
Key takeaways
• A video translation app is a three-stage pipeline: STT → MT → TTS. Apple’s Translation API covers common pairs on-device for free; Deepgram, OpenAI Realtime, ElevenLabs and Cartesia handle the rest in the cloud at $0.005–$0.12 per minute in 2026.
• Real-time live calls live or die at under 1.5 s glass-to-glass. Budget 200–400 ms STT + 100–300 ms MT + 90–300 ms TTS + 100–250 ms WebRTC. Stack the wrong tools and you pass 2 s before the first word lands.
• VOD dubbing is a different game. HeyGen, Synthesia and Rask AI generate lip-synced dubs in 130–175 languages at $0.20–$0.50 per minute. A 60-minute video in five languages: 4–6 hours, $250–$500 — versus $5,000–$20,000 for voice actors.
• On-device wins on privacy and unit economics; cloud wins on languages and voice quality. For HIPAA, GDPR or NDA-bound calls, default to Apple’s on-device stack on iOS 26 and later. For 50+ languages or premium voice cloning, go cloud.
• An MVP ships in 6–10 weeks for $30–60k. A full real-time platform with WebRTC, voice cloning and multi-party support is a 4–6 month build at $150–300k; infra runs $5–30k/month at SaaS scale.
Updated November 2026: Apple Live Translation and the CallKit translation API for calling apps, OpenAI gpt-realtime-translate at $0.034/min, Cartesia Sonic 3.6 and ElevenLabs Dubbing v2 pricing.
If you are scoping a video translation app in 2026, the hard part is no longer the AI. Accuracy on clear audio now sits at 90–98%. The hard part is choosing which of a dozen providers to wire together, hitting a latency target users won’t notice, and not overpaying for a stack you could have run on-device for free. Pick wrong and you ship a 2.5-second lag that feels broken; pick right and it feels like a human interpreter. This is the same decision tree we walk clients through during scoping.
More on the general case: our companion guide on real-time video translation across platforms covers the browser and web side; this one is the iOS-first build.
Just need to translate a few videos? Don’t build. An off-the-shelf video translation app is faster and cheaper: HeyGen or Synthesia for lip-synced dubs, Rask AI for batch, or Canva and VEED for quick social clips. Pay per minute and move on.
Building your own makes sense when translation is core to your product — you need your own latency budget, your own language set, on-device privacy, or a branded in-app experience the SaaS tools can’t give you. That’s the decision this guide is built for, and the five-question decision framework below settles it.
Why Fora Soft wrote this playbook
We’ve built video and real-time communication systems since 2005 — 250+ projects across 20 years. The most directly relevant case is the software we developed for Blaze Streaming Media’s Volo: a speaker talks, the audience scans a QR code, picks a language, and reads captions or hears a voiceover in the browser — no app install. Under the hood it’s Vonage for the speaker’s audio, Speechmatics for speech recognition, a socket.io caption fan-out, and Google Translate and Google TTS for translation and voiceover. Blaze runs Volo at conferences such as Black Hat USA, GDC and HIMSS.
For enterprise multilingual calling, our Nucleus on-prem WebRTC + SIP platform serves 5,000+ businesses and processes 600 million call minutes a month with real-time voice-to-voice translation, under SOC II, GDPR and HIPAA. TransLinguist runs remote interpretation and AI speech-to-speech translation on the platform we developed and won an appointment to the NHS (NOE CPC) language services framework. For sub-second broadcast we built Worldcast, delivering 1.5 Gb/s HD concert feeds at 0.4–0.5 s to 10,000 concurrent viewers.
This guide is that scoping conversation written down: which stack, which provider, which latency target, what budget. Read it end to end if you’re evaluating partners; jump to the tools matrix if you just need to choose between Apple Translation, OpenAI Realtime, Deepgram and the rest.
Building a video translation app?
30 minutes with a senior engineer who has shipped real-time translation for live events. We’ll review your latency budget, recommend a stack, and quote a build.
The state of iOS video translation in 2026
Five shifts between 2024 and 2026 reshaped what an iPhone can do with video translation.
1. Apple’s Translation API went production-ready. Since iOS 18 you get a TranslationSession from SwiftUI’s .translationTask modifier, get session-based translation with availability checks, and ship a fully on-device translator with no cloud bill. Apple handles the model download, language packs and updates. See the Translation framework docs.
