SpeechAnalyzer

SpeechAnalyzer is a tool for programmers that turns spoken language into written text. Apple released it in 2025 alongside the new versions of its operating systems for iPhone, iPad, and Mac. Anyone developing an app can use it to have a microphone or an audio file analyzed without having to build their own speech recognition. The name combines the English word for speech with the word for analyzer. One thing above all is important: the analysis happens on the device itself, not on a remote computer on the internet. The audio recording therefore never even leaves the phone or laptop.

Why Apple is replacing the old speech recognition

Apple’s previous system for speech recognition dated back to 2016 and was called SFSpeechRecognizer. It was designed for short commands, such as dictating a single message. With long recordings it became unreliable, and it often sent the audio track to Apple’s servers for evaluation. It is precisely these two weaknesses that SpeechAnalyzer addresses.

The difference becomes clear with recordings lasting an hour. A meeting, a lecture, or a podcast can be processed in one go. Developers measured early on that SpeechAnalyzer converts a multi-hour recording into text noticeably faster than comparable open systems running on the same device. For apps that generate notes from conversations, this is the decisive point.

Then there is data protection. If a doctor’s office, a lawyer, or a journalist records conversations, uploading them to foreign servers is sensitive. On-device processing avoids this problem. In addition, there are no ongoing costs per minute of audio, as is common with online services.

From microphone to finished sentence

SpeechAnalyzer itself does not recognize speech. It is more of a framework into which individual building blocks are plugged. The most important of these building blocks is called SpeechTranscriber and performs the actual conversion of audio into text. Other building blocks can analyze the same audio stream in parallel for other characteristics.

The process works like an assembly line. The app feeds audio in in small chunks, and text fragments continuously come out the other end. These text fragments come in two kinds. The first is a quick, preliminary suggestion that can appear on the screen immediately. The second is the final version that the model produces once it has heard enough context. That’s why in such apps you sometimes see a word that was just written change again.

The device downloads the language models needed for this once, as required. They then reside in the system and are shared by all apps, rather than being loaded separately in each app. A common misconception is that SpeechAnalyzer automatically recognizes who is speaking. It does not include such speaker separation. However, it does provide timestamps for individual words, which is useful for subtitles.

Apps that already use the technology

Apple uses SpeechAnalyzer in its own programs. The Notes app can transcribe voice recordings, and the Phone app displays a text version during calls. The automatic captions in FaceTime are also based on it.

More interesting is the effect on small providers. A notes app from a two-person team has been able to offer transcription since 2025 without renting expensive servers. In tech news, the term therefore often appears in connection with the competition among dictation apps. Compared to Whisper, a freely available speech recognition model from OpenAI, SpeechAnalyzer has the advantage of tight integration with the operating system. The disadvantage is obvious: it only runs on Apple devices.