Dual-Model Pipeline Tuning Guide

Traditional AI subtitle tools suffer from an awkward trade-off: line-by-line translation is either too slow for live audio playback, or disjointed without context—causing character names and idioms to fluctuate wildly from scene to scene.

CoWatcher resolves this via an asynchronous Dual-Model Pipeline.

---

1. How It Works

[Platform Subtitle Stream]
           │
           ▼
 ┌───────────────────────────┐
 │   Translation Model       │ ◄── Enriched with Show Glossary & Memory
 └─────────────┬─────────────┘
               │ Lookahead batch streaming
               ▼
   [Live Player Subtitle DOM]
               │
   (Async Background Audit & Extraction)
               ▼
 ┌───────────────────────────┐
 │   Supervisor Model        │ ──► Audits consistency, maintains names/glossary
 └───────────────────────────┘

1. Translation Model:

   • Role: Ultra-fast, responsive lookahead batch translation.
   • Mission: Pre-translates upcoming cues 10–30 seconds ahead of the video playhead so that playback is 100% latency-free.
   • Recommended: translategemma:4b (Ollama) or deepseek/deepseek-chat / google/gemini-2.0-flash-001 (OpenRouter).

2. Supervisor Model:

   • Role: Reasoning, contextual summarization, and consistency auditing.
   • Mission: Asynchronously warms up show context, reconciles character aliases, maintains terminology consistency, and flags difficult slang.
   • Recommended: qwen2.5:7b (Ollama) or google/gemini-2.0-flash-001 / anthropic/claude-3.5-haiku (OpenRouter).

---

2. Tuning Performance Parameters

Tune these options in the Dual-Model Pipeline Performance Tuning section of the popup:

Batch Size

• Recommended: 5 – 10 cues
• Determines how many consecutive cues are grouped into a single LLM prompt. Balances syntactic coherence with prompt round-trip time.

Lookahead Window

• Recommended: 15 – 30 seconds
• Determines how far ahead of the current playhead the engine pre-fetches and translates subtitles into local storage.

Concurrency Limit

• Local Ollama (GPU): Set to 1. Prevents GPU VRAM exhaustion or serialization bottlenecks.
• Cloud OpenRouter: Set to 2 – 4. Takes advantage of serverless parallelism to dramatically speed up buffer filling.