Can NanumPlayer Translate Video Subtitles Fully Offline?

TLDR

Local video can have translated subtitles offline after the speech models and a local Ollama model are downloaded.
NanumPlayer requires 64-bit Windows 10 or 11 and an AVX2-capable CPU; NVIDIA acceleration is optional.
Subtitles are regenerated by default. Use explicit SRT generation when you want files to keep.

Local playback needs two prepared AI components

NanumPlayer is a Windows player that recognizes dialogue and displays a translated subtitle track. Speech recognition runs on the PC; translation uses an Ollama-compatible server. Download the required models first and use a local translation model, and a local video can be watched offline.

The creator introduced the project on GeekNews after finding foreign videos frustrating to watch without subtitles. The player is free and open source, but its installer is not a complete bundle of AI models. Ordinary playback can work before translated subtitles are ready.

The privacy documentation distinguishes local and cloud translation. Audio is not uploaded. Subtitle text goes to the configured server, with the title and audio languages used for glossary work. A local Ollama model keeps that processing on the PC; choosing a cloud model sends text to its provider.

Connect speech recognition and translation separately

The setup guide requires 64-bit Windows 10 or 11 and a CPU supporting AVX2. To prepare translated subtitles:

Open Ctrl+N → AI component manager and download a Whisper speech model suitable for the audio language.
Install and run Ollama, download a local translation model, then select it under Ctrl+N → Translation model.
Check the translation server address. The default local endpoint is http://127.0.0.1:11434.
Select the target subtitle language and open a file with dialogue. Changes to model, server or target language apply to the next file opened.

Interface language is separate from subtitle language. If the video plays without translation, inspect the model and server status before treating playback itself as broken. Ctrl+K opens detailed status information.

The input path also depends on subtitles already present. When the audio language is known and a matching text subtitle exists, the player can translate that text instead of recognizing speech. An existing target-language subtitle is selected by default without AI processing.

A supported CPU can run recognition; speed is a separate question

The FAQ says an NVIDIA GPU is optional. Automatic mode falls back to the included CPU engine when GPU recognition is unavailable. That does not guarantee that every local model will keep pace with playback.

At the beginning of a file, a default 30-second pre-roll gives subtitles a head start. Initial model loading or slow hardware can extend the wait. A translation model that does not fit GPU memory can time out while speech recognition runs alongside it.

Try smaller speech and translation models separately to identify the constraint. The documented CUDA path requires NVIDIA compute capability 8.6 or higher, a current driver and CUDA 13 cuBLAS; the project points users to RTX 30-series or newer hardware. CUDA 12 alone is insufficient. Video upscaling is a separate feature with different requirements.

Seeking can also make subtitles arrive late in a section not yet analyzed. Judge both timing and language quality: reducing model size changes the workload, but can also change how accurately dialogue is recognized or translated.

Generate SRT files when you want subtitles to persist

With default caching behavior, subtitles are generated again on the next viewing. Generate full subtitles creates files explicitly, without playing the video.

Open Ctrl+N → Generate full subtitles…, select a source, languages and destination, then start. Sources include speech, existing text subtitles and OCR of image subtitles. The output includes the original-language SRT and translated files for selected languages. A first-ten-minutes preview lets you inspect the result before processing the whole film. Web videos are not supported by this export feature.

Online playback has its own limits: the FAQ excludes DRM-protected video and sites that do not expose a media address. Offline subtitle generation should not be confused with access to every streaming service.

Begin with a short local file, inspect recognition alongside translation, and find a model combination that keeps up. Once the models are prepared, reopen that same file without a network connection to check the offline workflow you actually intend to use.