Voice Bot Implementation in Mobile Apps

Imagine: you launch a voice assistant in a navigation app. The user says 'Route to Minsk, avoiding toll roads,' but the response comes after 4 seconds, and half the words are misrecognized—'Minsk' turns into 'mink.' This isn't a hypothesis; it's a typical scenario when integrating a voice bot without considering latency, ambient noise, and the quirks of Russian speech. We've encountered such cases in projects for logistics companies and retail: losing 30% of users on first interaction due to slow response.

To avoid this, we build a pipeline: Speech-to-Text → NLU/LLM → Text-to-Speech with total response time under 1.5 seconds. The key technique is streaming processing at each stage: start transcription before the phrase ends, generate response from the first tokens, and synthesize speech concurrently with output. We also tune configurations per device (iOS/Android) and acoustic environment (quiet office vs. noisy street).

Below is a breakdown of the main components and their impact on dialogue quality: STT engine, TTS synthesizer, audio session management, and wake word. We optimized each block on real projects handling up to 10,000 requests per day.

Why Latency Is the Main Enemy of a Voice Bot

A typical latency breakdown shows where time is lost:

The optimized variant is 2–3 times faster. The key is not waiting for each stage to finish:

• STT with shouldReportPartialResults = true—start processing before the phrase ends.
• LLM streaming—start synthesis as soon as the first tokens arrive.
• TTS streaming—start playback while the rest of the phrase is still being synthesized.

Choosing an STT Engine for Russian

For conversational dialogues with specialized vocabulary (medicine, law), we recommend Yandex SpeechKit—it's trained on a large Russian corpus and gives 15–20% fewer errors than native STT.

// iOS: AVAudioEngine → Yandex SpeechKit streaming
class VoiceBotRecorder {
    private let audioEngine = AVAudioEngine()
    private var recognitionStream: RecognitionStream?

    func startRecording() throws {
        let inputNode = audioEngine.inputNode
        let format = AVAudioFormat(commonFormat: .pcmFormatInt16,
                                   sampleRate: 16000,
                                   channels: 1,
                                   interleaved: false)!

        recognitionStream = speechKitClient.createStream(config: streamConfig)

        inputNode.installTap(onBus: 0, bufferSize: 4096, format: format) { [weak self] buffer, _ in
            guard let pcmData = buffer.int16ChannelData?[0] else { return }
            let bytes = Data(bytes: pcmData, count: Int(buffer.frameLength) * 2)
            try? self?.recognitionStream?.send(audio: bytes)
        }

        audioEngine.prepare()
        try audioEngine.start()
    }
}

TTS: Speech Synthesis That Doesn’t Annoy

• ElevenLabs—best quality, voice cloning for branding, streaming via WebSocket.
• OpenAI TTS—models tts-1 (fast) and tts-1-hd (high quality); for Russian, the nova voice sounds most natural.
• Yandex SpeechKit TTS—voices alena, filipp, jane; streaming via gRPC, naturalness on par with ElevenLabs for Russian-speaking audience.
• Native synthesis—AVSpeechSynthesizer (iOS) and TextToSpeech (Android)—free, offline, but robotic.

Choosing a TTS depends on budget and naturalness requirements: for simple responses, native is enough; for a 'consultant assistant' service, only cloud solutions will do. Contact us for help selecting the optimal TTS for your scenario.

Audio Management: Avoiding Echo and Interruptions

iOS. The AVAudioSession category must be .playAndRecord with the .defaultToSpeaker option. When playing TTS, temporarily deactivate the microphone: AVAudioSession.sharedInstance().setActive(false) before synthesis, true after.

try AVAudioSession.sharedInstance().setCategory(
    .playAndRecord,
    options: [.defaultToSpeaker, .allowBluetooth]
)

Android. Use AudioManager.requestAudioFocus() during playback, abandonAudioFocus() after. Bluetooth headsets require separate handling via BluetoothHeadset.

Barge-in (interruption). If the user starts speaking while the bot is still responding, detect speech → stop TTS → start recording. Voice Activity Detection can be implemented via AudioRecord.getMaxAmplitude() or using WebRTC VAD for greater accuracy.

Wake Word and Hands-Free Mode

For scenarios like 'driving navigation' or 'smart glasses,' we add a wake word, e.g., 'Hey Assistant,' that activates the bot without a button press. We use Picovoice Porcupine—supports custom wake words and runs entirely on-device. An alternative is OpenWakeWord (open source). Resource consumption: ~50 MB RAM, active only during listening.

What’s Included in the Work

1. Audit of current architecture—assessing requirements for language, latency, speaker diarization.
2. Engine selection—STT/TTS per task (budget, accuracy, offline mode).
3. Audio pipeline development—microphone capture, PCM/WAV encoding, streaming to cloud APIs.
4. NLU/LLM integration—from simple intent routing to response generation via OpenAI or Yandex GPT.
5. Latency optimization—streaming, caching, granular buffer tuning.
6. UI/UX—state visualization (listening, thinking, speaking), sound wave animation, error handling.
7. Testing—on real devices with different accents, background noises (cafe, street, subway).
8. Documentation and training—integration description, API keys, support procedures.
9. Post-launch support—monitoring, model fine-tuning for growing phrase base.

Our Experience and Guarantees

We’ve been working with voice interfaces for over 5 years—implementing more than 30 projects for iOS and Android in Russian and international companies. Our engineers are certified Swift and Kotlin developers with experience integrating Yandex SpeechKit, ElevenLabs, and OpenAI. We guarantee stable operation under loads up to 10,000 requests per day and an SLA on response time of no more than 2 seconds.

Timeline Estimates

• Basic version with native STT/TTS—1 week.
• With cloud engines (Yandex SpeechKit / ElevenLabs), streaming, and ~1s latency—3–5 weeks.
• With wake word and hands-free—+1–2 weeks.

Work Stages

1. Analysis and prototype—2–4 days.
2. Audio pipeline design—2–3 days.
3. Implementation—1–3 weeks.
4. Testing and debugging—3–5 days.
5. Deploy to App Store / Google Play—1–2 days.

Contact us to evaluate your project—we’ll find the optimal configuration for your budget and requirements. Request a demo of a working voice bot on real data.

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.