# AI Video Translator

> Desktop application for translating and dubbing video with AI: transcription →
> translation → text-to-speech → muxed output. Available on Windows (Microsoft Store); the macOS version
> (Apple Silicon, macOS 15 or later) is coming soon to the Mac App Store. No Linux, mobile
> or web version by design.

Positioning in one line: the video file never leaves the user's machine. Every
step can run on-device, or be handed to a cloud provider on the user's own API
key, and the two can be mixed inside a single project — a different vendor per
step, or even per speaker.

- **Pricing.** Free tier has every feature and exports the complete video and
  background audio, with dubbed voice on the first 10 speech segments. A Premium subscription,
  purchased in the store the app was downloaded from (the price is shown there),
  adds dubbed voice for every segment. That is the only difference. No watermark on any tier,
  free or paid. Commercial use is allowed on the free tier.
- **Providers.** Eight providers on Windows. The macOS build also
  includes a ninth provider, Apple (on-device), for speech recognition and for
  translation (Apple Foundation Models, where Apple Intelligence is available and
  turned on, and Apple Translation with language packs the user has installed in
  macOS) on macOS 26+ with supported hardware and languages. It also offers
  text-to-speech with Apple system voices already installed on the Mac (macOS 15+),
  without requesting voice downloads. Cloud: OpenAI, xAI (Grok), ElevenLabs,
  Google (Gemini), Inworld, Soniox — billed by the provider directly, at their
  public rates, with no markup. Paid through the user's own account and API
  key; we never proxy the request. Configurable: any OpenAI-compatible endpoint the
  user enters (LM Studio, Ollama, or a hosted aggregator), used for
  translation. Hosted endpoints send text to their host and may charge for usage;
  a server running on the user's computer keeps that text local. On-device models
  include Whisper, Omnilingual ASR, Qwen3.5, Qwen3.6, Gemma 4, Hy-MT2,
  Index-Translate, Supertonic 3, Kokoro, UVR-MDX-NET and Spleeter among 30+
  models, which need no key or internet after installation. Translation models
  range from Qwen3.5 0.8B to Gemma 4 12B and, for a very powerful computer with
  lots of memory, Index-Translate 35B-A3B, Qwen3.6 35B-A3B and Gemma 4 26B-A4B.
  Qwen3.5 0.8B is the
  default local translator on Windows. New Mac projects
  default to Apple Foundation Models when Apple Intelligence is available and
  supports both languages, otherwise to Gemma 4 E2B. For Apple speech recognition, macOS
  installs missing language assets from Apple after approval in the app, and
  manages their updates and retention. Apple's on-device language model for
  translation and in-app Help answers is likewise managed by macOS; the app does
  not download it, and it does not use Apple's server-side options such as
  Private Cloud Compute. Apple Translation languages and Apple system voices are
  installed and managed by macOS; the app does not download them.
- **Apple on Mac (macOS build only).** The Apple provider calls the models
  built into macOS, so on macOS 27 the app works with Apple's macOS 27 models.
  It needs no API key and adds no provider charge. On macOS 26+ with a supported
  Mac: Apple Speech transcription, with optional speaker labels produced by the
  app's own on-device speaker models (the source language must be chosen;
  automatic detection is not supported); Apple Translation, using translation
  languages installed in System Settings, without Apple Intelligence and without
  project context or glossary; Apple Foundation Models translation, which needs
  Apple Intelligence turned on and ready and receives the project context and
  glossary. On macOS 15+: speech with Apple system voices already installed on
  the Mac (Personal Voice is not used). The in-app Help answers questions with
  Apple Foundation Models from the bundled English help pages, citing its
  sources, and falls back to local search without Apple Intelligence. Help can
  be searched from the Shortcuts action "Search AI Video Translator help" and,
  on macOS 27, by asking Siri, which opens the Help panel with the query. The
  languages each Apple model covers come from macOS at run time; the app checks
  them before running. The Mac app itself requires Apple Silicon and macOS 15+.
