AI Voice Message Transcription in Mobile Apps

Voice messages are the worst format for quick information retrieval. Especially in corporate chat: a minute-long voice note instead of one line of text. We solve this by embedding transcription directly into your mobile app. The user speaks—the app converts speech to text with up to 95% accuracy. The text is synchronized with audio: tap any word to hear it. Technically, this means integrating with Whisper API or local solutions, implementing audio capture, converting to the optimal format (16 kHz mono MP3 32 kbps), sending to the server, and receiving a transcript with timestamps every 200–300 ms. Then post-processing, noise notation filtering, and displaying in an interactive UI. The entire cycle from button press to visible text takes 0.5 to 3 seconds depending on recording length. One of our projects in fintech showed a 60% reduction in meeting protocol processing time—saving approximately $2,000 per month.

How Audio Capture Works in the App

In practice, a mobile app works with two paths:

Recording inside the app. The user records directly in your app—native capture, full control over the format. We use AVAudioRecorder on iOS and MediaRecorder on Android. The audio sample rate of 16kHz is chosen based on the Nyquist theorem for speech (max 8kHz frequency content), and the MP3 codec at 32kbps ensures compression without significant loss of phonetic information.

Importing an external file. Get a WAV/MP3/OGG from a messenger via share sheet. On iOS—UTType.audio in UIDocumentPickerViewController. On Android—ACTION_GET_CONTENT with "audio/*". File format matters. OGG Opus (Telegram format) Whisper understands natively. AMR (old Android messengers)—needs conversion. On the server, ffmpeg handles conversion of any format:

import subprocess

def convert_to_mp3(input_path: str, output_path: str) -> None:
    subprocess.run([
        "ffmpeg", "-i", input_path,
        "-ar", "16000",      # 16kHz is enough for speech
        "-ac", "1",          # mono
        "-b:a", "32k",       # 32kbps for speech
        output_path
    ], check=True)

16kHz mono MP3 32kbps—optimal for Whisper: quality doesn't drop, file size is minimal.

Why Whisper API Is Not the Only Option

Whisper API: A 10-second message processes in 0.5–1.5 s. A 1-minute message in 3–8 s. This includes processing time on OpenAI servers plus network. Acceptable for the user if progress is shown. According to OpenAI Whisper GitHub repository, the model achieves state-of-the-art accuracy.

Deepgram Nova-2—real-time streaming transcription, latency <300 ms on short fragments. More expensive than Whisper, but faster. Deepgram Nova-2 offers latency under 300ms, which is up to 10x faster than Whisper API for short fragments.

Local Whisper (self-hosted). faster-whisper on GPU (RTX 3090) processes 1 minute of audio in 2–4 seconds. On CPU—15–30 seconds. If data cannot be sent to the cloud—the only option.

Client-side transcription on iOS. SFSpeechRecognizer—native Apple Speech framework, works on-device (since iOS 16), free, no data sent. But: supports only a limited set of languages, quality lower than Whisper, limit of 1 minute per request.

// iOS — local transcription via SFSpeechRecognizer
let recognizer = SFSpeechRecognizer(locale: Locale(identifier: "ru-RU"))
let request = SFSpeechURLRecognitionRequest(url: audioURL)
request.shouldReportPartialResults = true

recognizer?.recognitionTask(with: request) { result, error in
    guard let result else { return }
    DispatchQueue.main.async {
        self.transcriptText = result.bestTranscription.formattedString
    }
}

For short personal notes, SFSpeechRecognizer is a good option without server costs. For corporate meeting recordings—Whisper or Deepgram.

Comparison of Transcription Methods

How to Display Transcript with Timestamps

Simple transcription—just text. Good transcription on mobile:

• Interactive text with timestamps: tap a word → audio jumps to that moment
• Punctuation (Whisper restores it well, but not perfectly—sometimes post-processing needed)
• Paragraphs by pauses (Whisper segments audio—use segments for splitting)
• Copy all text button
• Search within transcript

For messenger-style functionality: transcript appears streaming—don't wait for full completion, show segments as they become ready.

Transcript Post-Processing

Whisper sometimes inserts [Music], [Applause] in Whisper notation, transcribes background noise. We filter them:

import re

def clean_transcript(text: str) -> str:
    # Remove Whisper notations like [Music], [Noise]
    text = re.sub(r'\[.*?\]', '', text)
    # Remove extra spaces
    text = re.sub(r'\s+', ' ', text).strip()
    return text

For business scenarios, LLM post-processing is useful: fix proper names, terms, add punctuation where Whisper made mistakes. This server-side transcription for mobile ensures high quality.

What's Included in the Work

• Source code of the transcription module for iOS and Android
• Documentation on architecture and REST API (if server side)
• Access to services (OpenAI, Deepgram) with ready keys
• Team training and consultations during integration
• 24/7 support after launch

How to Implement Transcription: Step-by-Step Plan

1. Analysis — discuss use cases, stack, latency and privacy requirements.
2. Design — architecture of capture, transcription, and display.
3. Implementation — integrate Whisper/Deepgram, code for iOS/Android, server-side conversion.
4. Testing — validate on real recordings, optimize for your case.
5. Deploy — release to App Store and Google Play, set up monitoring.
6. Documentation and training — hand over code, instructions, train your team.
7. Support — guarantee 24/7 stability after launch.

Timelines and Cost

Basic transcription via Whisper with plain text display — 1–2 weeks. Full tool with interactive text, timestamps, and post-processing — 3–4 weeks. Integration cost is determined individually.

With over 5 years of experience in mobile speech integration and 30+ successful implementations for fintech and healthcare clients, our team ensures reliable delivery. Contact us for a free assessment of your project. Request a demo version to test 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.