課題動画の文字起こしが面倒だったので、AIと文字起こしアプリを作った話 A developer built a local transcription app using AI tools to automate the tedious manual …

A developer has documented how they built a custom transcription application to automate what had become a repetitive chore: manually transcribing assigned lecture videos for coursework. The write-up matters because it illustrates a growing pattern in which individuals assemble local AI tools to solve narrow, personal productivity problems rather than reaching for a paid cloud service. It also shows how far the barrier to entry has fallen for tasks that once required specialized infrastructure.

The core problem is familiar to many students. Watching a video and typing out what is said is slow, error-prone, and tedious, especially when a course assigns lengthy recordings. Automatic speech recognition offers an obvious remedy, but off-the-shelf services can raise concerns around cost, upload limits, and the privacy of academic material. Running the transcription pipeline locally sidesteps several of those issues, keeping the audio on the developer's own machine and avoiding per-minute fees.

At the technical center of most projects like this is Whisper, the open speech recognition model released by OpenAI. Whisper is trained on a large volume of multilingual audio and can transcribe speech, and in some configurations translate it, across many languages. Because the model weights are available to download, developers can run it offline. A common practical choice is one of the optimized reimplementations, such as whisper.cpp or faster-whisper, which reduce memory use and speed up inference so that transcription is feasible on consumer hardware, including machines without a high-end GPU. Model size is a trade-off the builder typically has to weigh: larger variants are more accurate but slower, while smaller ones run quickly at some cost to precision.

The tag list for this piece points to a workflow that combines speech recognition with a local large language model. That combination is where the "AI app" framing becomes meaningful. Whisper produces a raw transcript, but raw transcripts are often imperfect. They can contain run-on text without punctuation, misheard technical terms, filler words, and no paragraph structure. Feeding that output into a locally hosted LLM, run through a tool such as Ollama or a similar runtime, allows a second pass that cleans up formatting, fixes obvious errors, inserts punctuation, and can even summarize the material. Handling both stages on the same device keeps the entire pipeline offline and under the user's control.

The account appears to emphasize assembly over invention. Rather than training a model, the developer combined existing components into a single application tailored to a specific need. This reflects a broader industry shift: capable open models and lightweight runtimes have made it practical for hobbyists and students to run inference on personal computers. Projects like Ollama, LM Studio, and llama.cpp have lowered the friction of downloading, quantizing, and serving models locally, while Whisper and its derivatives did the same for speech. Quantization, which reduces the numerical precision of model weights to shrink their footprint, is a key enabling technique that lets these models fit within the memory of ordinary laptops.

There are meaningful limitations to keep in mind. Automatic transcription accuracy varies with audio quality, background noise, accents, and domain-specific vocabulary, so a human review step is usually still advisable, particularly for academic work where precision matters. Local models generally lag the largest cloud-hosted systems in raw capability, and running them can be demanding on a laptop's processor, memory, and battery. Long recordings may take considerable time to process depending on the chosen model size and hardware. Users should also confirm that automating coursework transcription is consistent with their institution's policies before relying on it.

For readers considering a similar build, the takeaway is that the individual pieces are increasingly accessible, and the main work lies in wiring them together into something convenient to use. A typical version of this pipeline extracts audio from a video, often with a utility like FFmpeg, transcribes it with a Whisper variant, refines the text with a local LLM, and outputs a formatted document. The result is a small, self-contained tool that addresses one person's recurring frustration. More broadly, this kind of project signals how local AI is moving from experimentation toward everyday utility, giving users a private, low-cost alternative to subscription services for tasks that were impractical to automate only a few years ago.