Search

We propose a novel description-based controllable text-to-speech (TTS) method with cross-lingual control capability. To address the lack of audio-description paired data in the target language, we combine a TTS model trained on the target language with a description control model trained on another language, which maps input text descriptions to the conditional features of the TTS model. These two models share disentangled timbre and style representations based on self-supervised learning (SSL), allowing for disentangled voice control, such as controlling speaking styles while retaining the original timbre. Furthermore, because the SSL-based timbre and style representations are language-agnostic, combining the TTS and description control models while sharing the same embedding space effectively enables cross-lingual control of voice characteristics. Experiments on English and Japanese TTS demonstrate that our method achieves high naturalness and controllability for both languages, even though no Japanese audio-description pairs are used., Comment: Submitted to ICASSP 2025

We propose UNIVERSE++, a universal speech enhancement method based on score-based diffusion and adversarial training. Specifically, we improve the existing UNIVERSE model that decouples clean speech feature extraction and diffusion. Our contributions are three-fold. First, we make several modifications to the network architecture, improving training stability and final performance. Second, we introduce an adversarial loss to promote learning high quality speech features. Third, we propose a low-rank adaptation scheme with a phoneme fidelity loss to improve content preservation in the enhanced speech. In the experiments, we train a universal enhancement model on a large scale dataset of speech degraded by noise, reverberation, and various distortions. The results on multiple public benchmark datasets demonstrate that UNIVERSE++ compares favorably to both discriminative and generative baselines for a wide range of qualitative and intelligibility metrics., Comment: 5 pages, 5 figures, accepted at Interspeech 2024

This paper proposes an audio-conditioned phonemic and prosodic annotation model for building text-to-speech (TTS) datasets from unlabeled speech samples. For creating a TTS dataset that consists of label-speech paired data, the proposed annotation model leverages an automatic speech recognition (ASR) model to obtain phonemic and prosodic labels from unlabeled speech samples. By fine-tuning a large-scale pre-trained ASR model, we can construct the annotation model using a limited amount of label-speech paired data within an existing TTS dataset. To alleviate the shortage of label-speech paired data for training the annotation model, we generate pseudo label-speech paired data using text-only corpora and an auxiliary TTS model. This TTS model is also trained with the existing TTS dataset. Experimental results show that the TTS model trained with the dataset created by the proposed annotation method can synthesize speech as naturally as the one trained with a fully-labeled dataset., Comment: Accepted to INTERSPEECH 2024

We introduce LibriTTS-P, a new corpus based on LibriTTS-R that includes utterance-level descriptions (i.e., prompts) of speaking style and speaker-level prompts of speaker characteristics. We employ a hybrid approach to construct prompt annotations: (1) manual annotations that capture human perceptions of speaker characteristics and (2) synthetic annotations on speaking style. Compared to existing English prompt datasets, our corpus provides more diverse prompt annotations for all speakers of LibriTTS-R. Experimental results for prompt-based controllable TTS demonstrate that the TTS model trained with LibriTTS-P achieves higher naturalness than the model using the conventional dataset. Furthermore, the results for style captioning tasks show that the model utilizing LibriTTS-P generates 2.5 times more accurate words than the model using a conventional dataset. Our corpus, LibriTTS-P, is available at https://github.com/line/LibriTTS-P., Comment: Accepted to INTERSPEECH 2024

For the first time, we introduce a rapid wavelength-swept, passively mode-locked fiber laser in an all-polarization-maintaining and all-fiber configuration. Achieving an exceptional wavelength sweep rate of up to 19 kHz through external modulation of the LD driver pump current, this laser offers a high sweep rate, simple cavity design, cost-effectiveness, and excellent repeatability., Comment: 7 pages, 5 figures, 25 reference

The threshold pump power for modelocking decreased by 18% when the temperature was increased from 25 to 100 degrees C, where a SWCNT/PDMS coated tapered fiber was used as the saturable absorber in a fiber laser. Further, the pump power at which multi-pulse operation began decreased by 24%, and the pump power range over which fundamental modelocking could be maintained decreased by 59% over the same temperature range. This decrease in stability is attributed to the large thermo-optic coefficient of the PDMS polymer, which results in a 40% reduction of the overlap between the evanescent field and SWCNT coating of the taper fiber over a temperature range of 75 degrees C., Comment: 10 pages, 6 figures

