A sample app that implements an audio recording flow in SwiftUI: start, pause, resume, or discard a recording while a timer tracks the elapsed time, then send the recording to be transcribed on device — and read the resulting text in a sheet.

A screen recording of the full flow is included in the repo: Attendi_Sample_App.MP4.

Context

This sample app was created for a technical interview for a senior Mobile (iOS + Android) engineer role at Attendi, a Dutch health-tech company that makes voice-driven reporting software for healthcare professionals, based on a given set of requirements.

Requirements

Requirement	Version
Xcode	26 or later
Swift	6.3 or later
Platforms	iOS, macOS, and visionOS 26 or later

Running the app

Open Attendi.xcodeproj in Xcode, pick the Attendi scheme, and run it on an iOS device or simulator, a Mac, or an Apple Vision Pro — the same code base covers all three.

Two things to expect on a first run: the app asks for the microphone permission when the first recording starts, and the first transcription in a language needs a network connection — picking a locale from the toolbar menu downloads its speech model in the background, reported through the in-app notifications, and a recording sent before the download finishes simply waits for it. Every transcription after that happens entirely on device.

Project structure

The project separates the app shell from the feature code, which lives in a local Swift package:

Attendi/
├── Attendi.xcodeproj
├── Apps/
│   └── Attendi/                         # The app target (thin shell)
│       ├── Sources/
│       │   ├── App/                     # The @main entry point
│       │   ├── View Models/             # The attached services, the supported locales, and the download-to-notification mapping
│       │   └── Views/                   # The root view: hosts the feature, the locale picker, the notifications, and the transcription sheet
│       ├── Previews/                    # The PreviewPreinstalling service backing the previews
│       ├── Tests/                       # The AttendiTests bundle: the content view model suite, with its mock
│       ├── Catalogs/                    # The app's assets and localized strings
│       └── Resources/                   # The Info.plist, including the microphone usage description
└── Packages/
    └── Features/                        # Local Swift package with the feature code
        ├── Package.swift
        ├── Sources/
        │   ├── Notifying/               # The Notifying feature target
        │   │   ├── Models/              # The AppNotification model
        │   │   ├── Styles/              # The NotificationLabelStyle label style
        │   │   └── View Models/         # The Notifier view model
        │   └── Recording/               # The Recording feature target
        │       ├── Catalogs/            # The record, pause, and send icons; the localized strings
        │       ├── Errors/              # The AudioCapturingError, AudioTranscribingError, RecordingError errors
        │       ├── Models/              # The Transcription model
        │       ├── Protocols/           # The Capturing, Preinstalling, Transcribing protocols
        │       ├── Services/            # The AudioCapturing, AudioTranscribing, and AssetPreinstalling services, with its respective development dummies
        │       ├── Styles/              # The RecordingButtonStyle button style
        │       ├── View Models/         # The RecordingViewModel view model
        │       └── Views/               # The RecordingView view
        └── Tests/
            ├── Notifying/               # The notifier suite
            ├── Recording/               # The model, view model, and service suites, with their mocks
            └── Features.xctestplan

The app target only depends on the Features package and renders its public RecordingView view; all recording logic and UI live in the package's Recording target.

Architecture

The project follows an MVVM architecture on SwiftUI, split across a thin app target and a local Swift package:

Views render state and forward user events; they own no logic beyond layout, styling, and presentation, and decompose into private subviews and reusable styles.
View models (@Observable, @MainActor classes) own the state and every side effect of its transitions — the recording state machine and the notification lifecycle in the package; the locale and the mapping of download events to notifications in the app.
Services sit behind protocols (Capturing, Transcribing, Preinstalling) injected at initialization, so the same logic runs against the real audio backends, the preview dummies, or the test mocks.
Models are plain value types (Transcription, AppNotification) with injectable identifiers, so tests can construct comparable values.

The package boundary enforces the separation: the app depends on the Features library and composes its public surface, while everything else in the Recording target stays internal.

Platforms

A single multiplatform code base targets iOS, macOS, and visionOS 26. Platform differences are isolated behind conditional compilation: the shared AVAudioSession — its configuration, deactivation, and interruption observation — only exists on iOS and visionOS, and the navigation bar title display mode is skipped on macOS. Everything else, including the Liquid Glass styling, is shared.