2. The Foundation Models framework shipped in iOS 26. Apple opened its on-device ~3B-parameter model — the one behind Apple Intelligence — to any developer: free, private, offline, no API key, with structured @Generable output. iOS 26.4 added token counting and context inspection, and the model is gaining vision. That means privacy-preserving inference for medical, legal and financial translation without audio ever leaving the phone. Details in the Foundation Models docs.
3. Cloud STT and TTS broke the latency floor. Deepgram Nova-3 streams ASR under 300 ms and lists at $0.0077/min monolingual or $0.0092/min multilingual on Deepgram’s pricing page (November 2026). Cartesia advertises sub-90 ms model latency, and Sonic 3.6, generally available since August 27, 2026, speaks 44 languages and 61 locales according to the Cartesia changelog. OpenAI’s gpt-realtime-2.1 bundles STT + reasoning + TTS at roughly $0.06–$0.11 per minute with caching — down sharply from the $0.16–$0.18 the older model cost a year ago. On May 7, 2026 OpenAI added gpt-realtime-translate, a dedicated speech-to-speech model that takes 70+ input languages into 13 output languages for $0.034 per minute (OpenAI announcement, May 2026).
4. Lip-synced video dubbing got cheap. HeyGen and Synthesia generate lip-synced dubs in 130–175 languages at $0.20–$0.50 per minute — several times cheaper than two years ago. A 60-minute corporate video dubbed into five languages is 300 output minutes: $60–$150 in generation fees, or $250–$500 once you budget voice-clone setup, re-renders and a review pass. It takes 4–6 hours instead of a five-figure studio bill.
5. Live Translation became a system feature, and calling apps can plug into it. iOS 26 (September 2025), not iOS 18, added Live Translation to Phone, FaceTime and Messages, and Apple says it runs entirely on device. Phone and FaceTime translate one-to-one calls in English, French, German, Portuguese and Spanish. On a Phone call the other person can use any handset, because your iPhone does the translating. VoIP apps get one hook: CallKit’s CXSetTranslatingCallAction (iOS 26+) switches the same system translation on or off, and the system then supplies translated captions and a transcript (Apple developer docs). You can’t pick the voice, the model or the language set, so a custom latency target or a long-tail language still means your own pipeline. iOS 27 shipped in September 2026 and is now the current release. When you want these system pieces inside a real product, our AI mobile app development team builds on-device iOS features on Translation, Speech and Foundation Models.
Five use cases worth building for
1. Live meeting translation. Under 1.5 s glass-to-glass for 2–20 participants: international sales calls, multilingual standups, cross-border investor pitches, conference keynotes. It either closes deals or replaces interpreters at $200–$600/hour.
2. VOD dubbing and captioning. Batch processing with a 4–48-hour SLA: course platforms, corporate training, marketing video, podcast video. Economics: $250–$500 per hour of video for a full lip-synced dub in five languages, versus $5,000–$20,000 for a voice-actor workflow.
3. Live broadcast subtitles. 1–3 s is acceptable: concerts, sports, news. Worldcast delivers concert audio to 10,000 viewers at sub-second; layering Deepgram ASR for live captions adds about 500 ms.
4. Language learning and accessibility. On-device, offline-friendly, low latency. Apple’s Speech framework + Translation API + AVSpeechSynthesizer give you a full pipeline that runs offline, costs nothing per minute, and is COPPA- and GDPR-safe by default.
5. Multilingual customer-support video. Hybrid batch plus real-time: record support clips with auto-captions in five languages, then switch to live translation if the user escalates to an agent. This is where a video translation app pays for itself through deflected support tickets.
Architecture — the STT → MT → TTS pipeline
Every video translation app, real-time or batch, runs the same three logical stages: speech-to-text (STT), machine translation (MT), text-to-speech (TTS). What differs is where each stage lives — on-device or cloud — how aggressively each is streamed, and what you do with the audio afterward.

Figure 1. The three-stage pipeline every video translation app runs, with the on-device (free, private) lane above the cloud (broad coverage) lane and per-stage latency.
| Stage | On-device option | Cloud option | Latency |
|---|---|---|---|
| STT | Apple Speech, whisper.cpp | Deepgram Nova-3, Speechmatics, AssemblyAI | 200–400 ms streaming |
| MT | Apple Translation, Foundation Models | DeepL, Google, OpenAI | 100–300 ms |
| TTS | AVSpeechSynthesizer | Cartesia Sonic, ElevenLabs, OpenAI | 90–300 ms first chunk |
| Transport | On-device only (no network) | WebRTC: LiveKit, Twilio, Daily, mediasoup | 100–250 ms ingress+egress |
For a 1-to-1 live call the realistic total is 600–1,300 ms. For a 1-to-many broadcast you can afford 1–3 s and batch translation in 5–10 s windows, fanning captions out to browser tabs instead of opening a WebRTC connection per viewer.
