SpeechAnalyzer, the speech-to-text framework Apple introduced at WWDC 2025, transcribes on-device with a new proprietary model. In independent hands-on tests it matched Whisper's quality while running roughly twice as fast — making high-quality transcription a free, private, built-in baseline.
2026-07-13 · by Moe Ameen
At WWDC 2025 — the developer conference that began June 9, 2025 — Apple introduced SpeechAnalyzer, a new speech-to-text framework that ships in the iOS 26, iPadOS 26, macOS 26 (Tahoe), and visionOS 26 generation. It runs a new proprietary Apple transcription model entirely on-device, replaces the older `SFSpeechRecognizer` for new apps, and already powers system features like live transcription in Notes, Voice Memos, and call transcription. Apple detailed it in session 277, "Bring advanced speech-to-text to your app with SpeechAnalyzer."
What turned a developer session into a story was independent benchmarking against OpenAI's Whisper, the open-source model most transcription tools have been built on. In a widely shared hands-on test, 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 difference in quality. A separate word-error-rate benchmark on earnings-call audio placed Apple's SpeechTranscriber between Whisper base and Whisper small on accuracy while running at a much higher real-time speed factor than the small model.
The framework is modular and modern: a `SpeechTranscriber` module tuned for long-form audio, an actor-based Swift API that streams results as an `AsyncSequence` (fast "volatile" results replaced by stable "final" ones), and on-device processing so audio never leaves the machine and works offline. The language model isn't bundled with the OS — apps download and cache the packs they need through `AssetInventory` on first use — and language coverage is a growing set rather than Whisper's full range at launch. It is Apple-hardware-only. The takeaway most creators care about isn't the API surface: high-quality transcription is now fast, free, private, and built into the device.
When transcription is free and instant, the value quietly shifts from "getting the words" to "what you do with them" — and that's the half Kompozy owns. A transcript is a source, not a post; nobody watches a wall of text. So take the recording SpeechAnalyzer just transcribed on your Mac or iPhone and drop it into Kompozy. From the long audio, Clipped Shorts cuts the moments worth posting and burns in captions; from the transcript, Kompozy 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 held to one voice by your Persona Brief so a week of output reads as your brand, not raw dictation. Auto-captions themselves are already commodity (Kompozy burns them into every short); the durable value is the generation and the distribution around them.
That distribution is the point Apple's framework never reaches. SpeechAnalyzer is on-device and Apple-only and ends at the words; Kompozy carries them across nine social platforms plus blog and email from one queue, reframing to 9:16, 1:1, and 16:9 and fanning a single recording into 25–35 outputs, with Autopilot and a per-post review pipeline. It also generates what a transcription engine can't — Persona Shorts and HeyGen avatar video with a face-locked recurring identity. There's a same-week story to publish here too: "Apple's speech API now benchmarks against Whisper" is a query your audience is searching, and Kompozy turns your take on it into a captioned short, a carousel, a blog explainer, and platform-native posts in an afternoon.
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, and powers features like live transcription in Notes and Voice Memos.
In independent tests it ran faster: a SpeechAnalyzer-based tool transcribed a 34-minute video about 2.2× faster than a local Whisper Large V3 Turbo run, with comparable quality. On an earnings-call accuracy benchmark it sat between Whisper base and small. The trade-offs are Apple-only availability and a smaller launch language set.
It makes accurate transcripts and captions instant, private, and effectively free — so "getting the words out of a recording" stops being the bottleneck. The scarce work moves downstream: turning those words into clips, posts, a blog, and a newsletter, in your voice, published everywhere. That is what a content engine like Kompozy does.