Techniques

Swift 6 strict concurrency — async/await throughout, @MainActor isolation for the UI-facing types, Sendable protocols and models, AsyncStream for service events, and withTaskCancellationHandler for cancellable downloads. The Combine framework is not used.
Observation — view models are @Observable, so views react to exactly the properties they read.
On-device speech — the SpeechAnalyzer/SpeechTranscriber stack with AssetInventory asset management; no audio ever leaves the device.
Liquid Glass design — custom ButtonStyle and LabelStyle types over glassEffect backgrounds, scaling with Dynamic Type via @ScaledMetric.
Localization and accessibility — string catalogs with generated string symbols on both targets, localized accessibility labels on every icon-only control, timer values spelled out in full units, and the transient notifications announced to assistive technologies when posted.
Logging — OSLog subsystems for the app and the package, so failures that are not surfaced remain diagnosable.

Strategies

Model-owned side effects — a button press and a unit test drive the exact same transition code, keeping the views pure renderers and the state machine fully testable.
Serialized service access — a task chain guarantees the capture calls reach the service in transition order, no matter how quickly the user taps.
Drift-free timing — the recording time derives from a monotonic clock anchor instead of counted ticks, so scheduling latency never accumulates.
File-based audio hand-off — the capture and the transcription exchange a file URL, never an in-memory copy of the audio.
Preflighted downloads — speech model assets install when a locale is picked, as a cancellable task reported through notifications, instead of mid-transcription; locale reservations are recycled to respect the system's per-app quota.
Typed error surfacing — service failures map to a small user-facing error enum with tailored recovery, such as opening the app's settings for a denied permission, while the underlying errors go to the log.
Behavior-driven tests — Swift Testing suites drive the model exclusively through its public API, with parameterized cases across every state and mocks that record their invocations in call order.

The Recording feature

State machine

The flow is modeled as a state machine in RecordingView.Model, an @Observable, @MainActor view model:

State	Meaning
`notRecording`	Idle; no recording in progress.
`recording`	A recording is in progress and the timer is ticking.
`paused`	The recording is paused; it can be resumed, discarded, or sent.
`processing`	The sent recording is being stopped and transcribed.

The main button starts, pauses, and resumes a recording. While paused, two more buttons appear: discard throws the recording away, and send moves the flow into processing — stopping the capture and transcribing the audio — before returning to idle. The discard button stays available while processing, where it cancels the in-flight transcription and throws the recording away without publishing anything. The model owns every side effect of a transition (timer, capture, processing), so the state machine works the same whether it is driven by the view's buttons or by the unit tests.

The calls to the capturing service are serialized through a task chain: no matter how quickly the states change, start, pause, resume, and stop always reach the service in order. When the system interrupts the audio session mid-recording — a phone call or Siri, for example — the service emits an event and the model pauses exactly like a press of the main button would; once the interruption ends, the recording resumes by itself — but only when the system hints the capture may continue, and never a recording the user paused themselves.

Timer

The recording time is measured with a monotonic ContinuousClock: an anchor marks the start of the current recording stretch, and pausing folds the stretch into an accumulated duration, so scheduling latency never accumulates as drift. While recording, an async task republishes the measured time once per second. The view formats the count as mm:ss and animates digit changes with a numeric text content transition.

Errors

Failures of either service are published by the model as a RecordingError and presented by the view in an alert: a denied microphone permission additionally offers to open the app's settings, and unavailable speech model assets are distinguished from an ordinary transcription failure. The underlying errors are logged through OSLog under the Features.Recording subsystem.

Views

RecordingView — the public entry point of the feature. It renders the timer label and the control buttons for the model's current state, forwards every transcription to a closure given at initialization, and presents the error alert. All controls carry localized accessibility labels matching their current action, and the timer exposes its value to assistive technologies spelled out in full units.
RecordingButtonStyle — a custom ButtonStyle used by the controls: the label sits on a padded, circular red Liquid Glass background that reacts fluidly to presses, and dims while disabled. The label and its padding scale with Dynamic Type via @ScaledMetric, and an invertStyle flag controls whether the label's color scheme is inverted for contrast.

Services

The actual recording work is abstracted behind two protocols, injected into RecordingView at initialization and attached to its view model:

Capturing — captures the audio from a microphone, with throwing async start, pause, resume, and stop methods; stop hands the captured audio over as a file URL, avoiding any in-memory copies. An events stream (empty by default) lets a service report interruptions of the capture.
Transcribing — a callable service that transcribes a recorded audio file into text for a given locale, returning a Transcription.
Preinstalling — resolves the locales the transcriber supports and preinstalls the speech model assets of a locale ahead of its first transcription, emitting the started, cancelled, and failed downloads through an events stream instead of throwing.

