anonymization, and non-commercial release, with the additional requirement that users respect cultural norms and consent.

The child-ASR study introduces an additional interpretive constraint. Its scaling curves concern OWLS and Whisper, not XEUS itself, because XEUS is a fixed 19-layer model in that benchmark (Ying et al., 22 Aug 2025). The reported observation that WER improves reliably up to approximately 1 B parameters and then plateaus or degrades therefore does not constitute an empirical scaling law for XEUS. A plausible implication is that XEUS should be interpreted as a representation-learning component whose downstream behavior depends strongly on domain match and adaptation strategy, rather than as a member of the same scaling series studied for OWLS or Whisper.

7. Earlier X-ray-optics usage of XEUS

In the X-ray-optics literature summarized by Spiga, XEUS appears as a future X-ray telescope analyzed through the energy dependence of the Half-Energy Width (HEW) (Spiga, 2015). HEW at photon energy $d_k=d_v=128$1 is defined through the encircled-energy function

$d_k=d_v=128$2

with half-energy radius $d_k=d_v=128$3 satisfying $d_k=d_v=128$4, so that

$d_k=d_v=128$5

The total HEW is modeled as the quadrature sum of an approximately energy-independent figure-error term $d_k=d_v=128$6 and an energy-dependent X-ray-scattering term $d_k=d_v=128$7:

$d_k=d_v=128$8

Surface microroughness is described through a one-dimensional Power Spectral Density $d_k=d_v=128$9, where $K=2{,}048$0. For a double-reflection shell, the cutoff frequency $K=2{,}048$1 is obtained from

$K=2{,}048$2

and the corresponding scattering contribution is

$K=2{,}048$3

If the PSD follows a power law,

$K=2{,}048$4

then

$K=2{,}048$5

The source emphasizes that steeper PSDs, corresponding to larger $K=2{,}048$6, weaken the energy dependence of the scattering term.

The XEUS mirror-module baseline is Silicon Pore Optics (HPO) assembled into confocal nested shells up to 4.2 m diameter with 35 m focal length. Grazing-incidence angles range from $K=2{,}048$7 to $K=2{,}048$8. The assumed figure-error contribution is $K=2{,}048$9, consistent with an on-axis requirement of $$\mathcal{L}_{\mathrm{mask}} = - \sum_{t\in \mathcal{M}} \log p_{\theta}\bigl(z_t \mid \tilde{x}\bigr),$$0 at 1 keV, degrading to approximately $$\mathcal{L}_{\mathrm{mask}} = - \sum_{t\in \mathcal{M}} \log p_{\theta}\bigl(z_t \mid \tilde{x}\bigr),$$1 at 30 keV. The total module HEW is approximated by an effective-area-weighted shell average,

$$\mathcal{L}_{\mathrm{mask}} = - \sum_{t\in \mathcal{M}} \log p_{\theta}\bigl(z_t \mid \tilde{x}\bigr),$$2

Two representative PSD models are reported: $$\mathcal{L}_{\mathrm{mask}} = - \sum_{t\in \mathcal{M}} \log p_{\theta}\bigl(z_t \mid \tilde{x}\bigr),$$3 and $$\mathcal{L}_{\mathrm{mask}} = - \sum_{t\in \mathcal{M}} \log p_{\theta}\bigl(z_t \mid \tilde{x}\bigr),$$4, both normalized to measured roughness of commercial MEMC silicon wafers (Spiga, 2015). The simulated module HEW values are $$\mathcal{L}_{\mathrm{mask}} = - \sum_{t\in \mathcal{M}} \log p_{\theta}\bigl(z_t \mid \tilde{x}\bigr),$$5 at 1 keV for both PSDs; at 10 keV, approximately $$\mathcal{L}_{\mathrm{mask}} = - \sum_{t\in \mathcal{M}} \log p_{\theta}\bigl(z_t \mid \tilde{x}\bigr),$$6 for $$\mathcal{L}_{\mathrm{mask}} = - \sum_{t\in \mathcal{M}} \log p_{\theta}\bigl(z_t \mid \tilde{x}\bigr),$$7 and $$\mathcal{L}_{\mathrm{mask}} = - \sum_{t\in \mathcal{M}} \log p_{\theta}\bigl(z_t \mid \tilde{x}\bigr),$$8 for $$\mathcal{L}_{\mathrm{mask}} = - \sum_{t\in \mathcal{M}} \log p_{\theta}\bigl(z_t \mid \tilde{x}\bigr),$$9; at 20 keV, approximately $\tilde{x}$0 and $\tilde{x}$1; and at 40 keV, approximately $\tilde{x}$2 and $\tilde{x}$3. To satisfy the cited XEUS requirement of HEW $\tilde{x}$4 up to approximately 30 keV, the PSD must roll off steeply at high spatial frequency. The derived rms roughness tolerances are $\tilde{x}$5 for 1 mm–100 $\tilde{x}$6m, $\tilde{x}$7 for 100 $\tilde{x}$8m–10 $\tilde{x}$9m, $p_{\text{noise}}=0.2$0 for 10 $p_{\text{noise}}=0.2$1m–1 $p_{\text{noise}}=0.2$2m, and $p_{\text{noise}}=0.2$3 for $p_{\text{noise}}=0.2$4m$. In that literature, “XEUS” therefore designates an optics-constrained telescope concept rather than a speech representation model, despite the acronymic overlap.