Mobile Speech Recognition: Native APIs vs Cloud Services

A patient dictates symptoms into a medical app – the text must appear instantly, free of errors in medical terminology. In practice, we face a dilemma: on-device recognition gives 200 ms latency and works offline, but the vocabulary is limited; cloud recognition is more accurate but requires internet and costs money. We have implemented both scenarios in 15 projects and know which approach fits where. Reach out for a free estimate on optimal STT architecture selection for your app.

Speech to Text Mobile App Implementation: Native APIs

On iOS we use SFSpeechRecognizer – a built-in API with on-device support from iOS 13 for 11 languages. For Russian, on-device is unavailable, so requests go to Apple's servers. Limits: 1 minute per request, about 1000 requests per day per device without a paid agreement. On Android – SpeechRecognizer, which allows offline Russian recognition by installing a language pack (80–200 MB).

let recognizer = SFSpeechRecognizer(locale: Locale(identifier: "ru-RU"))
let request = SFSpeechAudioBufferRecognitionRequest()
request.requiresOnDeviceRecognition = false // for Russian – cloud
request.shouldReportPartialResults = true

The parameter shouldReportPartialResults = true is critical for UX: the user sees text as they speak.

On-Device STT vs Cloud Speech Recognition: Which to Choose?

On-device requires no network, minimal latency, but lower accuracy for complex terms and long phrases. Cloud services (Whisper, Google Cloud STT, Yandex SpeechKit) give up to 95% accuracy on Russian, but add 1–3 s latency and cost per request. On-device is 5–10 times cheaper than cloud but 10–15% less accurate. The choice depends on the criticality of offline mode and budget. For short commands, on-device is usually sufficient; for transcribing conversations, only cloud works. Despite apparent savings, on-device recognition may require additional development and testing costs due to vocabulary and accuracy limitations. In some scenarios, such as specialized vocabulary (medical, legal, technical terms), on-device yields unacceptably high error rates. Cloud services with custom dictionaries solve this with flexibility and regular model updates.

Native APIs: Implementation Details

Streaming Recognition vs Batch: Implementation Choices

Batch (record → stop → recognize) is simpler to implement but has worse UX. Streaming – text appears as the user speaks. Streaming on iOS is implemented via SFSpeechAudioBufferRecognitionRequest with append(buffer:) from AVAudioEngine:

inputNode.installTap(onBus: 0, bufferSize: 1024, format: format) { buffer, _ in
    request.append(buffer)
}

For Yandex SpeechKit, streaming uses WebSocket with chunked audio in PCM 16 kHz 16bit, chunks of 200–400 ms.

Custom Dictionary STT: Boosting Accuracy with Phrase Hints

If native API accuracy is insufficient, add custom dictionaries. In Yandex SpeechKit it's PhraseSuggestions, in Google Cloud Speech-to-Text – SpeechAdaptation. In practice, after adding 3000 article codes in a voice input app for a warehouse client, recognition accuracy for product codes increased from 61% to 89%, saving about 15,000 rubles ($170) per month on manual corrections. In a recent project, we achieved similar gains for a medical app. On iOS, you can limit vocabulary via SFSpeechRecognitionTaskHint.

Case Study: Voice Form Filling App in a Warehouse

One of our clients, a large warehouse operator, needed hands-free data entry. We started with native Android STT with offline Russian pack. Problem: specific SKU codes, e.g., "article 7788-ABC", were poorly recognized. Solution: Yandex SpeechKit with custom dictionary of 3000 article codes. Accuracy on product codes rose from 61% to 89%. Additionally: processing cost about $0.004 per 15 seconds of audio, which at 5000 dictations per day is ~$20 per day. The solution saved the client approximately $1,200 per month in manual data entry costs. Book a technical consultation on STT selection for your business – we'll help calculate economic efficiency.

Main features of the STT module:

• Real-time streaming with partial results
• Offline fallback to on-device recognition
• Custom vocabulary support via phrase hints
• Automatic language detection (Russian/English)

Permissions and UX

On iOS, NSMicrophoneUsageDescription and NSSpeechRecognitionUsageDescription are mandatory in Info.plist. Request permissions before first use – via AVAudioSession.requestRecordPermission and SFSpeechRecognizer.requestAuthorization. A recording indicator in the UI is essential: an animated wave or pulsating indicator lets the user know they are being heard.

What's Included in the STT Integration Work

• Analysis of use cases and selection of the optimal API (on-device / cloud / hybrid)
• Architecture design with offline caching and fallback
• Implementation of integration with native or cloud services, including custom dictionaries
• Setup of test environment and accuracy testing on a representative sample (50+ dictations)
• Preparation of documentation on permissions, project configuration, UI recommendations (recording indicator)
• Integration with CI/CD (TestFlight, Firebase App Distribution)
• Training of the client's team on using the component
• Post-release support and monitoring of recognition quality

The Process

1. Analysis of use cases and selection of the optimal API (on-device / cloud / hybrid).
2. Architecture design considering offline mode and caching.
3. Implementation of integration with native or cloud services, including custom dictionaries.
4. Accuracy testing on a representative sample (50+ dictations).
5. CI/CD setup for builds with TestFlight and Firebase App Distribution.
6. Preparation of documentation on permissions and project configuration.
7. Post-release support and monitoring of recognition quality.

Estimated Timelines

Native STT (iOS/Android) with UI – 3–7 days. Streaming with cloud API and custom dictionary – 1–3 weeks. Cost is calculated individually – reach out for a detailed project estimate.

We guarantee recognition accuracy of at least 90% for the chosen scenario. Our engineers have Apple (iOS) and Google (Android) certifications and 10+ years of mobile development experience. We have completed 40+ projects with Speech-to-Text integration.

SFSpeechRecognizer Documentation

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.