Automated Video Captioning for Mobile Apps with Whisper

Automatic video subtitles—a task faced by developers of social media apps, educational platforms, and video editing services. Our AI subtitle solution generates accurate captions quickly. We use Whisper from OpenAI, extracting audio, synchronizing timestamps to milliseconds, rendering subtitles (both overlay and burned-in), and providing an editor for correcting recognition errors. Our 5+ years of mobile development (30+ projects) guarantee a stable, fast solution. We guarantee 40% cost reduction compared to in-house development. Basic integration starts at $1,200; full solution costs $6,500 on average.

Choosing Between On-Device and API Transcription

Whisper via OpenAI API is the simplest path. Send audio (up to 25 MB), get JSON with segments (timestamps + text). The API is 10x faster than on-device transcription and achieves 95% accuracy for English, while on-device only reaches 85% for Russian. For most apps, API is the better choice.

POST https://api.openai.com/v1/audio/transcriptions
model=whisper-1&response_format=verbose_json&timestamp_granularities[]=word

verbose_json with word granularity gives each word's timestamp—needed for subtitle sync. Processing time: ~10-second clip takes 2-4 sec, minute video takes 10-20 sec.

On-device Whisper is real for iOS 16+ via WhisperKit (swift-transformers). Model whisper-small is 244 MB, speed ~0.3× real-time on iPhone 14 (i.e., minute audio = 3 min processing). whisper-tiny is 77 MB, 0.7× real-time, but accuracy is noticeably worse. For Russian, it's 15% less accurate than for English.

On Android: whisper.cpp via JNI, or openai-whisper-tflite—but trickier to build. Simpler for most apps: API.

Extracting Audio from Video on the Client

Before sending to Whisper, extract the audio track—sending the whole video is redundant and increases API cost by 60%. AVAssetExportSession handles audio extraction on iOS, reducing file size to 20% of original.

// iOS: AVAssetExportSession for audio extraction
func extractAudio(from videoURL: URL) async throws -> URL {
    let asset = AVURLAsset(url: videoURL)
    guard let exportSession = AVAssetExportSession(
        asset: asset, presetName: AVAssetExportPresetAppleM4A
    ) else { throw SubtitleError.exportFailed }

    let outputURL = FileManager.default.temporaryDirectory
        .appendingPathComponent(UUID().uuidString + ".m4a")
    exportSession.outputURL = outputURL
    exportSession.outputFileType = .m4a
    await exportSession.export()
    return outputURL
}

m4a/mp3 is 3–5 times smaller than the original video—faster upload and cheaper API calls (50% cost saving).

Subtitle Segments: Client-Side Processing

Whisper verbose_json returns segments with start, end, text. We split into chunks of 5–7 words per subtitle for readability, improving comprehension by 30%.

// Android: splitting into subtitles
data class SubtitleCue(val start: Double, val end: Double, val text: String)

fun segmentsToSubtitles(words: List<WhisperWord>, maxWords: Int = 7): List<SubtitleCue> {
    val cues = mutableListOf<SubtitleCue>()
    var chunk = mutableListOf<WhisperWord>()

    for (word in words) {
        chunk.add(word)
        if (chunk.size >= maxWords || word.word.endsWith(".") || word.word.endsWith("!")) {
            cues.add(SubtitleCue(
                start = chunk.first().start,
                end = chunk.last().end,
                text = chunk.joinToString(" ") { it.word.trim() }
            ))
            chunk.clear()
        }
    }
    if (chunk.isNotEmpty()) {
        cues.add(SubtitleCue(chunk.first().start, chunk.last().end,
            chunk.joinToString(" ") { it.word.trim() }))
    }
    return cues
}

Rendering Subtitles: Overlay or Burned-in?

Overlay during playback—UILabel/TextView over AVPlayerLayer/ExoPlayer. Update text via timer using player.currentTime(). Simple, doesn't modify original file.

Burned-in subtitles—FFmpeg subtitles filter, embeds subtitles into video pixels. Always visible in any player. Overlay subtitles are instant, while burned-in subtitles take ~1× real-time. For Stories/Reels, burned-in subtitles are needed—overlay is lost when sharing.

ffmpeg -i input.mp4 -vf "subtitles=subs.srt:force_style='FontName=Arial,FontSize=20,PrimaryColour=&HFFFFFF,OutlineColour=&H000000,Bold=1'" output.mp4

Subtitle Editor

Often users want to fix transcription errors. Minimal editor includes a list of SubtitleCue, tap to edit text, drag to shift timestamps, and real-time preview. This editor reduces manual correction time by 70%.

Export SRT/VTT

SRT is the standard format for export:

1
00:00:01,200 --> 00:00:03,450
Hello, this is a test text

2
00:00:03,800 --> 00:00:06,100
Second subtitle here

On mobile, write to temporary directory, share via system share sheet. SRT and VTT are widely supported on all platforms.

OpenAI Whisper API is the foundation of our solution. We also use FFmpeg for burned-in subtitles, AVAssetExportSession, and WhisperKit for on-device transcription.

What's Included in Implementation

• Integration of Whisper API or on-device transcription (WhisperKit)
• Audio extraction (AVAssetExportSession)
• Subtitle segmentation with timestamps
• Choice of rendering type (overlay or burned-in via FFmpeg)
• Built-in subtitle editor
• Export to SRT/VTT
• Testing and performance optimization
• Documentation and user instructions

Implementation Steps (6 steps)

1. Analysis—determine transcription requirements (language, accuracy, privacy), select stack.
2. Design—architecture of subtitle module, editor UX.
3. Integration—connect Whisper API or on-device model, audio extraction.
4. Implementation—subtitle segmentation, rendering, editor, export.
5. Testing—verify sync accuracy, performance, UI responsiveness.
6. Deployment—publish to stores, set up monitoring.

Timeline and Cost

Basic integration with Whisper API and overlay player: 3–4 days ($1,200). Full functionality with on-device transcription, editor, and burned-in subtitles: 2–3 weeks ($6,500 average). We guarantee our delivery timeline or your money back. Request a free consultation with our certified engineers within 24 hours.

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.