dubweave
The Problem
Streaming platforms frequently have content available only in English, with no dubbed or subtitled alternative. Cloud dubbing services are expensive, slow, and require sending audio to third parties.
The Solution
dubweave is a fully local AI dubbing pipeline: download → transcribe → translate → synthesize → assemble. Built with an empirical approach: every model decision is validated against measurable output quality.
Technical Highlights
- Whisper ASR for transcription with high accuracy across accents
- NLLB-200 / LLM translation with 8-utterance lookback for context-aware output
- Multi-model TTS: Kokoro, XTTS, Google, ElevenLabs, Gemini, Supertonic; switchable via config
- PT-PT → PT-BR normalizer: 36 regex rules, validated against 50-sentence corpus with 0 false positives
- Empirical calibration: measured 15.1 chars/sec actual vs 18.0 assumed, fixing 70% systematic overtrimming
- 4 completed autoresearch loops with KEEP/DISCARD decisions logged in TSV
- Pipeline resumability: interrupted runs continue from last checkpoint
Related Reading
Building an internal AI assistant for support engineers with MCP — Another MCP-based internal tool project that shares a similar philosophy of keeping data local and giving engineers direct access to the systems they trust.