Tools matrix — STT, MT, TTS providers head to head
Prices are 2026 list rates. Read the “where it breaks” column first — that’s where a stack quietly fails in production.
| Provider | Role | 2026 price | Where it wins | Where it breaks |
|---|---|---|---|---|
| Apple Translation | MT (on-device) | Free | Privacy, zero cost, offline | Limited pairs; iOS only |
| Deepgram Nova-3 | STT | $0.0077/min ($0.0092 multilingual) | Sub-300 ms, scales to thousands of streams | English-first; needs network |
| OpenAI Realtime 2.1 | STT+MT+TTS | $0.06–$0.11/min | One API, long-tail languages | Cost spikes uncached; opaque |
| Cartesia Sonic | TTS | Usage-based | Sub-90 ms model latency, voice cloning | 44 languages, fewer than ElevenLabs v2 |
| ElevenLabs | TTS + dubbing | $0.50/min (v1, 29 languages); $2.20/min (v2, 92 languages) | Best voice-clone fidelity, emotion | Pricier per minute |
| OpenAI gpt-realtime-translate | Speech-to-speech MT | $0.034/min | One call, 70+ input languages | 13 output languages; cloud only |
| HeyGen / Synthesia | VOD dub + lip-sync | ~$0.20–$0.50/min | 175 languages, lip-sync | Batch only, not real-time |
Reach for OpenAI Realtime when: you want one vendor for the whole loop, need long-tail languages, and can turn on prompt caching — without caching, a long call drifts toward $0.30/min. We broke the math down in our OpenAI Realtime pricing analysis.
Apple Translation Framework — the on-device default
Apple’s translation API for iOS is the Translation framework: TranslationSession translates text on device, free, inside your own app. For most iOS use cases inside its supported language set, Apple’s built-in framework is the right starting point. It’s free, runs locally, keeps text off your servers, which simplifies GDPR, HIPAA and COPPA reviews, and ships with the OS. The session API has been production-ready since iOS 18 and supports availability checks per language pair.
A minimal SwiftUI translation session
import SwiftUI
import Translation
struct TranslateView: View {
@State private var input = "Hello, how are you today?"
@State private var output = ""
@State private var config: TranslationSession.Configuration?
var body: some View {
VStack(spacing: 16) {
TextField("Source text", text: $input)
Text(output).foregroundStyle(.secondary)
Button("Translate to French") {
config = .init(source: Locale.Language(identifier: "en-US"),
target: Locale.Language(identifier: "fr-FR"))
}
}
.translationTask(config) { session in
output = (try? await session.translate(input).targetText) ?? ""
}
}
}
Reach for Apple Translation when: your pairs fit the supported set, you ship iOS-first, and privacy or per-minute cost rules out cloud APIs.
Real-time pipeline — the latency budget
A live translation system feels broken above ~1.5 s glass-to-glass. Below ~700 ms it feels like a real interpreter. Teams that miss the target usually made one of three mistakes: underestimating a stage, double-counting a buffer, or ignoring cold-model warmup.

Figure 2. The glass-to-glass latency budget. Best case lands at 510 ms; a realistic stack is 1,290 ms — still under the 1.5 s “feels broken” line, but only if you preload models.
| Stage | Best case | Realistic |
|---|---|---|
| Mic capture + WebRTC ingress | 60 ms | 150 ms |
| STT (streaming) | 200 ms | 400 ms |
| MT | 100 ms | 250 ms |
| TTS (first chunk) | 90 ms | 300 ms |
| WebRTC egress | 60 ms | 190 ms |
Best case: 510 ms. Realistic: 1,290 ms. Add 200–500 ms of cold-model warmup on the first request and you see why “real-time AI translation” demos disappoint in production. Preload models on launch, profile end-to-end on the target device, and don’t trust a vendor’s “sub-second” number until you measure it on your own network. We go deeper on the transport choices in our OpenAI Realtime with WebRTC guide.
Reach for a custom Deepgram + Cartesia stack when: you need sub-700 ms interpreter feel, control over each stage, and predictable per-minute cost — the managed all-in-one bundles trade that control for convenience.
