OffReco If you’re transcribing Japanese meetings, here’s how to think about the choice: for meeting-focused accuracy, Whisper medium / large-v3 (general-purpose models with more parameters that tend to stay stable in natural meetings with overlapping speech and noise), and for light, Japanese-specialized use, Kotoba Whisper v2.0 (Japanese-specialized, light and fast, and highly accurate on read-aloud audio). With OffReco, you can run either one fully on-device and switch between them instantly. This article lays out how the two differ and how to read accuracy and speed — with sources.
What’s actually different
In short, it’s the difference between a general-purpose multilingual model and a lightweight, Japanese-specialized one.
- Whisper large-v3: the top model in OpenAI’s Whisper series. A general-purpose model that covers many languages with a single network — highly accurate, but with a large parameter count, so it’s heavier to run.
- Kotoba Whisper v2.0: a Japanese-specialized model distilled from that large-v3. Per the HuggingFace model card, the student keeps large-v3’s full encoder, slims the decoder down to two layers, and is retrained on Japanese data (source: HuggingFace model card).
So Kotoba v2.0 sits where it “inherits large-v3’s Japanese ability while becoming lighter and faster.”
Accuracy, speed, and memory compared
Every figure below comes only from the HuggingFace model card.
| Aspect | Kotoba Whisper v2.0 | Whisper large-v3 |
|---|---|---|
| Type | Japanese-specialized (distilled from large-v3) | General-purpose multilingual |
| Japanese CER on read-aloud audio | Reported on par with or better than large-v3 (CER 9.2 on CommonVoice 8 Japanese) | Highly accurate general-purpose model |
| Speed | Reported ~6.3x faster than large-v3 | Larger and heavier to run |
| Inference library | faster-whisper (CPU inference) weights; no mlx build | Has an mlx build (fast on Apple Silicon GPU) |
The key point: according to the model card, Kotoba v2.0 reaches a Japanese error rate (CER) on par with or better than large-v3 while being about 6.3x faster on a read-aloud benchmark (CER 9.2 on CommonVoice 8 Japanese). It also ships weights for the high-speed faster-whisper library.
One caveat, though: that CER 9.2 and ~6.3x figure come from CommonVoice 8, a “read-aloud” benchmark — single speaker, low noise. In natural meetings with overlapping speech and background noise, general-purpose models with more parameters (medium / large-v3) can be more stable. A benchmark ranking doesn’t necessarily carry over to meetings.
A second practical difference: Kotoba only ships weights for faster-whisper (CPU); there’s no mlx build. So even on Apple Silicon it runs on the CPU rather than the GPU. By contrast, medium / large-v3 have mlx builds and run fast on the Apple Silicon GPU.
Note we deliberately don’t add any figures beyond these. Real-world accuracy and speed vary with your recording environment and hardware, so the surest test is your own meetings.
Which should you choose?
It comes down to your use case and hardware.
- You prioritize meeting accuracy (overlapping speech, noise): general-purpose models with more parameters — Whisper medium / large-v3 — tend to stay stable. On Apple Silicon, the mlx build runs fast on the GPU, so they’re practical even for meetings.
- You’re on Apple Silicon and want speed: the same mlx medium / large-v3 are easy to use. With the GPU available, even heavier models keep wait times down.
- Intel Mac / you want it as light as possible / you want to try a Japanese-specialized model: Kotoba Whisper v2.0 fits. It ships faster-whisper (CPU) weights and is accurate and light on read-aloud audio.
If you’re unsure, a practical order is to start with mlx medium / large-v3 for meetings, or with Kotoba v2.0 to try light/Japanese-specialized, then switch to the other and compare.
How this works in OffReco
What sets OffReco apart is that it runs both models fully on-device, and lets you choose.
- Both run locally: medium / large-v3 run fast on the Apple Silicon GPU (mlx), while Kotoba v2.0 runs on the CPU (faster-whisper). You can switch models in the setup screen (how to choose a model).
- Audio and transcript stay on your Mac: recording, transcription, and speaker separation all happen on the machine, so neither the audio nor the transcript leaves it. Transcription works in airplane mode (only the first-run model download needs a connection).
- Fully automatic, low barrier: it auto-detects meetings and transcribes when you stop recording. The first month is free, then ¥220/month or ¥2,200/year, on macOS 14.2 or later.
Meeting-focused: mlx medium / large-v3; light and Japanese-specialized: Kotoba v2.0. Either way you can use it without uploading to the cloud, and switch instantly to compare, so download it and check on your own meetings. Related reading: Choosing a transcription app that’s strong in Japanese (what Kotoba Whisper is) and running Whisper locally with no setup.