- **External control (Shortcuts, MCP).** The app owns the translation job;
  Shortcuts and AI agents request it. Mac only — use the Shortcuts app:
  "Translate video" starts a job, confirms acceptance and returns it (with
  Check translation, Get translated video and Cancel translation actions), and
  "Translate video and wait" returns the finished file and shows its path in a
  success dialog. Check/Get
  with no Translation job selected use the most recently added cached job and
  name its ID in the dialog; progress updates do not change creation order.
  With no known jobs, they ask you to start one with Translate video. Legacy
  cache entries without ordering are older than ordered entries; if all entries
  lack ordering, select a job in the action before running. Cancel with no job
  selected uses the sole non-terminal job; none reports no running translation,
  and several require selecting the job in the editor. No runtime job picker is
  used. The editor picker shows source basename, target language, state and
  native first-observed date/time (or date unavailable for legacy entries).
  Cancel translation runs without confirmation for explicit and sole-active jobs;
  use Shortcuts' own "Ask Before Running" for a guard. The app's
  new-project defaults are used unless a provider, model or voice is chosen.
  A missing choice fails immediately with a clear error: choose the provider,
  model or voice in the Shortcut, then run it again. There are no runtime
  questions. Asking Siri to translate a video will follow once Siri supports it.
  The result is saved once in Downloads, with a completion notification if
  notifications are allowed. Free exports dub only the first 10 speech segments
  and report this allowance. "Full dub required" refuses an over-limit run
  before translation or voice generation. MCP prepare_translation reports the
  allowance before an agent starts a job, and the result repeats it as a warning.
  Consent for external AI services is only ever given
  in the app; without it neither front sends anything to an external AI
  provider, and the request is refused until the user agrees in the app (MCP
  reports externalAiConsentRequired). Mac and
  Windows — MCP for AI agents (off by default; enable Settings → MCP first and copy host setup there): the app run with `--mcp` is a local stdio MCP
  server that Claude Code, Claude Desktop, Cursor and VS Code with Copilot can
  connect to (setup:
  https://aivideotranslator.app/docs/automation-mcp.html). An agent can list
  languages, providers, models and voices, prepare and start a translation,
  follow its progress, cancel it and search the app's Help. It can also list
  projects, inspect settings, speakers, transcripts, timing, verified audio
  status and source-video frames, and freely edit any registered project
  without opening it: segment text, timing, speakers, each speaker's voice and
  speed, and Even out acceleration (preview, then apply), with one shared undo
  step per batch. Free edits never start paid work. Reads and edits cover all
  registered projects accessible to the app, with no per-host/project permission.
  Explicit calls can retranslate or voice selected segments, retranscribe, or export an existing project. They start without GUI confirmation, may incur cloud charges, are not undoable, and never grant external-AI consent. Successful retranscription clears history; partial results remain after failure or cancellation. Paid work never automatically repeats after errors or restart. Export reports blocked row IDs without repairing them; the normal free-tier allowance warns while requireFullDub refuses an over-allowance export.
  Provider API keys, endpoints and provider configuration maps
  never leave the app, and a path typed in chat is not a file grant: the user
  selects the file or folder in the app with Browse. Compatibility with every
  MCP client is not promised. Shortcuts and Help guide:
  https://aivideotranslator.app/docs/apple-features-on-mac.html
- **MCP tools.** The server exposes 23 tools and the inert translate_video,
  edit_project and troubleshoot_job prompts. Editing proceeds through
  get_project, list_segments (selected fields and filters, including
  filter.timeMs) and optional get_frame inspection; reviewed free batches use
  edit_segments or edit_project, or even_out_acceleration (preview, then apply),