The real backends used by the app are:

AudioCapturing — records single-channel AAC into a temporary file through AVAudioRecorder, requesting the microphone permission and managing the shared audio session on the platforms that require it, and emitting an event when the system interrupts the session while recording.
AudioTranscribing — transcribes entirely on device through SpeechAnalyzer with a SpeechTranscriber module, in the closest supported equivalent of the requested locale, installing any missing speech model assets through the preinstalling service before the analysis.
AssetPreinstalling — downloads speech model assets through the system's AssetInventory, recycling the locale reservations beforehand since the system only permits a limited number per app, and discriminating a cancelled download from a failed one.

The internal DummyCapturing and DummyTranscribing fake the work for previews and default model values (the latter with a two-second delay and a dummy transcription); the unit tests inject configurable mocks, and other backends can be plugged in the same way without touching the feature's state machine.

The app shell

ContentView hosts the feature inside a navigation stack and presents the transcribed text of every processed recording in a modal sheet — or a content-unavailable message when the transcription is empty. A toolbar menu picks the locale of the spoken language from the locales the Preinstalling service supports, aligned at launch with the closest supported equivalent of the user's locale. Picking a locale preinstalls its speech model assets in the background through the service, so the first transcription does not have to download them mid-processing.

The preinstallation runs as a cancellable task owned by the model: picking another locale cancels a download still in flight before starting the new one. The model maps the service's started, cancelled, and failed download events to the Notifying feature's Notifier — supplying the localized message and the symbol for each — which presents them as transient in-app notifications: Labels wearing the NotificationLabelStyle, a capsule Liquid Glass banner whose icon is tinted by the notification's kind (green for info, yellow for a warning, red for an error). The banners overlay the feature just below the navigation bar, sliding in from the leading edge and dismissing themselves after a few seconds, and every posted notification is announced to assistive technologies. A download that completes posts none.

The app itself no longer touches the Speech or Accessibility frameworks: all speech model management lives behind the package's Preinstalling service, and the notification lifecycle — posting, announcing, and auto-dismissal — behind the Notifier.

Testing

The package is covered by Swift Testing suites. In the Recording target, RecordingViewModelTests.swift groups the view model's behaviors with nested @Suite types (initial state, button presses, computed properties, timer, capturer, errors, interruptions, and processing) and TranscriptionTests.swift covers the Transcription model. In the Notifying target, NotifierTests.swift covers the posting and auto-dismissal lifecycle with an injected, shortened dismissal delay, and AppNotificationTests.swift covers the AppNotification model. The state-dependent behaviors are exercised with parameterized tests across all four states, driven through the model's public press handlers, and the service mocks record their invocations in call order. AudioCapturingTests.swift covers the one piece of pure logic in the audio services — the parsing of audio session interruption notifications into capturing events, extracted so it compiles on every platform. The rest of the system-bound service code, AssetPreinstalling included, is deliberately untested: thin wrappers over device-global frameworks like AssetInventory and SpeechAnalyzer, whose behavior mocks cannot meaningfully verify.

Run the package tests with the Features scheme in Xcode (⌘U), or from the command line. The suites run on any of the three platforms; macOS needs no simulator:

xcodebuild test -scheme Features -destination 'platform=macOS'

For iOS, point the destination at a booted simulator, for example -destination 'platform=iOS Simulator,name=iPhone 16'.

The app has its own AttendiTests unit-test bundle, hosted by the app: ContentViewModelTests.swift drives ContentView.Model with a mocked Preinstalling service, covering the locale loading and alignment, the cancellable preinstallation guard, and the mapping of the download events to the notifier's notifications. Run it with the Attendi scheme in Xcode (⌘U).

Localization

Both the app and the feature package localize their user-facing text through string catalogs (Localizable.xcstrings, currently English only), resolved through Xcode's generated string symbols (LocalizedStringResource): the app's catalog covers the navigation titles, the transcription sheet, the locale picker, and the notification messages — the latter taking the locale's name as their argument — and the package's catalog covers the error alert and the accessibility labels.

Tooling

.swift-format — configuration for Apple's swift-format, used by Xcode's built-in formatter (4-space indentation, 200-column lines).