Need sub-second live translation on iOS?
We’ve shipped real-time AI translation for live events and platforms that carry 600M+ multilingual call minutes a month. 30 minutes is enough to scope your architecture.
Voice cloning and lip-sync for video dubbing
For VOD dubbing, raw text-to-speech in a stranger’s voice breaks the illusion. Modern pipelines clone the original speaker’s voice in the target language and re-animate the lips to match.
Voice-clone enrollment. Cartesia clones from about 10 seconds of audio, ElevenLabs from 30 seconds, ResembleAI from 1–5 minutes for higher fidelity. The clone keeps timbre, accent and prosody across languages — a US-accented English speaker cloned into French sounds like the same person speaking French, not a generic French voice.
Lip-sync. HeyGen, Synthesia and Akool map the phoneme sequence to mouth shapes. HeyGen reports voice cloning from a 30-minute sample at under 5% error and under 3% rhythm deviation. Under ~150 ms of audio-to-visual drift is invisible to most viewers; self-hosted options (SadTalker, Wav2Lip) work but cost engineering time and quality.
Quality bar. MCD (Mel-Cepstral Distortion) below 3.5 is human parity for a voice clone; MOS (Mean Opinion Score) above 4.2 means listeners rate the voice as natural. A/B-test with 50+ viewers per language before shipping — cultural acceptance varies sharply by market. Our AI dubbing and subtitle pipeline playbook walks the full batch flow.
Reach for HeyGen or Synthesia when: the job is VOD, you need lip-sync in 100+ languages, and a 4–48-hour turnaround is fine — do not build real-time infrastructure for a batch problem.
On-device privacy and Apple Foundation Models
Some translation never tolerates cloud audio: medical interpretation (HIPAA), legal proceedings (privilege), defense, financial advisory, and NDA-bound sales calls. For those, on-device is non-negotiable.

Figure 3. On-device versus cloud across seven axes. On-device wins privacy, cost, latency and offline; cloud wins language coverage and voice quality.
What runs on-device today. Apple’s Translation API, the Speech framework for STT, whose SpeechAnalyzer and SpeechTranscriber API (iOS 26+) transcribes long-form and live audio on device and can be preheated with prepareToAnalyze to cut the delay before the first result, AVSpeechSynthesizer for TTS (functional, less natural than ElevenLabs or Cartesia), and whisper.cpp ports for offline STT in 99 languages.
The Foundation Models framework. In iOS 26 Apple’s ~3B-parameter on-device model is a public API: summarization, rewriting and constrained translation run locally with no key and no network. iOS 27 (September 2026) went further: a public LanguageModel protocol lets a cloud provider such as Gemini or Claude back the same session, and the framework gained Vision-backed OCR and barcode tools. For sensitive workflows on iOS 26 and 27, make on-device the default and fall back to cloud only for unsupported languages.
The compliance payoff. If audio never leaves the device, GDPR, HIPAA and COPPA still apply to your app, but their scope shrinks to what the app itself stores and sends, and data-residency questions about recordings largely disappear. That shortens legal review a lot for healthcare, education and finance apps — the reason we default sensitive builds to Apple’s stack and add cloud only where the language set forces it.
Reach for Foundation Models when: you target iOS 26+, the audio is sensitive, and your pairs are common enough that on-device quality holds — it removes the cloud bill and the data-processing agreement in one move.
Cost model — what a video translation app actually runs
These figures are conservative because we use Agent Engineering to scaffold the WebRTC plumbing, paywall and analytics — legacy shops typically quote 30–50% higher for the same scope.

Figure 4. Dubbing one 60-minute video into five languages: the AI pipeline lands at $250–$500 against $5,000–$20,000 for a voice-actor studio — a 10–80× gap.
| Build | Timeline | Build cost | Monthly infra |
|---|---|---|---|
| On-device MVP (Apple stack) | 6–8 weeks | $30–45k | ~$0 translation |
| Cloud real-time MVP | 8–10 weeks | $45–60k | $5–12k |
| VOD dubbing platform | 10–14 weeks | $60–110k | $8–20k |
| Full real-time platform (WebRTC + clone) | 4–6 months | $150–300k | $10–30k |
Worked example — a real-time conference for 100 attendees, one hour. Deepgram STT at $0.0077/min is $0.46. Apple Translation is $0. Cartesia TTS for one speaker fanned to 100 listeners is ~$4.50. A LiveKit SFU for the room is ~$24. Total: about $29 per hour. Scale to thousands of attendees and the SFU bill would dominate — so the trick is to broadcast a single translated audio stream over a CDN instead of unicasting to every viewer.