Repetition-rate-scanned lasers are practical in accordion frequency comb generation that serves as a variable gearbox connecting optical and radio wave domains. Rapid and wide-range scanned repetition rate can benefit versatile purposes, however scanning robustness remains unsecured that typically requires complicated feedback loops. Recently, multiplexed lasers have been demonstrated with the nature of common-noise rejection among simultaneously emitted combs. Here, we propose a bidirectional polarization-multiplexed fiber laser that delivers synchronized pulses with rapid-scanned and reference-free repetition rates. Benefiting from the all polarization-maintaining fiber configuration, the laser shows good robustness and inter-comb coherence. As rapid as 493.5 kHz/s scanning rate over 329-kHz scanning range of fundamental repetition rate is realized. The 1-hour and 1-day maximal variations of difference frequency are merely 0.52 Hz and 5.46 Hz. The capability to rebuilt steady state after mode hopping is also demonstrated. These results provide a promising solution for developing high-performance accordion-frequency laser sources., Comment: 9 pages

We demonstrate a 783-MHz fundamental repetition rate mode-locked Er-doped all-fiber ring laser with a pulse width of 623 fs. By using carbon nanotubes (CNT) saturable absorber (SA), a relatively low self-starting pump threshold of 108 mW is achieved. The laser has a very compact footprint less than 10 cm * 10 cm, benefiting from the all-active-fiber cavity design. The robust mode-locking is confirmed by the low relative intensity noise (RIN) and a long-term stability test. We propose a new scheme for generating high repetition rate femtosecond optical pulses from a compact and stable all-active-fiber ring oscillator., Comment: 11 pages, 5 figures, 1 table

For the first time, we present the pump power-controlled wavelength-tunable mode-locked fiber laser in the L-band (1565 nm to 1625 nm), achieved by all-polarization maintaining (all-PM) nonlinear polarization rotation (NPR). The wavelength of the laser can be tuned over 20 nm, from 1568.2 nm to 1588.9 nm simply by controlling the pump power from 45 mW to 115 mW. In contrast to conventional wavelength tuning mechanisms such as optical bandpass filters, our tuning method is non-mechanical and electrically controllable, featuring simplicity and cost-effectiveness in a superior all-fiber design., Comment: 5 pages, 4 figures

We propose PromptTTS++, a prompt-based text-to-speech (TTS) synthesis system that allows control over speaker identity using natural language descriptions. To control speaker identity within the prompt-based TTS framework, we introduce the concept of speaker prompt, which describes voice characteristics (e.g., gender-neutral, young, old, and muffled) designed to be approximately independent of speaking style. Since there is no large-scale dataset containing speaker prompts, we first construct a dataset based on the LibriTTS-R corpus with manually annotated speaker prompts. We then employ a diffusion-based acoustic model with mixture density networks to model diverse speaker factors in the training data. Unlike previous studies that rely on style prompts describing only a limited aspect of speaker individuality, such as pitch, speaking speed, and energy, our method utilizes an additional speaker prompt to effectively learn the mapping from natural language descriptions to the acoustic features of diverse speakers. Our subjective evaluation results show that the proposed method can better control speaker characteristics than the methods without the speaker prompt. Audio samples are available at https://reppy4620.github.io/demo.promptttspp/., Comment: Accepted to ICASSP 2024

We investigated the inner buried nucleus of a nearby luminous infrared galaxy NGC 4418 using high-resolution spectroscopy of fundamental carbon monoxide (CO) ro-vibrational absorptions around $4.67 \mu$m for the first time. This method allowed us to examine the physical and kinematical properties in the hot inner region of this nucleus. We detected a series of both very deep (partly saturated) $^{12}$CO and moderately deep (optically thin) $^{13}$CO absorption lines and inferred a large column density ($N_\mathrm{H2}=(5\pm3)\times10^{23}$ cm$^{-2}$ in front of the $5 \mu$m photosphere) of warm ($T_\mathrm{ex}\simeq170$ K) molecular gas by assuming an isothermal plane-parallel slab illuminated by a compact background MIR-emitting source. We modeled that the warm CO absorber almost covers the central heating source and that it is an inner layer around the $5 \mu$m photosphere (at $r=$several pc) of a compact shroud of gas and dust ($d\sim100$ pc). The width of the absorption lines ($110$ km s$^{-1}$) and their small deviation from the systemic velocity ($<10$ km s$^{-1}$) are consistent with a warm and turbulent layer with little bulk motion in the radial direction., Comment: 18 pages, 3 figures, 2 tables, accepted for publication in ApJ