  each with the latest expectedRevision and a new requestKey. Verify surviving
  rows with list_segments and undo free changes through project_history.
  edit_project updates settings and speakers without processing; speed belongs
  to a speaker, so one line at another speed needs a separate speaker assigned
  with edit_segments, and source replacement stays in the app form.
  even_out_acceleration widens voiced segments into pauses using disk-measured
  audio; a preview writes nothing. Each mutating call has one shared undo entry
  and applied, notApplied and uncertain receipts. All replies include status,
  summary, data, warnings and next. The paid tools retranslate_segments,
  revoice_segments, retranscribe_project and export_project require an
  explicit user request, can cost money, start without GUI confirmation, are
  not undoable and must never be retried automatically. While MCP is on, the
  host can receive full transcripts, settings, frames as JPEG images and file
  paths under its own terms; provider keys remain private.
- **Languages.** 70+ are offered as a translation source or target — a
  conservative floor under what the app's own language registry holds. Transcription and
  text-to-speech cover fewer, and which ones depends on the chosen provider and
  voice; on-device coverage is narrower still.
- **Features.** Multi-speaker diarization, source separation (vocals vs
  background), a segment-by-segment editor with confidence scoring, voice
  emotion tags, per-project glossary and context hints, transcription progress
  tracking, and completed translation batches retained after an interrupted
  translation with failure and retry guidance. CPU / GPU Test compares local
  translation on Windows and in the macOS build. That Mac build uses
  Metal by default for every on-device translation model (Qwen3.5, Qwen3.6,
  Gemma 4, Hy-MT2 and Index-Translate), with CPU as an alternative and fallback;
  this does not describe the other on-device engines. On Windows, translation
  models use a discrete GPU through Vulkan, except Index-Translate, Gemma 4
  26B-A4B and Qwen3.6 35B-A3B, which run on the CPU.
  In the macOS build, Apple on-device speech recognition can also label speakers
  (speaker diarization), and the in-app Help can answer questions on-device
  where Apple Intelligence is available; Help can also be searched from Shortcuts
  and, on macOS 27 and later, by asking Siri.
- **Deliberately out of scope.** No lip-sync — the video track is never
  modified. No subtitle import or export in the current version. No voice
  cloning inside the app, though a clone already created on a provider account
  (e.g. an ElevenLabs custom voice) shows up as a selectable voice.
- **Privacy.** Source video is never uploaded to us and we run no server that
  could receive it. Cloud providers receive the audio or text required for the
  selected step. On-device AI keeps all media local. Optional pseudonymous
  usage statistics send only action names, fixed codes, enum identifiers,
  boolean state flags, counters and durations to Mixpanel after explicit consent.
  No account required. Siri on Mac opens Help with a query and receives help
  passages. Automation through Shortcuts on Mac and the local MCP server on
  Mac/Windows returns help passages, job status, file paths and the result file
  location to the requesting assistant/host. MCP project reads
  also return project context, glossary, terms, settings, speaker voice choices,
  full source/target transcripts, timing, audio status, free edit receipts and session history. Hosts can immediately edit any registered accessible project, with one shared undo step per changed batch. Free edits never start paid work. MCP returns path,
  MIME type, size and warnings for exports, without video bytes. The get_frame tool returns source-video frames as JPEG image bytes, with dimensions and requested time, without changing projects or history. Frames are returned to the requesting host and are excluded from logs, crash reports and usage statistics. Provider keys never cross
  either front. Existing file access, external-AI consent and free-tier export
  rules still apply; automation adds no premium gate. Local translation does
  not promise local processing by the assistant: its own terms govern what it
  receives and any subsequent reading or sharing of the result file.