A decision framework — pick your iOS translation stack in five questions
Five questions decide the stack. Answer them in order; the first “yes” that constrains you usually settles the rest.

Figure 5. The five-question decision tree, from “is the audio sensitive?” down to scale, with the stack each branch points to.
1. Is the audio sensitive? Yes (medical, legal, financial, NDA) means on-device first: Apple Translation + Speech + Foundation Models. No means cloud APIs open up more languages and quality.
2. What latency matters? Under 700 ms (interpreter feel) means Cartesia TTS + Deepgram or Speechmatics STT + Apple MT. Around 1.5 s (fine for meetings) means the OpenAI Realtime bundle. Batch (VOD) means Whisper or AssemblyAI + DeepL + ElevenLabs or HeyGen.
3. How many language pairs? A handful, EU or English-pivot, means Apple Translation works. 50+ languages means cloud MT (DeepL, Google, OpenAI). Long-tail languages mean a frontier LLM for zero-shot.
4. Live or VOD? Live means WebRTC plus streaming STT/MT/TTS on LiveKit, Twilio or mediasoup. VOD means HeyGen or Synthesia for full lip-sync, or ElevenLabs plus AVPlayer for voice-only.
5. Scaling to thousands of concurrent listeners? Yes means broadcast translated audio over HLS or LL-HLS — never unicast TTS per viewer. No (1-to-1, small group) means a per-viewer SFU is fine.
App Store rules and StoreKit 2
Privacy disclosure. If audio leaves the device for cloud STT/MT/TTS, declare it in your App Store privacy report under “Audio Data” — linked or not linked to the user, depending on whether you store identity. Apple flags missing disclosures during review.
StoreKit 2 monetization. Auto-renewable subscriptions for a premium tier (cloud translation, voice cloning, dubbing) plus pay-per-use credits via consumable IAP. Apple takes 30% standard, 15% under the Small Business Program (below $1M/year) and on year-two subscriptions. In the US you may also link out to an external checkout under the 2025 Epic v. Apple injunction. Apple asked the Supreme Court to review it in May 2026, so confirm the current commission terms before you model margins, and weigh the saving against a rougher checkout.
Background audio. For translation that continues when the app is backgrounded, set the right AVAudioSession category, declare the audio background mode, and add a Now Playing entry so iOS doesn’t kill the session.
Children’s apps. COPPA bars collecting personal data from under-13 users. On-device translation with no logging is COPPA-safe by default; cloud APIs need parental-consent flows and data-processing agreements.
KPIs every video translation app should track
Quality KPIs. Word Error Rate (target under 5% on clean speech, under 15% on noisy). BLEU on translation (25–35 baseline, 50+ near human parity). MOS on TTS (above 4.2 reads as natural). Audio-to-visual drift on dubbed video (under 150 ms is invisible).
Business KPIs. Glass-to-glass latency (under 1,500 ms for live calls). Session completion rate (above 85% for live calls; users ditch broken sessions fast). Trial-to-paid conversion (25–40% for SaaS pricing). Languages used per user, a proxy for engagement.
Reliability KPIs. STT/MT/TTS API failure rate (under 0.5%). Cold-model start time (under 500 ms via preload). Crash-free user rate (above 99.9%; elite is 99.93%+ — see our iOS optimization playbook).
Mini case — real-time translation software for Blaze Streaming Media’s Volo
Situation. Blaze Streaming Media wanted live captions and translation for in-person conferences, with nothing for attendees to install.
Plan. We developed a browser app that attendees open from a QR code, a producer console and an admin panel. Speech recognition runs on Speechmatics, translation on Google Translate and voiceover on Google TTS; Vonage carries the speaker’s audio, and captions reach phones over WebSockets.
Outcome. Blaze runs Volo at conferences such as Black Hat USA, GDC and HIMSS. Want a similar assessment? Grab 30 minutes.
Want a fixed-fee quote for your video translation app?
Share your concept and target latency. We’ll come back with a stack, a milestone plan and a price — usually 30–40% below legacy shops because Agent Engineering scaffolds the WebRTC, paywall and STT/TTS wiring for us.
