---

📖 Project Introduction

Disruptive Intelligent Subtitle Engine. Based on Multi-Agent Collaboration technology, it automates the entire process, delivering near-human level translation works.

✨ Highlights

• Unified Style: Generates a full-film style guide (tone/style/addressing strategy) first, avoiding the "two different translators" feeling.
• Consistent Terminology: Automatically generates and enforces a glossary, unifying names/places/proper nouns and pronoun gender (he/she).
• Context Disambiguation: Uses a sliding window to reference previous translations + future previews, reducing ambiguous references and sentence segmentation errors.
• Semantic Audit: Built-in review loop specifically checking for mistranslations/omissions/hallucinations; also adheres to enjambment protocols, not randomly completing half-sentences.
• More Human-Like: Enjambment + multi-line sliding translation, making the tone more coherent and reducing "machine translation flavor".
• Multi-Format Subtitles: Supports multiple common subtitle format inputs (auto-detection).
• Token Saving: Compact subtitle format, reducing Token consumption and improving processing efficiency.

Quick Start

Environment Requirements

• .NET SDK 10.0 (This project TargetFramework is net10.0)
• Available LLM API Key and Endpoint (customizable in the startup entry)

Run

Execute in the project root directory:

cd src
dotnet restore
dotnet run --project .\Subtitles.Translate.Agent\Subtitles.Translate.Agent.csproj

Interactive Use

After startup, the program will prompt sequentially:

• Enter local subtitle file path (supports drag and drop to terminal)
• Enter target language (Enter defaults to: Simplified Chinese)
• Enter API Key (prompts if not configured)

Output Files

• After translation, it will generate in the same directory as the original subtitle: OriginalFileName.<targetLanguage>.srt
• Defaults to writing monolingual translation (GenerateTranslatedSrt()); for bilingual subtitles, change the entry point to GenerateBilingualSrt()

Custom Models and Endpoints

Entry configuration is located at AgentSystemConfig initialization in Program.cs:L90-L105, where ModelId, Endpoint, ApiKey etc. can be modified.

// src/Subtitles.Translate.Agent/Program.cs
var systemConfig = new AgentSystemConfig();
systemConfig.AddDefaultConfig(new AgentConfig
{
ModelId = "gpt-oss-120b",
ApiKey = “apiKey”,
Endpoint = "https://<your-endpoint>/v1"
});

Modification example (using long-context models for Agent 1/2, lower cost model for translation):

var systemConfig = new AgentSystemConfig();
systemConfig.AddConfig(nameof(Step1_DirectorAgent), new AgentConfig
{
ModelId = "gemini-3-flash",
ApiKey = “apiKey”,
Endpoint = "https://<your-endpoint>/v1"
});
systemConfig.AddConfig(nameof(Step2_GlossaryAgent), new AgentConfig
{
ModelId = "gemini-3-flash",
ApiKey = “apiKey”,
Endpoint = "https://<your-endpoint>/v1"
});
systemConfig.AddConfig(nameof(Step3_TranslatorAgent), new AgentConfig
{
ModelId = "gpt-oss-120b",
ApiKey = “apiKey”,
Endpoint = "https://<your-endpoint>/v1"
});

Recommended Models (Cost/Context Trade-off)

• Agent 1 (Director) / Agent 2 (Glossary): Recommend gemini-3-flash (longer context, lower cost, better for global scanning and terminology extraction)
• Translation Stage (Translator): Can use lower cost models, e.g., gpt-oss-120b

1. Step1_DirectorAgent (Global Understanding / Style Guide)

• What it does: Reads the full subtitle first, generating a "full-film guidebook" for subsequent translations to strictly follow, solving style drift and addressing confusion.
• Input: Full subtitles (compact formatted), target language, and other request parameters.

2. Step2_GlossaryAgent (Terminology Extraction / Consistency Constraint)

• What it does: Based on Step 1’s strategy and full subtitles, extracts key entities and builds a controlled glossary, ensuring "one noun, one translation".
• Input: Step 1 Style Guide + Full Subtitles.
• Output: Character table (including aliases and gender inference), location table, terminology table (including domain and definition), and provides a Markdown version directly embeddable in prompts.

3. Step3_TranslatorAgent (Sliding Window Translation / Strong Format Validation)

• What it does: Translates subtitles in batches using a sliding window, referencing previous context and future previews to reduce segmentation, reference, and ambiguity errors.
• Input: Step 1 Style Guide + Step 2 Glossary + Previous Translated Context + Current Batch + Future Preview.
• Output: Line-by-line initial translation (forced to keep original IDs and quantity consistent); optionally triggers Step 4 for semantic audit before writing final draft.

4. Step4_ReviewerAgent [To Be Open Sourced] (Semantic Audit / Back-Translation Protocol)

• What it does: Only performs "audit-level" corrections for semantic accuracy, specifically checking for mistranslations, omissions, and hallucinations; no polishing, no terminology beautification.
• Input: Step 3 Initial Translation Batch.
• Output: Line-by-line PASS/FIXED, error reason (critique), and final adopted translation (final_translation), strictly aligned with Step 3 quantity.

5. Step5_PolisherAgent [To Be Open Sourced] (Terminology Compliance + Flow Polishing)

• What it does: Without breaking timeline cut points, first enforces terminology and pronoun corrections, then performs more native expression polishing and rhythm optimization.
• Input: Step 2 Glossary + Step 1 Style Guide + Current Batch Translation + Previous Polished Result (for coherent transition).
• Output: polished_text, optional note (slang/terminology explanation), optimization_tag (terminology correction/context polishing/style adaptation/no change).

6. Step6_TimingAdjusterAgent [To Be Open Sourced] (Reading Comfort Timeline Fine-tuning)

• What it does: Automatically extends end_time based on translation length and next sentence start time to improve readability; only changes end time, does not touch start time, overlap not allowed.
• Input: Translated text, original start/end, next sentence start (including 50ms safety buffer).
• Output: KEEP/EXTEND, adjusted_end, reason, and applies adjustments back to subtitle object.

📅 Open Source Plan

• February 2026: Open Step6_TimingAdjusterAgent
• March 2026: Develop Windows / macOS / Web UI

🙏 Acknowledgements

This project uses the following excellent open-source projects:

• Subtitle Edit (libse): Powerful subtitle editing and processing core library.
• Microsoft Agents: Foundation framework for building intelligent Agents.
• Mscc.GenerativeAI: Provides .NET support for Google Gemini models.