Saijo retweeted
私のチームが @merl_newsと一緒に取り組んで来た、タスク汎用音源分離が広報発表されました!開発の裏側をnoteの記事にしてもらったので、是非ご覧ください。今月の音響学会では、@kohei_1979 がこの技術の詳細を発表します。
Saijo retweeted
🔥Announcing the VoiceMOS Challenge 2026🔥
We are REVERTING BACK to the VoiceMOS Challenge, as there are still unsolved problems in the evaluation of speech itself 🤨
Pre-registration: forms.gle/L6YdkUf1PJdSSwLU7
Website: sites.google.com/view/voicem…
See the tracks in the thread!
Saijo retweeted
We are looking for a postdoctoral researcher in speech and audio processing, with a possible start in the Fall 2026 semester. If you are interested in working with us, please apply through the following form: forms.gle/gfENMMrRf1nmnTJeA
Saijo retweeted
テックブログを公開しました。
「Scaling Speech AI」の下、1Bから3Bへと音声言語モデルをスケールさせた際のTTS性能へ影響を検証しました。
日本語特有の読みや表記揺れ、表現の広がりがみられるに加え、現状の課題についても整理しています。
日本語音声生成・SpeechLM・TTSに関心のある方はぜひご覧ください。
blog.dubguild.com/melte/llm-…
1B/3Bモデルの構築にあたって実施した、データ前処理・事前学習・事後学習の詳細も、今後順次公開していく予定です。
続報もお待ちいただければ幸いです。
先日博士課程を修了しました! 3年間多くの方々にお世話になりました。ありがとうございました。
修了記念に公聴会のスライドを公開しました: speakerdeck.com/kohei_1979/u…
今後も研究は続けていく予定なので、引き続きよろしくお願いいたします。
“Input-Adaptive Spectral Feature Compression by Sequence Modeling for Source Separation”, w/ @yoshipon0520, was accepted to IEEE TASLPRO!
Introduced feature compression by seq. modeling + psychoacoustic inductive bias, outperforming band-splitting
arxiv.org/abs/2602.08671
Codes have been published!
github.com/b-sigpro/spectral…
Saijo retweeted
Kohei Saijo, Yoshiaki Bando, "Input-Adaptive Spectral Feature Compression by Sequence Modeling for Source Separation," arxiv.org/abs/2602.08671
ICASSP 2025で発表した Task-aware Unified Source Separation のコードが出ました
プロンプトで音源の種類を指定して、プロンプトの数だけ音源分離/抽出できます
色々な発展性があると思うので、ぜひ試してみてください!
最近出た FlexIO の基盤にもなってます
arxiv.org/abs/2510.21485
We published the code of “Task-aware Unified Source Separation (TUSS)”, presented at ICASSP 2025!
TUSS can separate arbitrary number of sounds just by specifying the desired source type(s).
Code and Pre-trained Models: github.com/merlresearch/unif…
Paper: arxiv.org/abs/2410.23987
5
1,053
We published the code of “Task-aware Unified Source Separation (TUSS)”, presented at ICASSP 2025!
TUSS can separate arbitrary number of sounds just by specifying the desired source type(s).
Code and Pre-trained Models: github.com/merlresearch/unif…
Paper: arxiv.org/abs/2410.23987
TF-Locoformerのリポジトリに、EUSIPCO 2025で発表したNoPE (No Positional Encoding) バージョン追加しました
少し軽量になって、推論時の入力長に対する外挿性能が改善しました
もし性能いい音声強調/音源分離モデル探していたら見てみてください🙌
Happy to see the TF-Locoformer repo got 100 stars!
We added the NoPE variant. It performs better while being more efficient.
Try it if you’re looking for a high-performing speech enhancement/separation model!
Repo: github.com/merlresearch/tf-l…
Paper: arxiv.org/abs/2504.19605
Happy to see the TF-Locoformer repo got 100 stars!
We added the NoPE variant. It performs better while being more efficient.
Try it if you’re looking for a high-performing speech enhancement/separation model!
Repo: github.com/merlresearch/tf-l…
Paper: arxiv.org/abs/2504.19605
Saijo retweeted
Our ICASSP submission entitled "FlexIO: Flexible Single- and Multi-Channel Speech Separation and Enhancement" is now on arXiv! FlexIO can handle arbitrary numbers of microphones (inputs) and speakers (outputs) within a single model.
arXiv: arxiv.org/abs/2510.21485
Saijo retweeted
Yoshiki Masuyama, Kohei Saijo, Francesco Paissan, Jiangyu Han, Marc Delcroix, Ryo Aihara, Fran\c{c}ois G. Germain, Gordon Wichern, Jonathan Le Roux, "FlexIO: Flexible Single- and Multi-Channel Speech Separation and Enhancement," arxiv.org/abs/2510.21485
Saijo retweeted
🚀 Join the ICASSP 2026 URGENT Challenge! Advance Universal, Robust & Generalizable Speech Enhancement.
🗣 Track 1: Universal Speech Enhancement
🎧 Track 2: Speech Quality Assessment
🔗 urgent-challenge.github.io/u…
#ICASSP2026 #SpeechEnhancement #AI #AudioProcessing
Saijo retweeted
End of the beginning! Thank you for the EMMA team and the JSALT organizers.
The lectures and presentations on CSLP Youtube channel: youtube.com/@jhuclsp?si=oEus…
My lecture slide for speech enhancement and separation: merl.com/publications/docs/T…
少し先ですが、以下の論文を発表予定です
よろしくお願いします!
Interspeech: arxiv.org/abs/2505.23212
EUSIPCO: arxiv.org/abs/2504.19605
WASPAA: arxiv.org/abs/2505.07631
One paper got accepted to WASPAA 2025!
“Is MixIT Really Unsuitable for Correlated Sources? Exploring MixIT for Unsupervised Pre-training in Music Source Separation”
📃Paper: arxiv.org/abs/2505.07631
We will publish the code and pre-trained models soon.
See you in Tahoe!
Saijo retweeted
Our paper "Is MixIT Really Unsuitable for Correlated Sources? Exploring MixIT for Unsupervised Pre-training in Music Source Separation" written by @kohei_1979 is accepted for IEEE WASPAA 2025!
Congratulations🎉🎉🎉
I’ll be presenting a work on unified source separation model at #ICASSP tomorrow morning!
The culmination of my 10-month internship at MERL🔥
📃 Task-Aware Unified Source Separation
📅 Apr. 10, 9:42-10:00am (oral)
📍MR1.02 (Audio Source Separation III)
















