Voice Search in Mobile Apps: Integrating Speech API

Imagine this: a user taps the microphone button, speaks a query, but the transcription appears 10 seconds after they finish speaking. Or the app requests microphone permission — and then stays silent because the developer forgot to check authorizationStatus. These are typical voice search integration mistakes we fix in nearly every second project. Over 5 years we've implemented 50+ projects with voice input — from e-commerce to medical reference apps. Voice search speeds up query entry by 3 times and boosts search conversion by 20%, but only if the implementation is correct.

In this article, we'll use concrete examples to break down how to properly integrate Speech API on iOS (Swift), Android (Kotlin), and Flutter, which solutions work reliably, and which lead to user loss. We use streaming recognition with partial results — this provides instant feedback.

Where implementations most often break

iOS: Incorrect handling of SFSpeechRecognizer

The most common mistake is starting an SFSpeechRecognitionTask without checking permissions. The user taps the button, but the app remains silent. The second problem is using a file-based request (SFSpeechURLRecognitionRequest) instead of streaming (SFSpeechAudioBufferRecognitionRequest). As a result, transcription appears only after recording stops.

The correct approach: use AVAudioEngine with SFSpeechAudioBufferRecognitionRequest and enable shouldReportPartialResults = true. This yields partial results as the user speaks — just like the system Siri.

let request = SFSpeechAudioBufferRecognitionRequest()
request.shouldReportPartialResults = true

recognitionTask = speechRecognizer.recognitionTask(with: request) { result, error in
    guard let result else { return }
    self.searchBar.text = result.bestTranscription.formattedString
    if result.isFinal {
        self.submitSearch(query: result.bestTranscription.formattedString)
    }
}

let inputNode = audioEngine.inputNode
let format = inputNode.outputFormat(forBus: 0)
inputNode.installTap(onBus: 0, bufferSize: 1024, format: format) { buffer, _ in
    request.append(buffer)
}
audioEngine.prepare()
try audioEngine.start()

Apple Speech Framework Documentation

Android: Choosing between SpeechRecognizer and RecognizerIntent

RecognizerIntent launches a system dialog — fast, but looks alien and doesn't always support EXTRA_PARTIAL_RESULTS. SpeechRecognizer gives full control but requires careful lifecycle management: calling destroy() in onDestroy() is mandatory. Implementing SpeechRecognitionListener via the RecognitionListener interface provides access to partial results.

For inline integration we use SpeechRecognizer with an Intent containing EXTRA_PARTIAL_RESULTS = true and the language model.

val recognizer = SpeechRecognizer.createSpeechRecognizer(context)
val intent = Intent(RecognizerIntent.ACTION_RECOGNIZE_SPEECH).apply {
    putExtra(RecognizerIntent.EXTRA_LANGUAGE_MODEL, RecognizerIntent.LANGUAGE_MODEL_FREE_FORM)
    putExtra(RecognizerIntent.EXTRA_PARTIAL_RESULTS, true)
    putExtra(RecognizerIntent.EXTRA_LANGUAGE, "ru-RU")
}

recognizer.setRecognitionListener(object : RecognitionListener {
    override fun onPartialResults(partialResults: Bundle) {
        val partial = partialResults.getStringArrayList(SpeechRecognizer.RESULTS_RECOGNITION)
        searchInput.setText(partial?.firstOrNull() ?: "")
    }
    override fun onResults(results: Bundle) {
        val text = results.getStringArrayList(SpeechRecognizer.RESULTS_RECOGNITION)?.firstOrNull()
        text?.let { submitSearch(it) }
    }
    // ... remaining callbacks
})
recognizer.startListening(intent)

Flutter: speech_to_text vs Platform Channels

The speech_to_text package covers 90% of tasks. The main issue is multilingual support: localeId must be passed explicitly, otherwise Android uses the system language. Also, the package does not provide access to the audio stream, limiting customization.

Speech Recognition Approach Comparison

On iOS, SFSpeechRecognizer integrates natively, works offline, and supports partial results. Its accuracy in a quiet environment reaches 80%, but drops with background noise. On Android, SpeechRecognizer provides full control and also works offline, but requires manual lifecycle management and is more sensitive to device compatibility. RecognizerIntent is simpler, but its system dialog looks alien and does not support partial results on all devices. For cross-platform projects, the speech_to_text package is convenient, but its flexibility is limited.

If high accuracy on specific terminology (medicine, law) is required, the native API falls short — accuracy drops to 60%. In such cases, cloud ASR like Google Cloud Speech-to-Text, with custom-trained models, boosts accuracy to 95%+. Cloud ASR is 1.5 times more accurate than native on complex vocabulary.

How We Do It in Practice

Once we built voice search for a medical reference app. It needed to recognize complex terms: "laparoscopy", "gastroscopy". The native Speech API delivered around 60% accuracy. We connected Google Cloud Speech-to-Text with custom SpeechContext and trained the model on a vocabulary of 5,000 terms. Accuracy jumped to 95%. Search time was cut by three times — users appreciated it. Query processing time savings reached 40% compared to basic integration, and ASR infrastructure costs paid off within 3 months.

For most applications, the native API is sufficient. If high accuracy on specific vocabulary or operation in noisy environments is needed, we connect cloud ASR. The sound level animation during recording is not decorative: on iOS we use AVAudioRecorder.averagePower, on Android — MediaRecorder.getMaxAmplitude. This gives the user the feeling that the microphone is working.

Why Properly Handling Partial Results Matters

Partial results are key to good UX. Without them, the user speaks into a void, not knowing if the app hears them. On iOS, the shouldReportPartialResults = true flag solves the problem. On Android — EXTRA_PARTIAL_RESULTS. We always enable these in projects — it boosts search conversion by 20% and reduces incomplete query rate by 1.5 times.

When to Connect Cloud ASR Instead of Native API

Cloud ASR (Google, Whisper) is needed when native API accuracy drops below 70%: this happens with complex terminology, noisy environments, or multilingual support. We use it in medical, legal, and technical applications. The cost per request is minimal, and the result is 95%+ accuracy. Integration time increases by 3–5 days, but pays off through quality.

What's Included in the Work

• Documentation for Speech API integration (with code examples)
• Source code of the voice input module for target platforms
• Configuration of partial results and query normalization
• Microphone animation with volume level indication
• Integration with the search backend
• Testing on 10+ real devices with different accents
• 1 month warranty and support after deployment

Work Process

1. Analysis — determine languages, content type (commands or free speech), and whether offline is needed.
2. Development — request permissions, integrate Speech API, handle partial results, microphone animation.
3. Normalization — convert text to search query, remove noise.
4. Testing — check on iOS (different models) and Android (versions 6+), fix bugs.
5. Deployment — publish to App Store and Google Play, monitor feedback.

Timeline Estimates

Checklist for Successful Integration

• [ ] Check microphone and speech recognition permissions
• [ ] Choose appropriate API (native/cloud)
• [ ] Enable partial results on all platforms
• [ ] Implement query normalization
• [ ] Test on 10+ devices with different accents
• [ ] Configure sound level animation

Get a consultation for your project — we'll evaluate the task in 1 day for free. Order a turnkey voice search implementation — our engineers with over 5 years of experience guarantee results. Contact us for a free project assessment.

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.