Mobile real-time translation: providers, debounce, and cache

Clients often come with a task of real-time translation in a mobile app — for chat, text input, or documents. Each scenario requires its own approach: 200–400 ms latency for chat, debounce for live input, batch with progress bar for documents. Over many years of experience, we have implemented over 30 projects with translation for medical tourism, e-commerce, and messengers. We guarantee stable performance and compliance with App Store Review Guidelines. We reduce integration costs through thoughtful architecture. In this article, we break down the stack, architecture, and practical techniques.

How to Choose a Translation Provider?

We are certified Google Cloud developers and have experience integrating all three major providers. Comparison table:

Google Cloud Translation is the standard for production. Supports formal/informal registers, REST and gRPC. DeepL is the choice when translation quality of complex texts (German, French) is critical. ML Kit is the foundation for offline scenarios: model ~15 MB per language pair, downloaded once.

Why Debounce Matters?

The main mistake of live translation is sending a request on every keystroke. At a typing speed of 200 characters per minute, that's 3–4 requests per second, of which 90% will be wasted before getting a response. The correct pattern is debounce with cancellation of previous pending requests.

@Published var inputText: String = ""

inputText
    .publisher
    .debounce(for: .milliseconds(500), scheduler: RunLoop.main)
    .removeDuplicates()
    .filter { $0.count >= 3 }
    .flatMap(maxPublishers: .max(1)) { [weak self] text -> AnyPublisher<String, Never> in
        guard let self else { return Empty().eraseToAnyPublisher() }
        return self.translationService.translate(text)
            .replaceError(with: "")
            .eraseToAnyPublisher()
    }
    .receive(on: RunLoop.main)
    .assign(to: &$translatedText)

flatMap(maxPublishers: .max(1)) — this is switchMap: it cancels the previous request on new input. Without this, old responses can overwrite the current translation.

On Android (Kotlin Flow):

val translatedText: StateFlow<String> = inputText
    .debounce(500)
    .filter { it.length >= 3 }
    .distinctUntilChanged()
    .flatMapLatest { text ->
        flow { emit(translationRepo.translate(text)) }
            .catch { emit("") }
    }
    .stateIn(viewModelScope, SharingStarted.Lazily, "")

flatMapLatest — equivalent of switchMap, cancels the previous coroutine.

Without cancellation, during fast typing the pipeline gets overwhelmed: each new character triggers a translation, but responses arrive in arbitrary order. The user sees translation "flickering". Using flatMapLatest guarantees that on new input the previous translation is cancelled, and only the latest result reaches the UI.

Integration with Google Cloud Translation v3

suspend fun translate(text: String, targetLang: String = "ru"): String {
    val body = JSONObject().apply {
        put("q", text)
        put("target", targetLang)
        put("format", "text")
    }
    val response = httpClient.post("https://translation.googleapis.com/language/translate/v2") {
        header("Authorization", "Bearer $accessToken")
        contentType(ContentType.Application.Json)
        setBody(body.toString())
    }
    return response.body<TranslationResponse>().data.translations[0].translatedText
}

For access_token in production — GCP service account, JWT signing on the backend. The mobile client receives a short-lived token via its own /api/translate-token endpoint. No GCP API key in APK/IPA.

ML Kit for Offline Scenarios

// iOS: Google ML Kit Translate
let options = TranslatorOptions(sourceLanguage: .english, targetLanguage: .russian)
let translator = Translator.translator(options: options)

translator.downloadModelIfNeeded { error in
    guard error == nil else { return }
    translator.translate("Hello world") { result, error in
        print(result ?? "")
    }
}

The model is downloaded once over Wi-Fi. After that, it works offline. Device latency is 20–50 ms per phrase. Ideal for messengers, offline video subtitles, and travel apps.

Translation Caching

Repeated requests for the same text waste money. Cache at the SQLite level (Room/CoreData) with key sha256(source_text + target_lang). TTL 7 days for regular content, no TTL for static UI strings. At the HTTP level, Cache-Control for GET requests. According to Google Cloud Translation documentation, GET requests with parameters are supported, allowing caching at the URLCache / OkHttp Cache level. This significantly reduces API costs — caching improves efficiency by up to 10x over uncached translation.

Case Study

In a project for our client, a medical tourism app (iOS + Android), we translated clinic descriptions and patient reviews (en→ru, ru→en, de→ru). Cloud translation for content on load, ML Kit offline for the consultation interface. Debounce of 700 ms for the search bar. Translation cache in Room reduced API requests by 73% within the first week after launch, and translation display speed increased by 40%. The API budget savings were significant — over $500 per month for our client.

Process

1. Analytics: study usage scenarios (chat, live, batch), latency and language requirements.
2. Design: select provider and architecture (cloud/offline/hybrid), design debounce and cache.
3. Implementation: integration via backend, writing translation module, setting up request cancellation.
4. Testing: check latencies under different loads, translation quality, offline behavior.
5. Deployment: publish to App Store / Google Play, monitor errors via Crashlytics.

What Is Included

• Source code for iOS and/or Android with integration of the chosen provider.
• Documentation on architecture, access keys, and operation.
• Backend token setup and API access.
• Training your team on using the solution.
• Warranty support for 1 month after delivery.

Timelines and Cost

Basic integration of one provider with debounce and cache takes 5 to 7 days and costs starting from $1,500. Adding ML Kit offline, language auto-detection, and formatted text takes 4 to 6 days. Exact timelines and cost are calculated individually, considering your budget. Get a consultation on provider selection and translation architecture. Contact us to evaluate your project.

Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models

We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.

How to Choose Between CoreML and TFLite for On-Device Inference?

CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.

TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.

For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.

What If You Need Text Generation On-Device?

Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.

llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.

On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.

Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.

How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?

On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.

For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.

Integrating OpenAI API and Other Cloud Models

For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.

Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.

What Is Included in the Work (Deliverables)

• Trained and quantized model for the target device (documentation with metrics)
• SDK for integration (Swift/Kotlin/Flutter) with call examples
• Performance tests on 3–5 real devices
• Instructions for OTA model updates
• Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
• 2 weeks of technical support after release

Typical Project Pipeline

1. Task analysis — measure latency, privacy, size, supported devices.
2. Model prototyping — in Python, evaluate accuracy on target data.
3. Conversion and quantization — for CoreML/TFLite with validation.
4. Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
5. Testing — on real devices, measure FPS, RAM, battery.
6. Deployment — via TestFlight / Firebase App Distribution, monitor metrics.

Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.

Why We Take on Complex Cases?

10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.

Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.