Five pitfalls that wreck video translation apps
1. Latency stacking. Adding each vendor’s marketing number gives you “sub-second” on paper and 2.5 s in production. Measure end-to-end on the target device, on a real LTE connection, with a cold-started model. Then tune.
2. Voice mismatch across languages. A generic French voice for a US-English speaker kills the illusion. Clone the speaker (Cartesia 10 s, ElevenLabs 30 s) so the same person sounds like themselves in every language.
3. Lip-sync drift on long video. A perfect 3-second clip can drift 200–300 ms over 10 minutes. Test long-form output with 50+ viewers per language; reject takes that drift past 150 ms.
4. Cold-model warmup. The first request after launch can take 2–5 s while a model loads. Preload in the background on launch and show a “ready” state before letting the user start.
5. Cultural localization gaps. Literal MT of idioms, jokes and brand names reads as nonsense. For marketing and entertainment, pair MT with light human review on the highest-value clips — we covered the math in our hybrid AI + human translation guide.
AI in iOS video translation — the 2026 frontier
Multilingual frontier LLMs. Models trained natively on 100+ languages now do zero-shot translation for rare pairs (Welsh to Tamil, Khmer to Estonian) at quality that used to need custom training, for roughly $0.001–$0.02 per ~500 words.
Instant voice cloning. Cartesia clones a voice in about 10 seconds, ElevenLabs in 30. A user can upload one voicemail and get a translated clone of themselves speaking 30+ languages minutes later.
On-device inference. Apple Foundation Models, Qualcomm’s AI Engine and NVIDIA Jetson all push 3–7B-parameter models to the edge. On iPhone that already means on-device STT through SpeechAnalyzer and on-device translation through the Translation framework, which removes the cloud bill for supported languages.
Streaming inference primitives. Sonic 3.6, Deepgram Nova-3, streaming Whisper and OpenAI Realtime cut end-to-end latency 3–5× versus batch APIs. Together, sub-700 ms speech-to-speech is now real for a 1-to-1 call — the point where users stop noticing the AI.
When NOT to build a custom video translation app
If you only need one common pair on iOS 18+. The system Translate app handles English to Spanish natively, and since iOS 26 Live Translation covers one-to-one calls in Phone and FaceTime. A wrapper around it rarely earns its keep.
If your audience is under 500 monthly actives. Below that, engineering and infra dwarf revenue. Validate demand on a no-code MVP (a hosted SFU plus an LLM) first.
If your volume is one-shot. A single conference, launch video, or training video: pay HeyGen or Synthesia per minute. Don’t build infrastructure for one use.
If you can’t commit to ongoing cost ops. Real-time infrastructure needs continuous tuning of vendor pricing and language availability. No budget for a quarterly review means you should ship on a managed platform instead.
FAQ
How much does it cost to build a video translation app?
An on-device MVP on Apple’s stack ships in 6–8 weeks for $30–45k with near-zero translation cost. A cloud real-time MVP is $45–60k over 8–10 weeks. A full real-time platform with WebRTC, voice cloning and multi-party support is $150–300k over 4–6 months, plus $5–30k/month infra at SaaS scale.
Can I translate a video in real time?
Yes. A streaming STT + MT + TTS pipeline over WebRTC delivers live translation in 600–1,500 ms glass-to-glass. Below ~700 ms it feels like a human interpreter. The keys are streaming (not batch) APIs, preloaded models, and broadcasting one translated stream when you have many listeners.
Can AI translate a video for free?
On iOS, yes for common language pairs: Apple’s Translation API, Speech framework and AVSpeechSynthesizer run on-device at no per-minute cost. Free tiers of cloud tools exist but cap minutes and languages; true multi-language, lip-synced dubbing costs $0.20–$0.50 per minute.
How do I translate what someone is saying in a video?
Transcribe the speech to text (STT), translate the text (MT), then either show captions or synthesize a voice in the target language (TTS). For pre-recorded video, add lip-sync with HeyGen or Synthesia; for live video, stream all three stages over WebRTC.
What is the latency of the Apple Translation API?
On-device translation runs about 50–200 ms per sentence on modern Apple silicon, with no network round-trip. Combined with the Speech framework for STT, a fully on-device pipeline stays well under a second for supported pairs.
How much does the OpenAI Realtime API cost per minute in 2026?
