XYZ

Nobody

[XYZ @XyzVision]

@denkiwakame 603 pages...😂

[XYZ @XyzVision]

@Daichi_Ishii 感觉可以上上看

[XYZ @XyzVision]

@denkiwakame already broken down

[XYZ @XyzVision]

@denkiwakame Let's do it next year.

[XYZ @XyzVision]

@Tdash_Koz 😍😍😍

[XYZ @XyzVision]

@denkiwakame It's time to prepare ソレスタルビーイング.

[Sakana AI @SakanaAILabs]

We’re excited to introduce KAME: Tandem Architecture for Enhancing Knowledge in Real-Time Speech-to-Speech Conversational AI, accepted at #ICASSP2026! 🐢

Blog https://t.co/arVz1TGpJJ
Paper https://t.co/0EwpyRXeCs

Can a speech AI think deeply without pausing to process?

In real conversation, we don’t wait until we’ve fully worked out what we want to say—we start talking, and our thoughts catch up as the sentence unfolds.

Fast speech-to-speech models achieve this, but their reasoning tends to stay shallow. Cascaded pipelines that route through a knowledgeable LLM are smarter, but the added latency breaks the flow—they fall back to "think, then speak."

In our new paper, we propose a way to break this trade-off. We call it KAME (Turtle in Japanese).

A speech-to-speech model handles the fast response loop and starts replying immediately. In parallel, a backend LLM runs asynchronously, generating response candidates that are continuously injected as "oracle" signals in real time.

This shifts the AI paradigm from "think, then speak" to "speak while thinking."

The backend LLM is completely swappable. You can plug in GPT-4.1, Claude Opus, or Gemini 2.5 Flash depending on the task without changing the frontend. In our experiments, Claude tended to score higher on reasoning, while GPT did better on humanities questions.

Try the model yourself here: https://t.co/uDA0nvvjhS

[XYZ @XyzVision]

Also, there are resources in Japan.

[Akari Asai @AkariAsai]

Not many PhD students know about compute grants, but they can make a huge difference. During my PhD, I got access to Stability AI's HPC cluster through a small proposal and used it for Self-RAG training.
Great practical post by @_emliu!

[XYZ @XyzVision]

@wataru_hirota does this project welcome external cooperation?

[XYZ @XyzVision]

@shyyhs I think it’s also okay to just use a few of the budget since the time is short.

[XYZ @XyzVision]

@shyyhs robot arm++

[XYZ @XyzVision]

@lyu_boxuan speech有中文 文档有中文 唯独界面没有

[XYZ @XyzVision]

@wataru_hirota Next step, distill experts skills.

[XYZ @XyzVision]

@tatsukamijo @sherryyangML Congratulations!!!

[XYZ @XyzVision]

@caesar_wanya HCI + Speech，最高！

[XYZ @XyzVision]

@denkiwakame 🥰

[XYZ @XyzVision]

@caesar_wanya At the same time, the acc rate of last year’s winter exam for KU’s IST course is almost 2 or 3 out of 80. Then they canceled written exam.

[马东锡 NLP @dongxi_nlp]

非常有趣的工作！

如果 LLMs 必须按照 token-by-token 做 CoT 生成，那么每一步的生成，是否可以承载超过一个 token 的信息量？

Think wider. Think Shorter！