SpeechAnalyzer

On-device speech-to-text on Apple platforms via the iOS 26 Speech framework (SpeechAnalyzer + SpeechTranscriber), the modular async/await successor to SFSpeechRecognizer built for long-form and live transcription. The deep API reference — architecture, every module, file and live pipelines, asset/locale management, result handling, SwiftUI integration, SFSpeechRecognizer migration — lives in references/guide.md. This file is the decision and discipline layer: read it first, open the guide for specifics.

This skill covers turning audio into text. It does not cover capturing that audio — the mic tap, AVAudioEngine, session category, and format wiring belong to avfoundation-audio. SpeechAnalyzer consumes AVAudioPCMBuffers; getting them is that skill's job.

Dials

Set these explicitly at the start; they change what "correct" means.

1. SOURCE — file (pre-recorded AVAudioFile; prefer the structured analyzeSequence(from:) and .offlineTranscription preset) · live (mic buffers fed through an AsyncStream<AnalyzerInput>; autonomous start(inputSequence:) + .progressiveLiveTranscription).
2. RESULTS — final-only (commit only result.isFinal; simplest, no flicker) · volatile (enable .volatileResults, show interim guesses as a replaceable preview, commit on final). Add attributeOptions: [.audioTimeRange] only if you need word/segment timing.
3. LOCALE_STRATEGY — current-only (resolve supportedLocale(equivalentTo: .current), download on demand) · reserved (pin specific languages with AssetInventory.reserve(locale:) so the system won't reclaim the model; bounded by maximumReservedLocales).

When to use

Building or reviewing any transcription, dictation, live-caption, or audio-to-text feature on an iOS 26+ target. If the target must support iOS 25 or earlier, or watchOS, SpeechAnalyzer is unavailable — fall back to SFSpeechRecognizer. If you only need to capture audio (not transcribe it), use avfoundation-audio.

Core rules