Apple's on-device speech-to-text framework — a new proprietary transcription model, introduced at WWDC 2025, that benchmarks against OpenAI's Whisper.
Last verified · 2026-07-13 · by Moe Ameen
Apple SpeechAnalyzer is the speech-to-text framework Apple introduced at WWDC 2025 (session 277, "Bring advanced speech-to-text to your app with SpeechAnalyzer," during the conference week that began June 9, 2025). It ships in the iOS 26, iPadOS 26, macOS 26 (Tahoe), and visionOS 26 generation as the modern successor to the older `SFSpeechRecognizer`, and it runs on a new proprietary Apple transcription model. The same engine powers system features like live transcription in Notes, Voice Memos, and call transcription.
The framework is built around a small set of modules. `SpeechAnalyzer` is the engine that manages an audio session; `SpeechTranscriber` is the long-form module tuned for extended audio like lectures, meetings, and podcasts; `DictationTranscriber` handles short utterances (the closer analog to the old dictation API); and `SpeechDetector` does voice-activity detection. It is an actor-based Swift API using modern concurrency — you feed it audio and read results back as an `AsyncSequence`, where early "volatile" results arrive fast and are then replaced by stable "final" ones.
The headline is that it runs on-device. Transcription happens locally, so audio does not leave the machine, and it works offline. The language model isn't pre-installed with the OS — the first time an app uses a given language, the assets are downloaded through `AssetInventory` and cached. Language support is a growing set rather than the full Whisper range at launch; developers check `SpeechTranscriber.supportedLocales` and can preload packs for offline use.
What drew attention was speed. In independent hands-on tests, a small command-line tool built on SpeechAnalyzer transcribed a 34-minute video in about 45 seconds — roughly 2.2× faster than a local Whisper Large V3 Turbo run, with no noticeable drop in quality. On a standard word-error-rate benchmark (a subset of the earnings22 earnings-call dataset), Apple's SpeechTranscriber landed near a 14% error rate at a high real-time speed factor — between Whisper base and Whisper small on accuracy, while running well faster than the small model. The honest framing: it is an excellent, fast, private on-device transcription engine for developers building Apple apps. It is not a content tool — it produces a transcript and stops.
SpeechAnalyzer answers one question extremely well — "what did I say?" — and it hands you a transcript. But a transcript isn't content; nobody watches a wall of text. That gap is exactly where Kompozy takes over. Record your talk, interview, or podcast, let an app built on SpeechAnalyzer transcribe it on-device, then bring the recording and its transcript into Kompozy. From the long recording, Kompozy's Clipped Shorts finds the moments worth cutting and produces vertical shorts with burned-in captions; from the transcript text, it drafts a Blog Article and an Email Newsletter, pulls the sharpest lines into Quote Graphics and a brand-exact Carousel, and writes native Text Posts — all governed by your Persona Brief so the whole batch reads in one consistent voice instead of raw dictation.
The multiplier is that one spoken recording becomes a full published week. Apple's framework stops at the words on the page and is Apple-device-only; Kompozy carries those words the rest of the way — captioning, reframing to 9:16, 1:1, and 16:9, and fanning a single source across nine social platforms plus a blog and email newsletter from one queue, with Autopilot and a per-post review pipeline. It even generates the formats a transcription engine can't touch at all: Persona Shorts and HeyGen avatar video with a face-locked recurring identity, so the same idea can ship as a talking-head clip too. Use SpeechAnalyzer to capture the words for free, privately, and fast; use Kompozy to turn those words into finished, scheduled, on-brand posts everywhere.
SpeechAnalyzer is Apple's on-device speech-to-text framework, introduced at WWDC 2025 and shipping in iOS 26, iPadOS 26, macOS 26 (Tahoe), and visionOS 26. It runs a new proprietary Apple transcription model, replaces the older SFSpeechRecognizer for new apps, and includes modules like SpeechTranscriber for long-form audio.
On speed, it's very competitive: hands-on tests measured a SpeechAnalyzer-based tool transcribing a 34-minute video about 2.2× faster than a local Whisper Large V3 Turbo run with no noticeable quality drop. On accuracy, one earnings-call benchmark put Apple's model between Whisper base and Whisper small, while running faster. Its main trade-off is fewer launch languages and Apple-only availability.
Yes. Transcription happens locally, so audio stays on the device and works without a network connection. The language model is not pre-installed — an app downloads and caches the needed language assets through AssetInventory the first time it uses them.
No. It produces a transcript or live captions and nothing more — it writes no posts, makes no video or images, and publishes nowhere. To turn a transcript into clips, carousels, a blog, a newsletter, and scheduled posts across platforms, you use a content engine like Kompozy.
Transcribe your recording on-device, then bring the recording and transcript into Kompozy. Kompozy cuts captioned Clipped Shorts from the audio, drafts a blog and newsletter from the transcript, builds carousels and quote graphics from the best lines in your Persona Brief voice, and schedules and publishes them across nine platforms plus blog and email.