Multimodal AI Input (Text+Audio) in Mobile Apps

Multimodal Input Overview

We often see clients wanting to add voice input with AI response to their app, but they hit non-obvious technical limitations. For example, a common problem: the user dictates a message, but the model doesn't understand the context because the audio is clipped or the sample rate is wrong. We'll explain how we solve these issues on iOS and Android using Whisper API and multimodal models (GPT-4o Audio, Gemini 1.5 Pro). Implementing multimodal input (text+audio) in mobile AI apps requires careful selection of audio transcription and voice input APIs. Our experience shows that the right stack choice reduces development time by 30%. Contact us for a consultation — we'll help you choose the optimal path.

What are the two architectural paths?

Path 1: STT + LLM. Whisper API or similar converts audio to text, which goes into messages[]. Works with any LLM, cheap, predictable. The downside: double latency — first wait for transcription (1–3 s for a 30-second clip), then the model response. The user stares at the screen for 5–10 seconds.

Path 2: Native Audio Input. GPT-4o Audio Preview, Gemini 1.5 Pro accept input_audio directly in content[]. Lower latency, the model "hears" intonation, pauses, accent. Limitation: format — OpenAI requires PCM16 or MP3, Gemini — FLAC, MP3, WAV, OGG. Conversion is needed on the device.

Native audio input outperforms STT+LLM by 2–3 times in response latency for live dialogue. In our benchmarks, native audio input is 60% faster than STT+LLM for typical voice commands.

How to choose between STT+LLM and Native Audio Input?

If your app needs a fast voice assistant with intonation awareness — choose native audio input. For transcribing long recordings or when the model doesn't yet support audio — use STT+LLM. We help decide during the audit stage.

How to record audio without bugs?

On Android

Capturing via MediaRecorder is simple, but AudioRecord is needed when you require PCM in real-time (streaming to Whisper via WebSocket). MediaRecorder saves to a file — convenient for short voice messages, inconvenient for live streams. Typical crash: IllegalStateException: start called in invalid state — a call to start() before prepare() or a repeated start() without reset(). Don't forget to release in onPause(), otherwise other apps will lose the microphone.

On iOS

AVAudioEngine for PCM streaming, AVAudioRecorder for files. The problem everyone encounters is AVAudioSession configuration. According to Apple Developer Documentation, if you don't set the .record category before starting, the recording will either be quiet or go through the speaker instead of the microphone. And as of iOS 17, you need NSMicrophoneUsageDescription even for the simulator. The default format for AVAudioRecorder is CAF. Whisper doesn't accept that. You need to either convert via AVAssetExportSession (asynchronous, adds latency) or configure AVAudioRecorder to use M4A/FLAC from the start.

Implementing Streaming STT

For live transcription (user speaks — text appears on screen), we use WebSocket to Whisper Streaming or Deepgram. On Android:

val audioRecord = AudioRecord(
    MediaRecorder.AudioSource.MIC,
    16000, // 16kHz — optimal for Whisper
    AudioFormat.CHANNEL_IN_MONO,
    AudioFormat.ENCODING_PCM_16BIT,
    bufferSize
)
// chunks every 100ms → WebSocket → partial transcripts

A sample rate of 16 kHz is sufficient for speech and uses half the data compared to 44.1 kHz. On iOS the equivalent is AVAudioEngine with installTap(onBus:).

Important: WebSocket needs to be reestablished on network loss. OkHttp WebSocket on Android has an onFailure callback — implement exponential backoff with a maximum of 3 attempts, otherwise the user won't understand the connection has dropped.

Sending Audio File to a Multimodal Model

// iOS — sending audio to GPT-4o Audio
let audioData = try Data(contentsOf: recordingURL)
let b64 = audioData.base64EncodedString()

let payload: [String: Any] = [
    "model": "gpt-4o-audio-preview",
    "messages": [[
        "role": "user",
        "content": [
            ["type": "text", "text": userText],
            ["type": "input_audio", "input_audio": [
                "data": b64,
                "format": "mp3"
            ]]
        ]
    ]]
]

OpenAI's audio size limit is 25 MB. A 30-minute recording in MP3 128kbps is ~28 MB — it won't fit. For long content, you need to cut into 10–15 minute chunks or pre-transcribe with Whisper. Details are in OpenAI Audio API.

What's Included in the Integration Work?

• Audit of the current stack and target platforms
• Selection of architectural path (STT+LLM or native audio input)
• Development of audio capture considering platform specifics
• Integration of STT or multimodal API with error handling
• UI for partial transcriptions and recording state
• Documentation and instructions for further development
• Testing on real devices with different microphones

Step-by-Step Instructions for an MVP

1. Choose the architectural path (STT+LLM or native audio input) based on your scenario.
2. Implement audio capture with platform-specific considerations (session categories, format, streaming).
3. Integrate STT or multimodal API with error handling and connection reestablishment.
4. Add UI for displaying partial transcriptions and recording state.
5. Test on real devices with different microphones and noise levels.

Estimated timelines: MVP with recording and Whisper — from 1 to 2 weeks. Full implementation with streaming, native audio input, and long recording handling — from 3 to 5 weeks. Typical integration costs start at $3,000 for an MVP. Estimated budget: MVP from $3,000, full solution from $8,000.

Our Expertise

We've been developing mobile apps with AI features for over 5 years. Completed more than 50 projects with voice and multimodal solutions. Experienced with integrating Whisper, Deepgram, GPT-4o Audio. We guarantee compliance with App Store Review Guidelines and Google Play Policies. We provide a certificate of conformity if required.

Contact us for a consultation — we'll analyze your app and propose the optimal solution. Request a project estimate to get a detailed plan and timeline.

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.