A user dictates a command, waits for a response — and after 3 seconds gets the wrong thing. Sound familiar? In a voice assistant, speed and recognition accuracy are paramount. We solve this by building a pipeline from VAD, STT, NLU, and TTS with a total latency from the end of the phrase to the response ≤1.5 seconds. This is a technical constraint we overcome by optimizing each stage: from voice detection to speech synthesis. We handle the entire cycle: design, development, integration with your CRM or ERP, and app store publication. Over 7+ years, we have deployed voice assistants in 15+ projects for iOS and Android. Evaluate our approach — contact us for a consultation.
How the Voice Assistant Pipeline Works
Microphone → VAD → STT → NLU → Logic → TTS → Speaker. Each component contributes to latency.
VAD — Voice Activity Detection, cuts silence. We use WebRTCVAD or SileroVAD (ONNX/TFLite, ~1 MB). This reduces empty STT requests and saves traffic.
Example VAD configuration on Android
```kotlin
val vad = SileroVAD.create(context)
vad.start { frame ->
if (vad.isVoice(frame)) {
// send audio to STT
}
}
```STT — Speech-to-Text. Options: native SFSpeechRecognizer (iOS) or Android Speech for simple scenarios; for high accuracy Russian — Yandex SpeechKit or OpenAI Whisper API. Apple Developer Documentation recommends using SFSpeechRecognizer for basic commands. Cloud recognition costs ~$0.006 per audio minute for Whisper API; on-device is free. A typical request lasts 3–5 seconds, so costs are minimal.
Why Latency ≤1.5s Is Critical
Users expect instant response. Delay over 2 seconds feels like a hang. We achieve this through parallel requests, local processing, and caching frequent intents. Each component has its typical delay:
| Component |
Typical Latency |
Cost (per audio minute) |
| VAD (Silero) |
30–50 ms |
$0 (on-device) |
| STT (Whisper API) |
200–400 ms |
~$0.006 |
| NLU (Rasa) |
200–400 ms |
$0 (self-hosted) |
| TTS (Yandex SpeechKit) |
200–500 ms |
~$0.002 |
Total — up to 1.5 seconds. Replacing NLU with an LLM (GPT-4) can increase latency to 4 seconds, unacceptable for real-time.
Intent Recognition: What Actually Works
For a limited domain (smart home, internet banking) — Rasa NLU or Dialogflow with 50–200 training examples per intent. For an open domain — LLM with function calling. Rasa NLU is better than Dialogflow for confidential data since it runs on your server and does not send speech to the cloud.
| Feature |
Rasa NLU |
Dialogflow |
LLM (GPT-4) |
| Privacy |
Full |
Google Cloud |
Cloud (prompt) |
| Domain Accuracy |
90%+ |
85%+ |
95%+ (but slower) |
| Setup Difficulty |
Medium |
Low |
High (prompt engineering) |
| Latency |
200–400 ms |
500–800 ms |
1–4 sec |
Case Study from Our Practice
Corporate assistant for field employees: voice task creation in CRM without unlocking the phone. Stack: SileroVAD on-device -> Yandex SpeechKit streaming -> Rasa NLU (self-hosted, 23 intents) -> CRM REST API -> Yandex SpeechKit TTS. Latency: median 1.1 s, p95 2.3 s. Rasa NLU provided full data control. The client estimated time savings of ~25% for employees.
How to Implement a Voice Assistant: Step-by-Step Plan
-
Scenario Analysis — define command list and contexts (up to 3 days).
-
Component Selection — STT, NLU, TTS considering language, privacy, and budget.
-
Pipeline Integration — connect modules, tune VAD parameters and timeouts.
-
Testing on Real Data — record dialogues, A/B tests, optimization.
-
App Store Release — prepare metadata, test with TestFlight/Internal Track.
When On-Device STT Is Needed?
If the app must work offline or requires minimal latency — choose on-device. Accuracy is lower (80–90%), but latency is 300–500 ms and no API call costs. For Russian, on-device still lags behind cloud solutions but works for a limited set of phrases.
What’s Included in the Work
- Architecture documentation and API specifications
- Configured CI/CD for build and deployment
- Source code and repository access
- Training your team on the voice pipeline
- Post-release support for 2 weeks
Estimated Timelines
- Basic pipeline STT + NLU + TTS: 2–3 weeks
- With wake word and context: 4–6 weeks
- With integration into existing infrastructure: individually defined
Cost is determined after analyzing your requirements. Get a consultation for your scenario — contact us. We will evaluate your project in 1–2 days.
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.
| Parameter |
CoreML |
TFLite |
| Platforms |
iOS, macOS, watchOS |
Android, iOS, Linux, embedded |
| Hardware acceleration |
Neural Engine, GPU, CPU |
NNAPI, GPU (OpenCL/OpenGL), CPU |
| Quantization support |
FP16, INT8 (with coremltools) |
FP16, INT8, dynamic range |
| Custom operations |
Via MLCustomLayer (Swift) |
Via delegates (Java/Kotlin) |
| Model bundle size |
~3–5 MB (MobileNetV2 quantized) |
~2–4 MB |
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.
| Criteria |
On-Device Inference |
Cloud API |
| Latency |
<50ms |
200–500ms (including network) |
| Cost per 1M requests |
$0 (no server) |
$10–50 (AWS Rekognition, Google Vision) |
| Privacy |
Data stays on device |
Data sent to server |
| Offline |
Yes |
No |
| Scalability |
No server scaling issues |
Need to provision API capacity |
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
-
Task analysis — measure latency, privacy, size, supported devices.
-
Model prototyping — in Python, evaluate accuracy on target data.
-
Conversion and quantization — for CoreML/TFLite with validation.
-
Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
-
Testing — on real devices, measure FPS, RAM, battery.
-
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.
