Ajilamu

Aji Lhamu Mask Drama · Modern Precision Dubbing

Give a film another tongue, and keep its rhythm.

Ajilamu dubs creator video into another language. It holds every line inside the slot the original speaker left for it. The system measures each synthesized take against its slot, repairs misfits, and records every API cost.

The architecture comprises a Go server, a Svelte 5 workspace, and an append-only ledger in ClickHouse.

---

The Core Problem

Traditional dubbing either stretches dialogue awkwardly or lets speech drift across mouth movements. Ajilamu treats dialogue timing as a hard physical constraint. Fit is two-sided. A line that underruns breaks lip sync as severely as a line that overruns.

flowchart LR
  A[Source Video] --> B[Gemini 3.8 Flash
Watch & Segment]
  B --> C[Gemini 3.8 Flash
Translate to Slot]
  C --> D[Google Cloud Chirp 3 HD
Voice Synthesis]
  D --> E[ffprobe
Measure Signed Delta]
  E --> F{Fit Check
Within +8% or -5%?}
  F -- Yes: within budget --> G[ffmpeg atempo
Time Stretch]
  F -- No: beyond budget --> H[Gemini Rewrite
Shorter / Fuller]
  H --> D
  G --> I[ffmpeg Mux
Preserve Audio Bed]
  I --> J[Dubbed Video]

  classDef default fill:#090d15,stroke:#334155,color:#f1f5f9
  classDef highlight fill:#1e293b,stroke:#f59e0b,color:#f8fafc
  classDef accent fill:#082f49,stroke:#06b6d4,color:#38bdf8
  class B,C,D highlight
  class E,F,G,H accent

---

Core Capabilities

Two-Sided Fit Loop

Underrun and overrun both break lip sync. Ajilamu measures the signed delta of every take. It time-stretches up to 8 percent long or 5 percent short, and triggers targeted Gemini rewrites beyond those thresholds.

Immutable Ledger

Every take, edit, and charge appends to ClickHouse. The system stores waveform peak arrays directly in the database. Creators can inspect history and time-travel to earlier commits.

Lossless Audio Bed

Original background audio, music, and ambient noise survive everywhere outside dialogue slots. The export preserves source sample rate and channel layout.

Nanodollar Precision

The internal cost engine tracks charges in nanodollars. A 75-second sample dub costs $0.023414 across segmentation, translation, and synthesis calls.

---

Explore the Documentation

• Getting Started: Discover the origins of the name and the core philosophy.
• Quickstart: Clone the repository, configure credentials, and run your first dub locally.
• Pipeline Architecture: Read how video demux, speech synthesis, and audio muxing cooperate.
• Fit Loop: Study the mathematical repair matrix and signed delta thresholds.
• Immutable Ledger: Inspect the ClickHouse schema, DAG commits, and time-travel replay.
• Cost Accounting: Review itemized rate cards and billing verification.