We propose a lightweight end-to-end text-to-speech model using multi-band generation and inverse short-time Fourier transform. Our model is based on VITS, a high-quality end-to-end text-to-speech model, but adopts two changes for more efficient inference: 1) the most computationally expensive component is partially replaced with a simple inverse short-time Fourier transform, and 2) multi-band generation, with fixed or trainable synthesis filters, is used to generate waveforms. Unlike conventional lightweight models, which employ optimization or knowledge distillation separately to train two cascaded components, our method enjoys the full benefits of end-to-end optimization. Experimental results show that our model synthesized speech as natural as that synthesized by VITS, while achieving a real-time factor of 0.066 on an Intel Core i7 CPU, 4.1 times faster than VITS. Moreover, a smaller version of the model significantly outperformed a lightweight baseline model with respect to both naturalness and inference speed. Code and audio samples are available from https://github.com/MasayaKawamura/MB-iSTFT-VITS., Comment: Accepted to ICASSP 2023

Several fully end-to-end text-to-speech (TTS) models have been proposed that have shown better performance compared to cascade models (i.e., training acoustic and vocoder models separately). However, they often generate unstable pitch contour with audible artifacts when the dataset contains emotional attributes, i.e., large diversity of pronunciation and prosody. To address this problem, we propose Period VITS, a novel end-to-end TTS model that incorporates an explicit periodicity generator. In the proposed method, we introduce a frame pitch predictor that predicts prosodic features, such as pitch and voicing flags, from the input text. From these features, the proposed periodicity generator produces a sample-level sinusoidal source that enables the waveform decoder to accurately reproduce the pitch. Finally, the entire model is jointly optimized in an end-to-end manner with variational inference and adversarial objectives. As a result, the decoder becomes capable of generating more stable, expressive, and natural output waveforms. The experimental results showed that the proposed model significantly outperforms baseline models in terms of naturalness, with improved pitch stability in the generated samples., Comment: ICASSP 2023

The decay of the primordial isotopes $^{238}\mathrm{U}$, $^{235}\mathrm{U}$, $^{232}\mathrm{Th}$, and $^{40}\mathrm{K}$ have contributed to the terrestrial heat budget throughout the Earth's history. Hence the individual abundance of those isotopes are key parameters in reconstructing contemporary Earth model. The geoneutrinos produced by the radioactive decays of uranium and thorium have been observed with the Kamioka Liquid-Scintillator Antineutrino Detector (KamLAND). Those measurements have been improved with more than 18-year observation time, and improvements in detector background levels mainly by an 8-year almost rector-free period now permit spectroscopy with geoneutrinos. Our results yield the first constraint on both uranium and thorium heat contributions. Herein the KamLAND result is consistent with geochemical estimations based on elemental abundances of chondritic meteorites and mantle peridotites. The High-Q model is disfavored at 99.76% C.L. and a fully radiogenic model is excluded at 5.2$\sigma$ assuming a homogeneous heat producing element distribution in the mantle., Comment: 13 pages, 2 figures, accepted for publication in Geophysical Research Letters on Aug 4th, 2022

Data augmentation via voice conversion (VC) has been successfully applied to low-resource expressive text-to-speech (TTS) when only neutral data for the target speaker are available. Although the quality of VC is crucial for this approach, it is challenging to learn a stable VC model because the amount of data is limited in low-resource scenarios, and highly expressive speech has large acoustic variety. To address this issue, we propose a novel data augmentation method that combines pitch-shifting and VC techniques. Because pitch-shift data augmentation enables the coverage of a variety of pitch dynamics, it greatly stabilizes training for both VC and TTS models, even when only 1,000 utterances of the target speaker's neutral data are available. Subjective test results showed that a FastSpeech 2-based emotional TTS system with the proposed method improved naturalness and emotional similarity compared with conventional methods., Comment: Accepted to INTERSPEECH 2022

