Skip to content

AI Subtitle Translation

WhisperJAV can translate Japanese subtitles to other languages using AI language models. Translation works as a standalone tool or integrated into the transcription pipeline.

Supported Providers

Provider Type API Key Required Best For
Ollama Local No Privacy, no cost, easy setup. Recommended for local use.
Local LLM Local No Legacy local option (llama-cpp). Consider Ollama instead.
DeepSeek Cloud Yes Cost-effective, good CJK quality
Gemini Cloud Yes Good multilingual support
Claude Cloud Yes High quality
GPT Cloud Yes Widely available
OpenRouter Cloud Yes Access to many models
GLM Cloud Yes Chinese-related tasks
Groq Cloud Yes Fast inference
Custom Cloud Varies Any OpenAI-compatible endpoint

Setting Up a Provider

Cloud Providers

  1. Get an API key from the provider's website
  2. In the GUI, select the provider from the dropdown
  3. Enter your API key in the field
  4. Click Test Connection to verify

Tip

API keys are saved locally and never sent anywhere except the provider's API endpoint.

Ollama is the easiest way to run local translation. Install Ollama, then:

# CLI: translate with Ollama (auto-detects GPU, picks best model for your VRAM)
whisperjav-translate -i subtitles.srt --provider ollama

# Use a specific model
whisperjav-translate -i subtitles.srt --provider ollama --model gemma3:12b

# List locally available Ollama models
whisperjav --list-ollama-models

OllamaManager auto-starts the server, detects your GPU, and recommends a model:

VRAM Recommended Model
CPU only qwen2.5:3b
8 GB qwen2.5:7b
12 GB gemma3:12b
16 GB+ qwen2.5:14b

Local LLM (Legacy)

The Local provider runs a llama-cpp server on your machine. No API key needed, but requires:

  • A GPU with ~8GB VRAM
  • The [llm] extra installed (pip install whisperjav[llm])
  • A GGUF model file (downloaded automatically on first use)

Note

Consider switching to Ollama — it's easier to set up, more reliable, and supports more models.

Translation Tone

Tone Description
Standard Clean, natural translations suitable for general audiences
Adult-Explicit Specialized instructions tuned for JAV dialogue with appropriate vocabulary

Two Ways to Translate

Method 1: Translate During Transcription

Transcribe and translate in one workflow:

  1. Set up your transcription on the Ensemble tab
  2. Check "AI-translate" after the merge strategy
  3. Select provider and model
  4. Click Start

Translation runs automatically after transcription completes.

Method 2: Translate an Existing SRT

Use the standalone translation tab:

  1. Go to AI SRT Translate (Tab 4)
  2. Add your .srt file
  3. Configure provider, model, target language, and tone
  4. Click Start

Advanced Settings

Setting Default What It Does
Movie Title (empty) Gives the AI context about the content
Actress Names (empty) Helps the AI handle character names correctly
Plot Summary (empty) Additional context for better translation
Scene Threshold 60 sec Groups subtitles into scenes for batch processing
Max Batch Size 30 Subtitles per API call (lower = fewer token issues)
Max Retries 3 Retry count for failed API calls

Improving Translation Quality

Filling in the Movie Title and Actress Names fields significantly improves translation quality. The AI uses this context to make better word choices and handle names consistently.

Batch Size Tuning for Local LLMs

Local LLMs have limited context windows compared to cloud APIs. If you see errors like "Hit API token limit" or "No matches found in translation text", your batch size is too large for your model's context window.

WhisperJAV auto-adjusts the batch size based on context window size, but you can also set it manually:

# CLI: set batch size explicitly
whisperjav-translate -i subtitles.srt --provider local --max-batch-size 10

# Or configure it persistently
whisperjav-translate --configure
# When prompted for "Max batch size", enter your preferred value

Recommended batch sizes by model context window:

Context Window Auto-Cap Recommended Manual Notes
8K (8192) 11 8–12 gemma-9b, small models
16K (16384) 27 20–27 Most mid-range models
32K+ 30 30 Large context models, cloud APIs

Note

The default batch size of 30 is designed for cloud APIs with 128K+ context windows. For local models, the auto-cap handles most cases automatically. Only set it manually if you still see token limit errors.

CLI Translation

# Translate with Ollama (local, recommended)
whisperjav-translate -i subtitles.srt --provider ollama

# Translate with DeepSeek (cloud)
whisperjav-translate -i subtitles.srt --provider deepseek --api-key YOUR_KEY

# Translate with adult tone
whisperjav-translate -i subtitles.srt --provider gemini --tone adult

# Translate to Portuguese
whisperjav-translate -i subtitles.srt --target-language portuguese

# Translate to Chinese
whisperjav-translate -i subtitles.srt --target-language Chinese

# Local LLM with reduced batch size (for 8K context models)
whisperjav-translate -i subtitles.srt --provider local --max-batch-size 10