📖 英文と日本語訳（一文ずつ）

🎧 通し読み（全文）

リスニング・シャドーイング用の全文です。

Sakana AI has announced KAME (Tandem Architecture for Enhancing Knowledge in Real-Time Speech-to-Speech Conversational AI), a new architecture for real-time speech conversational AI. This study has been accepted to ICASSP 2026, and the paper, inference code, fine-tuning code, and models have been made publicly available. Conventionally, there are two primary methods for achieving voice interaction with AI. One is the Speech-to-Speech (S2S) method, which generates speech directly from audio; while it offers fast response times, it tends to be inferior in terms of knowledge. The other is a cascaded approach that connects speech-to-text (STT), large language models (LLM), and text-to-speech (TTS) in series; while superior in terms of knowledge, it faced challenges with response latency. KAME employs a mechanism that links an S2S model with a backend large language model to combine the strengths of both. Specifically, the S2S model manages a high-speed response loop, while the backend LLM operates asynchronously on a slower cycle, injecting response guidelines known as "oracle signals" into the S2S side as they become available. As the user’s speech progresses, the backend LLM is repeatedly invoked for the growing partial transcript, incrementally refining the response signal. This is said to realize a new paradigm of "thinking while speaking" rather than "thinking before speaking." The research team evaluated KAME by comparing it with Moshi, an existing full-duplex S2S model. To ensure fairness, both systems started from the same Moshi base model and were trained on the same data. While Moshi is trained as a standard SFT model, KAME’s Moshi-based frontend is trained to receive oracle signals from the backend model. The evaluation adopted a methodology in which MT-Bench text prompts were converted into audio queries, after which the resulting audio responses were transcribed and scored by an LLM judge. KAME, utilizing gpt-4.1 as its backend, outperformed Moshi in the reasoning problems presented in the demo, as well as in STEM and humanities subjects. Furthermore, KAME features the easy replacement of backend LLMs, allowing users to select models such as GPT-4.1, Claude-Opus-4-1, or Gemini-2.5-Flash depending on the application. According to the examples in the article, Claude-opus-4-1 tends to achieve high scores in reasoning tasks, while GPT-4.1 is reported to show a tendency for high scores in the humanities. It is concluded that the tandem structure allows for flexible backend switching, enabling the strengths of individual LLMs to be leveraged without altering the dialogue framework.

📝 学習のヒント

1. 1まず英文を読む — 知らない単語にあたりをつけてから音声へ。
2. 2一文ずつ確認 — 日本語訳と照合し、構文を理解する。
3. 3通し読み Normal で耳を作る — 内容を追いながらリピート。
4. 4Fast でシャドーイング — 口を慣らし、リスニング速度を上げる。
5. 5翌日に復習 — 1日空けて再聴すると長期記憶に定着しやすい。