We report the first measurement of the zodiacal light (ZL) polarization spectrum in the near-infrared between 0.8 and 1.8 $\mu$m. Using the low-resolution spectrometer (LRS) on board the Cosmic Infrared Background Experiment (CIBER), calibrated for absolute spectrophotometry and spectropolarimetry, we acquire long-slit polarization spectral images of the total diffuse sky brightness towards five fields. To extract the ZL spectrum, we subtract contribution of other diffuse radiation, such as the diffuse galactic light (DGL), the integrated star light (ISL), and the extragalactic background light (EBL). The measured ZL polarization spectrum shows little wavelength dependence in the near-infrared and the degree of polarization clearly varies as a function of the ecliptic coordinates and solar elongation. Among the observed fields, the North Ecliptic Pole shows the maximum degree of polarization of $\sim$ 20$\%$, which is consistent with an earlier observation from the Diffuse Infrared Background Experiment (DIRBE) aboard on the Cosmic Background Explorer (COBE). The measured degree of polarization and its solar elongation dependence are reproduced by the empirical scattering model in the visible band and also by the Mie scattering model for large absorptive particles, while the Rayleigh scattering model is ruled out. All of our results suggest that the interplanetary dust is dominated by large particles., Comment: 15 pages, 15 figures, accepted by ApJ

Scientific communities are increasingly adopting machine learning and deep learning models in their applications to accelerate scientific insights. High performance computing systems are pushing the frontiers of performance with a rich diversity of hardware resources and massive scale-out capabilities. There is a critical need to understand fair and effective benchmarking of machine learning applications that are representative of real-world scientific use cases. MLPerf is a community-driven standard to benchmark machine learning workloads, focusing on end-to-end performance metrics. In this paper, we introduce MLPerf HPC, a benchmark suite of large-scale scientific machine learning training applications driven by the MLCommons Association. We present the results from the first submission round, including a diverse set of some of the world's largest HPC systems. We develop a systematic framework for their joint analysis and compare them in terms of data staging, algorithmic convergence, and compute performance. As a result, we gain a quantitative understanding of optimizations on different subsystems such as staging and on-node loading of data, compute-unit utilization, and communication scheduling, enabling overall $>10 \times$ (end-to-end) performance improvements through system scaling. Notably, our analysis shows a scale-dependent interplay between the dataset size, a system's memory hierarchy, and training convergence that underlines the importance of near-compute storage. To overcome the data-parallel scalability challenge at large batch sizes, we discuss specific learning techniques and hybrid data-and-model parallelism that are effective on large systems. We conclude by characterizing each benchmark with respect to low-level memory, I/O, and network behavior to parameterize extended roofline performance models in future rounds.

We conducted systematic observations of the H I Br$\alpha$ (4.05 $\mu$m) and Br$\beta$ (2.63 $\mu$m) lines in 52 nearby ($z<0.3$) ultraluminous infrared galaxies (ULIRGs) with AKARI. Among 33 ULIRGs wherein the lines are detected, three galaxies show anomalous Br$\beta$/Br$\alpha$ line ratios ($\sim1.0$), which are significantly higher than those for case B (0.565). Our observations also show that ULIRGs have a tendency to exhibit higher Br$\beta$/Br$\alpha$ line ratios than those observed in Galactic H II regions. The high Br$\beta$/Br$\alpha$ line ratios cannot be explained by a combination of dust extinction and case B since dust extinction reduces the ratio. We explore possible causes for the high Br$\beta$/Br$\alpha$ line ratios and show that the observed ratios can be explained by a combination of an optically thick Br$\alpha$ line and an optically thin Br$\beta$ line. We simulated the H II regions in ULIRGs with the Cloudy code, and our results show that the high Br$\beta$/Br$\alpha$ line ratios can be explained by high-density conditions, wherein the Br$\alpha$ line becomes optically thick. To achieve a column density large enough to make the Br$\alpha$ line optically thick within a single H II region, the gas density must be as high as $n\sim10^8$ $\mathrm{cm}^{-3}$. We therefore propose an ensemble of H II regions, in each of which the Br$\alpha$ line is optically thick, to explain the high Br$\beta$/Br$\alpha$ line ratio., Comment: 28 pages, 15 figures, accepted for publication in ApJ

A theoretical model of the high-beta equilibrium of magnetospheric plasma was constructed by consistently connecting the (anisotropic pressure) Grad-Shafranov equation and the Vlasov equation. The Grad-Shafranov equation was used to determine the axisymmetric magnetic field for a given magnetization current corresponding to a pressure tensor. Given a magnetic field, we determine the distribution function as a specific equilibrium solution of the Vlasov equation, using which we obtain the pressure tensor. We need to find an appropriate class of distribution function for these two equations to be satisfied simultaneously. Here, we consider the distribution function that maximizes the entropy on the submanifold specified by the magnetic moment. This is equivalent to the reduction of the canonical Poisson bracket to the noncanonical one having the Casimir corresponding to the magnetic moment. The pressure tensor then becomes a function of the magnetic field (through the cyclotron frequency) and flux function, satisfying the requirement of the Grad-Shafranov equation. Numerical solutions have been obtained to interpret the experimental data of the RT-1 laboratory magnetosphere., Comment: 19 pages, 8 figures

