Loki Studio v1.1.0

← All updates

NEW FEATURES

### Translated Metadata Generation

• Metadata Tab now supports generating titles, descriptions, and tags in any of the 35 supported languages
• Translation dropdown in Metadata Tab lets you select target language
• Generated metadata is automatically translated using NLLB-200
• Translated metadata is saved alongside original metadata
• Perfect for creators targeting international audiences

### Translated Metadata Upload to YouTube

• Upload Tab now uploads translated metadata to YouTube
• When translated metadata exists, it's automatically included in the upload
• YouTube will display localized titles/descriptions based on viewer's language
• Supports all YouTube-supported languages

### NLLB Translation Model Selection

• Added support for all 3 NLLB-200 model sizes in Application Settings: * NLLB-200 Distilled 600M (~2.5 GB) - Fast, good for high-resource languages * NLLB-200 Distilled 1.3B (~5.5 GB) - Balanced quality/speed (Recommended) * NLLB-200 3.3B (~13 GB) - Best quality, slower
• New Translation section in Application Settings with: * Model size selector dropdown * Download button with progress indicator * Model description and size information
• Models are downloaded to %LOCALAPPDATA%\LokiStudio\Models\nllb\

IMPROVEMENTS

### Translation Quality

Based on NLLB-200 research, implemented key quality improvements:

1. Sentence-Level Translation

• NLLB-200 is trained on sentence-aligned parallel data - Input text is now automatically split into sentences before translation - Each sentence is translated individually, then rejoined - Significantly improves translation quality for multi-sentence input

2. Anti-Hallucination Protection

• Changed no_repeat_ngram_size default from 0 to 3 - Prevents repetition loops like "the the the" or repeated phrases - Well-documented technique from CTranslate2/fairseq community

3. Sentence Splitting Heuristics

• Splits on standard punctuation: . ! ? - Splits on CJK sentence-ending punctuation - Avoids splitting on abbreviations: Dr., Mr., Ms., vs., Jr., Sr., St., No. - Avoids splitting on decimals: "3.5" stays intact

### Caption Display - One Word Pop Style

• Fixed: Last word no longer hangs awkwardly until next spoken segment
• Added 0.8 second maximum display duration cap for any single word
• Non-last words are bounded by the next word's start time
• Fix applies to both live preview and exported ASS caption files

SUPPORTED TRANSLATION LANGUAGES (35)

Latin Script: English, Spanish, French, German, Italian, Portuguese, Dutch,

Polish, Romanian, Swedish, Danish, Norwegian, Finnish, Czech,

Hungarian, Turkish, Vietnamese, Indonesian, Malay

Cyrillic: Russian, Ukrainian, Bulgarian, Belarusian

Asian: Chinese (Simplified), Japanese, Korean, Thai

Other Scripts: Greek, Arabic, Persian, Hebrew, Hindi, Bengali, Tamil, Urdu

TECHNICAL DETAILS

### Files Modified

• loki_studio-cpp/src/tabs/MetadataTab.cpp * Added translation language selector * Integrated NLLB translation into metadata generation pipeline * Save translated metadata to metadataXX.json files
• loki_studio-cpp/src/tabs/UploadTab.cpp * Load and upload translated metadata to YouTube * Support for YouTube's localized metadata API
• loki_studio-cpp/src/services/CaptionService.cpp * formatOneWordPop(): Added maxWordDisplaySec = 0.8f cap * generateAssFile(): Added maxWordDisplayMs = 800 cap for ASS export
• Muninn-Faster-Whisper/include/muninn/translator.h * TranslationOptions::no_repeat_ngram_size default changed to 3
• Muninn-Faster-Whisper/src/translator.cpp * Added split_sentences() function for sentence-level translation * Added join_sentences() function for rejoining translated sentences * Modified translate_batch() to use sentence splitting
• loki_studio-cpp/src/core/AppSettingsManager.h/cpp * Added translationModel setting
• loki_studio-cpp/src/ui/views/ApplicationSettingsView.qml * Added Translation section with model selector
• loki_studio-cpp/src/ui/views/MetadataView.qml * Added translation language dropdown
• loki_studio-cpp/src/tabs/ApplicationSettingsTab.cpp * Added onDownloadTranslationModelClicked() handler * Added onTranslationModelChanged() handler

DOCUMENTATION

• Created Docs/NLLB_Translation_Technical_Guide.md * Comprehensive documentation of NLLB-200 research findings * Best practices for translation quality * Language code mappings * Known issues and solutions

KNOWN ISSUES

• Translation model download requires stable internet connection
• Large models (1.3B, 3.3B) require significant disk space and VRAM
• First translation after model load may take a few seconds to initialize

DEPENDENCIES

No new dependencies added. This release uses the existing:

• CTranslate2 for NLLB inference
• SentencePiece for tokenization

UPGRADE NOTES

• Existing installations will continue using 600M model by default
• Users can upgrade to larger models via Application Settings > Translation
• Translation settings are preserved during upgrade
• Existing metadata files remain compatible

← All updates  ·  Source-of-truth changelog on GitHub