Source: https://fnflow.app/apple-speechanalyzer-vs-whisper

# Apple SpeechAnalyzer vs Whisper
FnFlow ships both engines, so we've seen them side by side. In our tests Whisper large-v3-turbo did better on technical terms, and SpeechAnalyzer's advantages are that it needs no model in your app and is managed by the system. We've only compared a small number of recordings, so treat this as an early read rather than a benchmark.

## The same sentence through both

A 7.6-second sentence full of work jargon:

| Engine | Result |
|---|---|
| SpeechAnalyzer (macOS 26) | Let's sync with Priya about the Kubernates migration and the Q3OKRs, then push the hot fix to staging before the demo on Friday. |
| Whisper large-v3-turbo | Let's sync with Priya about the Kubernetes migration and the Q3 OKRs, then push the hotfix to staging before the demo on Friday. |

Real output, unedited. Synthetic voice, M4 Pro, October 2026. One sentence is an example, not a benchmark, and we make FnFlow, which defaults to Whisper.

In one Russian recording we looked at, the gap was wider. SpeechAnalyzer gave “Правильно понимаю я могу поставить долларов показать приложении…”, and even the small Whisper model built into FnFlow gave a usable “Я же правильно понимаю, я могу поставить доллар за это приложение, да?” (roughly: “So can I price this app at a dollar?”). That's one sample in one language, so it can't tell you how often this happens.

## What each one is

|  | SpeechAnalyzer | Whisper (whisper.cpp) |
|---|---|---|
| Available on | macOS 26, iOS 26 and Apple's other platforms | Any Mac you can build it on; many other systems |
| Model | Downloaded and managed by the system | A file you ship or download, 32 MB to 3.1 GB |
| Language | Locale chosen by the app | Detected, or set with a flag |
| Live results | Volatile, then final | In FnFlow, final text after you release the key |
| Made for | Long-form and distant audio such as lectures and meetings, per Apple | General transcription |

From Apple's WWDC25 session 277 and the whisper.cpp README, checked 6 November 2026.

## Choosing as an app developer

SpeechAnalyzer is the easy path if you target macOS 26 and later. There's no model to bundle, and live, volatile results come with the API. Apple notes that an app can only support a limited number of languages at a time. Whisper is the option if you need older systems, control over the model or your own vocabulary hints. [Running Whisper locally](run-whisper-locally-mac) has the setup.

## Choosing as a user

Try your own speech. Open FnFlow's Settings → Recognition on macOS 26, dictate as usual, and re-run a few takes from history with the other engine. Pay attention to the names you say every day. That will tell you more than our sentence can.

## Questions

**Is SpeechAnalyzer on-device?**
Yes. Apple says transcription is entirely on device, though the models have to be downloaded first. The API manages that through AssetInventory, so the model doesn't add to your app's size.

**Which macOS does it need?**
macOS 26, alongside iOS 26 and Apple's other platforms. That's why FnFlow also runs Whisper: on macOS 15, whisper.cpp is the only engine.

**Does SpeechAnalyzer detect the language?**
In the WWDC sample, you pass a locale when you create the transcriber. Whisper can detect the language from the audio.

**Can I compare them in FnFlow?**
Yes, on macOS 26. Every dictation is kept with its audio, so you can re-run a take with the other engine and look at the two results.