A typical agent on gpt-realtime-2.1 runs about $0.06–$0.11 per minute with prompt caching on, and $0.02–$0.05 on the mini model. Without caching, long calls can reach $0.18–$0.46. Token rates are $32/$64 per million audio in/out on the flagship.
Can I do real-time translation entirely offline on iPhone?
Yes, for supported languages. Apple’s Speech framework, Translation API and AVSpeechSynthesizer, plus the iOS 26 Foundation Models framework, run a full pipeline with no network. A whisper.cpp port extends offline STT to 99 languages at higher latency.
How long does it take to dub one hour of video into five languages?
Roughly 4–6 hours end to end with HeyGen, Synthesia or Rask AI, including voice cloning and lip-sync, at $250–$500 total once re-renders and a review pass are budgeted. A human voice-actor workflow for the same output runs $5,000–$20,000 and takes days.
Does iPhone have built-in live call translation?
Yes. Live Translation arrived in iOS 26; iOS 18 did not have it. The feature in the Phone app translates one-to-one calls on device, speaks the translation to the other person and shows you a transcript; FaceTime shows translated captions. Call languages are English, French, German, Portuguese and Spanish.
Can my own calling app use Apple’s Live Translation?
Partly. On iOS 26+ a CallKit app can switch system translation on or off with CXSetTranslatingCallAction, and the system supplies captions and a transcript. You can’t choose the voice, model or languages, so anything beyond that needs your own STT → MT → TTS pipeline.
How do I translate a video on iPhone?
For a few clips, use an app: HeyGen or Synthesia for lip-synced dubs, Rask AI for batch jobs. To build video translation into your own app, chain speech-to-text, Apple’s Translation framework and text-to-speech, as in the pipeline above.
What is OpenAI gpt-realtime-translate?
A speech-to-speech translation model OpenAI released on May 7, 2026. It accepts 70+ input languages, speaks 13 output languages and costs $0.034 per minute, less than running gpt-realtime-2.1 as a translator.
Which transport is right — WebRTC, WebSocket, or HTTP?
WebRTC for live, low-latency 1-to-1 or small-group calls. WebSocket for streaming captions to many viewers (the software we developed for Volo sends captions to phones over socket.io). HTTP for batch VOD jobs where a few seconds of turnaround is fine.
What to read next
Real-time AI
OpenAI Realtime API Pricing: The Real Cost Per Minute
2026 token math and per-minute costs from thousands of measured sessions.
Transport
OpenAI Realtime with WebRTC, SIP and WebSockets
Sub-200 ms voice + video pipelines bridging browsers, telephony and AI agents.
AI translation
Real-Time Video Translation: The 2026 Playbook
The cross-platform companion to this iOS build guide.
Translation
Hybrid AI + Human Translation Services
When MT is enough, when human review pays off, and the cost math behind it.
iOS streaming
iOS Video Streaming App Development in 2026
AVPlayer, Mux, Cloudflare Stream, FairPlay and ABR ladders for an OTT iOS build.
Ready to ship a video translation app users actually trust?
In 2026 the right stack is a three-stage pipeline tuned to your use case: Apple Translation and Foundation Models on-device for common pairs and privacy-sensitive work; Deepgram or Speechmatics + Apple MT + Cartesia or ElevenLabs for sub-1.5 s real-time calls; HeyGen or Synthesia + ElevenLabs for VOD dubbing with lip-sync. Hit under 1.5 s glass-to-glass, under 5% WER and above 4.2 MOS, and the rest of the product follows.
Whether it’s live conference translation, multilingual support video, language learning, accessibility, or VOD dubbing, we’ve built it at conference, enterprise and consumer scale. We’d rather show you how the pieces fit in 30 minutes than write another paragraph.


The handbook · 7 volumes
Fora Soft Video Engineering Handbook
A video translation app is three models in a row, and the user only judges the voice that comes out the end. We wrote seven volumes on the video stack; start with Volume 7 on real-time pipelines, live captions and cost per stream-hour, then Volume 6 on ASR, TTS and generative video, and Volume 5 on timestamps and lip sync. Written by Nikolay Sapunov, CEO at Fora Soft.
Browse the handbook on Amazon →7 volumes · 9,000+ pages · Kindle & paperback
Let’s scope your video translation build
A 30-minute call with a senior engineer who builds real-time translation for live events. Bring your concept and target latency; we’ll come back with a stack, milestones and a price.
Building for a regulated vertical? Our AI integration services and video conferencing engineering teams have shipped HIPAA and GDPR translation into production.