黒電話を分解して、ローカルLLM×ずんだもんと通話できるマルチモーダルAIシステムを作ってみた➁This follow-up article details the construction of a multimodal AI system that…

A do-it-yourself project published on Qiita continues its walkthrough of converting a vintage Japanese black rotary telephone into the front end for a conversational AI system, pairing a locally hosted large language model with the Zundamon synthetic voice. The topic is worth attention because it shows, step by step, how inexpensive edge hardware and freely available speech and language tools can be combined into an interactive exhibit — here aimed at a school or cultural festival (gakuensai) display with a deliberate entertainment slant.

At the heart of the build is a disassembled 600-A2-CL, one of the standard black telephones once common in Japanese households and offices. Instead of preserving its telephone function, the maker repurposes the handset, dial and internal contacts as a physical interface for a modern software stack. Picking up the receiver and speaking into it become the natural input gestures, and the system replies through synthesized speech, so a familiar analog object turns into the gateway for a chatbot-style interaction.

The "multimodal" description refers to the way the project blends several input and output channels rather than relying on text alone. Voice captured through the handset is processed and passed to the language model, and the model's text response is rendered back as audio. Zundamon, the voice used for those replies, is a well-known character associated with the VOICEVOX text-to-speech ecosystem, which is widely used in Japanese hobbyist and content-creation communities because it is free for many uses and produces recognizable, expressive speech.

Running the language model locally is a defining choice. Rather than sending prompts to a cloud API, the system appears to host the model on the device itself, which the project's tags indicate is built around a Raspberry Pi. Local inference carries trade-offs relevant to any similar build: it avoids per-request cloud costs and keeps data on-device, and it can work without a reliable network connection — a practical advantage in a crowded festival hall — but it is constrained by the limited memory and compute of small single-board computers. Projects of this kind typically rely on quantized, smaller-parameter models and optimized runtimes to fit within those limits, sometimes at the expense of response speed or answer quality.

A working system of this type generally chains together three stages: speech recognition to transcribe what the caller says, the language model to generate a reply, and text-to-speech to voice it. Each stage adds latency, so tuning the pipeline for a responsive back-and-forth is often as important as the model choice itself. As a follow-up to the first article, this installment is likely to focus on the later or more refined stages of that integration — wiring the reused telephone hardware to the compute board, handling audio in and out, and connecting the software components — though readers seeking the full background may need the earlier post for the initial design and parts list.

The project sits within a broader wave of interest in on-device AI. Tools such as Ollama, llama.cpp and similar runtimes have made it far easier for individuals to run open-weight models on modest hardware, while the Raspberry Pi remains a popular platform for physical-computing and hardware-hacking experiments. Combining these with Japanese-language voice synthesis reflects a growing pattern of makers building playful, self-contained AI appliances instead of depending solely on hosted services.

For a festival exhibit, the appeal is as much about presentation as engineering. A decades-old rotary phone that answers back in a familiar character voice is an immediately legible demonstration of how modern AI can be embedded in unexpected obj