## Pages

- [Home](https://aivideotranslator.app/): product overview, how the pipeline
  works, provider mixing, pricing, and the full FAQ.
- [Terms & Privacy](https://aivideotranslator.app/legal.html): End User Licence
  Agreement, privacy policy, support contacts, and the separate policy covering
  this website.
- Mac App Store listing: coming soon; the macOS build will be installed from it.
- [Microsoft Store listing](https://apps.microsoft.com/detail/9N6M49T2H4PP):
  where the Windows build is installed from.

## Help

Step-by-step help pages, one Markdown file per topic. Each page is also published as HTML at the same address with a `.html` extension, and states the platform and app version it was reviewed against.

- [Help index](https://aivideotranslator.app/docs/index.md): every help page, with a one-line summary.
- [Getting started](https://aivideotranslator.app/docs/getting-started.md): Install the app, answer the first-launch questions, then take one video from transcription through editing, translation and voices to an exported file, with the exact free-tier allowance.
- [Apple Intelligence unavailable](https://aivideotranslator.app/docs/apple-intelligence-unavailable.md): What to check when Apple Foundation Models reports that Apple Intelligence is not available on this Mac, and which other provider to use instead.
- [App basics](https://aivideotranslator.app/docs/basics-for-beginners.md): What to click first, what words like transcription, segment, model, provider and API key mean, whether it costs money, whether it is safe, whether you need the internet, how long it takes and where the finished video goes.
- [Providers and API keys](https://aivideotranslator.app/docs/providers-and-api-keys.md): Which AI services the app can use for transcription, translation and voices, which of them need your own API key, how to add, check, change or remove a key, how keys are stored, and how to use a local OpenAI-compatible server such as LM Studio or Ollama.
- [On-device models](https://aivideotranslator.app/docs/on-device-models-and-hardware.md): The downloadable on-device models for transcription, translation, voices and source separation with their approximate sizes, what the badges mean, memory and GPU use on Windows and Mac, the hardware compatibility test, where models are stored, and how downloads, resuming and deleting work.
- [Apple features](https://aivideotranslator.app/docs/apple-features-on-mac.md): How the Mac app uses Apple's on-device speech recognition, Apple Foundation Models and Apple Translation, and Apple system voices, which macOS versions they need, how to install missing languages and voices, what stays on your Mac, and how Help works with Apple Intelligence and Siri, and video translation works in Shortcuts.
- [Voices and speakers](https://aivideotranslator.app/docs/voices-and-speakers.md): How to give each speaker in a video their own voice, how automatic speaker detection (diarization) works, voice speed, voice tags, what happens when a translated line is longer than the original, and how to regenerate the voice of a single line.
- [Editing and translation](https://aivideotranslator.app/docs/editing-and-translation-quality.md): How to correct the transcript and translation, change segment timing, add, merge and delete segments, use undo and history, translate everything or only what is missing, and improve accuracy with the project context, glossary and terms fields.
- [Export and audio tracks](https://aivideotranslator.app/docs/export-and-audio-tracks.md): What Create Final Video produces, where the file is saved, output formats and audio codecs on Windows and Mac, adding the dub as a second audio track or replacing the original, keeping background music with source separation, and what the export checks before it starts.
- [Local MCP assistant](https://aivideotranslator.app/docs/automation-mcp.md): Connect an assistant, inspect and edit projects, explicitly translate, voice, retranscribe or export, track jobs, and understand costs, file access and consent.
- [Privacy and data](https://aivideotranslator.app/docs/privacy-and-data.md): What stays on your computer and what can leave it, how to use the app on-device only, usage statistics and crash reports and how to turn them off, where projects, models, settings and keys are stored, how to delete everything, and notes for company IT.
- [Subscription and free tier](https://aivideotranslator.app/docs/subscription-and-free-tier.md): Exactly what the free tier includes, what the subscription adds, plans and prices, trials, watermarks, commercial use, Windows versus Mac purchases, restoring and managing a subscription, and what to do when a purchase does not show as active.
- [Troubleshooting](https://aivideotranslator.app/docs/troubleshooting.md): What common messages mean and what to do, including FFmpeg on Windows, missing or moved files, disk space, memory, model downloads, network and certificate errors, no speech found, a slow computer, and how to send a report with the app log.
- [What's new](https://aivideotranslator.app/docs/whats-new.md): Changes by app version, as shown in the app’s What’s New dialog.
- [All help pages in one file](https://aivideotranslator.app/llms-full.txt): the full text of the pages above.

## Help versions

1. Ask the user for the app version: Settings → About App → App Version; use the part before `+` (e.g. `1.39.0+3` → `1.39.0`).
2. If that version is in the list below, read https://aivideotranslator.app/docs/<version>/llms-full.txt (single pages: https://aivideotranslator.app/docs/<version>/<page>.html or .md).
3. The latest version (1.39.0) is always at https://aivideotranslator.app/llms-full.txt and https://aivideotranslator.app/docs/.
4. If the user's version is older than every listed version, answer from the latest version and suggest updating the app (Mac App Store or Microsoft Store).
5. If it is newer than every listed version (a test build), use the latest version.
6. To see what changed between the user's version and the latest, read https://aivideotranslator.app/docs/whats-new.md (each version's own list is in https://aivideotranslator.app/docs/<version>/whats-new.md).

Release dates are listed separately for macOS and Windows. A platform marked not released has no release date yet.

[Version index (JSON)](https://aivideotranslator.app/docs/versions.json)

- [1.39.0](https://aivideotranslator.app/docs/1.39.0/) (latest) — macOS: not released; Windows: not released. [Full help](https://aivideotranslator.app/docs/1.39.0/llms-full.txt)

## Contact

- Support and general enquiries: ai.video.translators@gmail.com