Research topics
xxx
To appear.
Dialogue-history-aware speech synthesis (2022)
To appear.
Automatic prediction of synthetic speech quality (2022)
To appear.
Self-supervised audio restoration (2022)
To appear.
Emotion prediction of nonverbal vocalization (2022)
To appear.
Speech pseudonymization (2022)
To appear.
Personalized filled-pause prediction (2022)
To appear.
Emotion-controllable speech synthesis (2021)
To appear.
Digital speech makeup (2021)
To appear.
Audiobook speech synthesis considering cross-sentence contexts (2021)
To appear.
Accent modeling for dialect speech synthesis (2021)
To appear.
Cross-lingual TTS as domain adaptation (2021)
To appear.
Computational speech chain (2021)
To appear.
Onoma-to-wave (2021)
To appear.
VAE-based voice conversion (2021)
To appear.
Inter-singer similarity in Japanese song (2020)
To appear.
Crowd-powered anonymization (2020)
To appear.
Stereo electrolarynx (2020)
To appear.
VAE-based accent retrieval (2020)
To appear.
manga2voice (2020)
To appear.
Lightweight voice anonymization (2020)
To appear.
Phoneme-balanced singing voice corpus (2020)
To appear.
Cross-lingual TTS via domain adaptation (2020)
To appear.
Real-time, full-band, light-weight voice conversion (2020)
To appear.
HumanGAN (2020)
To appear.
Mask-shaped voice conversion device (2019)
To appear.
Speaker V2S (verification-to-synthesis) attack (2019)
To appear.
Speaker representation using human perception (2019)
To appear.
Real-time DNN-based voice conversion (2019)
To appear.
Unsupervised subword tokenization for end-to-end speech synthesis (2019)
To appar.
心地よく使える高品質音声変換を目指し，リアルタイム広帯域音声変換と，変換エラーの定量化・緩和を目指します．
個別化言語教育に向け，外国人による非流暢な日本語音声から，その人の声色で，その人より流暢な日本語音声を合成します．
Neural double-tracking (2019)
人間の「歌声の間違い方」を学習して，自然な多重録音感を付与します． It provides natural "doubletrackedness" to singing voices.
tamaru19icassp_ndt_poster.pdfNMF/VAE-based anomaly detection under noisy environments (2018)
雑音環境下における異常音検出のため，VAEとNMFに基づく異常検出法を提案します． VAE/NMF-based detection of anomaly audio under noisy environments.
aiba18otogaku_noise-nmf.pdfMMD-based data augmentation for speaker recognition (2018)
話者照合の精度向上のために，MMDに基づくデータ拡張を行います．Introduce MMD-based data augmentation for improving speaker verification accuracy.
shiota18apsipa_asv-augmentation.pdfAutomated jazz harmonization (2018)
ジャズ音楽のための自動和声付けを行います．DNN-based jazz harmonization
Crowdworkers' creditability for sound localization evaluation (2018)
クラウドソーシングに基づく音像定位評価のために，参加者の信頼度を導入します．Introduce listeners' creditability for sound localization evaluation using crowdsourcing.
takamichi18asja_crowd_poster.pdfLow-musical-noise speech enhancement (2018)
処理による不快な人口雑音 (ミュージカルノイズ) の発生を抑圧するDNN音声強調です．DNN-based speech enhancement to reduce harmful noise (a.k.a., musical noise).
mizoguchi18ea11_kurtosis-ratio_poster.pdfSpeech-based automated dementia detection (2018)
音声の長期変動成分により，認知症を自動検出します．Automated dementia detection using long-term components of speech spectra.
sonobe18asjs_dementia_poster.pdfPhase modeling with directional-statistics DNN (2018)
方向統計DNNを用いて，音声信号の位相情報をモデル化します．Modeling phase spectrograms of speech signals based on directional-statistics DNN
takamichi18iwaenc_phase_poster.pdfMulti-dialect speech synthesis (2018)
日本の色んな方言を1人の声で話してくれる音声合成です．Multi-Japanese-dialect speech synthesis with one speaker's voice.
akiyama18apsipa_dialect_slide.pdfPre-processing for vocoder-free speech synthesis (2017)
WaveNetなどのボコーダフリー音声合成の学習精度を上げるための前処理法です．A pre-processing method for improving training accuracy of vocoder-free speech synthesis such as WaveNet.
敵対的学習 (GAN) を導入して「声のなりすまし検出器」を騙すように音声合成をアップデートします． Introducing generative adversarial networks, we train speech synthesis to deceive anti-spoofing verification (security technology to distinguish natural/synthetic speech.)
計算機が人間の様に「喋るたびにちょっとずつ違う」音声 (一期一会な音声) を生成します．Computers can speak speech that has human-like inter-utterance randomness.
takamichi17interspeech_poster.pdf音声変換の品質を競うVoice conversion challenge 2016にエントリーし，話者再現度に関して世界最高精度を獲得しました．Report of Voice Conversion Challenge 2016. Our system achieved the 1st place in speaker similarity.
kobayashi2016interspeech.pdf言語学習者の声で，その人が将来話すであろう流暢な音声を作り出します．Speech technologies for language learning. We can synthesize your fluent (native-like) speech from your non-fluent (non-native) speech.
テキスト音声合成の品質を競うBlizzard Challenge 2015の結果報告です．3つのインド言語において，世界最高の音質と評価されました．Report of Blizzard Challenge 2015 which compares performance of Text-to-speech synthesizer. Our system achieved the best speech quality in three Indian languages.
音声特徴量の時間変動を表す変調スペクトルを持ちいて，高品質音声合成を実現します．Speech synthesis incorporating the modulation spectra that represent temporal fluctuation of speech parameters.