Understanding the inner structure of the clumpy molecular torus surrounding the active galactic nucleus is essential in revealing the forming mechanism. However, spatially resolving the torus is difficult because of its size of a few parsecs. Thus, to probe the clump conditions in the torus, we performed the velocity decomposition of the CO rovibrational absorption lines ($\Delta{v}=0\to 1,\ \Delta{J}=\pm 1$) at $\lambda\sim 4.67\,\mathrm{\mu{m}}$ observed toward an ultraluminous infrared galaxy IRAS 08572+3915 NW with the high-resolution spectroscopy ($R\sim 10{,}000$) of Subaru Telescope. Consequently, we found that each transition had two outflowing components, i.e., (a) and (b), both at approximately $\sim -160\,\mathrm{km\,s^{-1}}$, but with broad and narrow widths, and an inflowing component, i.e., (c), at approximately $\sim +100\,\mathrm{km\,s^{-1}}$, which were attributed to the torus. The ratios of the velocity dispersions of each component lead to those of the rotating radii around the black hole of $R_\mathrm{rot,a}:R_\mathrm{rot,b}:R_\mathrm{rot,c}\approx 1:5:17$, indicating the torus where clumps are outflowing in the inner regions and inflowing in the outer regions if a hydrostatic disk with $\sigma_V\propto R_\mathrm{rot}^{-0.5}$ is assumed. Based on the kinetic temperature of components (a) and (b) of $\sim 720\,\mathrm{K}$ and $\sim 25\,\mathrm{K}$ estimated from the level population, the temperature gradient is $T_\mathrm{kin}\propto R_\mathrm{rot}^{-2.1}$. Magnetohydrodynamic models with large density fluctuations of two orders of magnitude or more are necessary to reproduce this gradient., Comment: Accepted for publication in ApJ, 27 pages, 13 figures, 6 tables, Minor changes after proof (v3)

The KamLAND-Zen 800 experiment is searching for the neutrinoless double-beta decay of $^{136}$Xe by using $^{136}$Xe-loaded liquid scintillator. The liquid scintillator is enclosed inside a balloon made of thin, transparent, low-radioactivity film that we call Inner Balloon (IB). The IB, apart from guaranteeing the liquid containment, also allows to minimize the background from cosmogenic muon-spallation products and $^{8}$B solar neutrinos. Indeed these events could contribute to the total counts in the region of interest around the Q-value of the double-beta decay of $^{136}$Xe. In this paper, we present an overview of the IB and describe the various steps of its commissioning minimizing the radioactive contaminations, from the material selection, to the fabrication of the balloon and its installation inside the KamLAND detector. Finally, we show the impact of the IB on the KamLAND background as measured by the KamLAND detector itself., Comment: 23 pages, 16 figures, to be submitted to JINST

There are many theories where a dark matter particle is part of a multiplet with an electrically charged state. If WIMP dark matter ($\chi^{0}$) is accompanied by a charged excited state ($\chi^{-}$) separated by a small mass difference, it can form a stable bound state with a nucleus. In supersymmetric models, the $\chi^{0}$ and the $\chi^{-}$ could be the neutralino and a charged slepton, such as the neutralino-stau degenerate model. The formation binding process is expected to result in an energy deposition of {\it O}(1--10 MeV), making it suitable for detection in large liquid scintillator detectors. We describe new constraints on the bound state formation with a xenon nucleus using the KamLAND-Zen 400 Phase-II dataset. In order to enlarge the searchable parameter space, all xenon isotopes in the detector were used. For a benchmark parameter set of $m_{\chi^{0}} = 100$ GeV and $\Delta m = 10$ MeV, this study sets the most stringent upper limits on the recombination cross section $\langle\sigma v\rangle$ and the decay-width of $\chi^{-}$ of $2.0 \times 10^{-31}$ ${\rm cm^3/s}$ and $1.1 \times 10^{-18}$ GeV, respectively (90\% confidence level).