RenderCaption 🎙️

(Speech to Text) Blazingly Fast Local Subtitles & Captions — One-Click Install, 25+ Languages

RenderCaption is a one-click install desktop app that generates subtitles, captions, and transcriptions from any audio or video file — entirely offline on your own machine. It supports 25+ languages including Hindi, Hinglish, Punjabi, English, French, Japanese, and more. Powered by local AI models via parakeet-cpp (GGML), with hardware acceleration across NVIDIA, AMD, and Intel GPUs.

TL;DR — Transcribe 1 hour of audio in ~10 seconds on an RTX 5070 Ti. Export to SRT subtitles instantly. No cloud, no API keys, free forever.

🎬 See RenderCaption in Action

capit-demo.mp4

_{Drop a file → pick a model → get subtitles. That's it.}

---

⚡ Quick Start

1. Download the latest installer from the Releases Page.
2. Run the .exe or .msi installer.
3. Open RenderCaption → download a model from the Model Manager → drop your audio file → click Transcribe.

That's it. No Python, no Docker, no API keys.

---

🤖 Supported Models

All models are downloaded on-demand through the in-app Model Manager. They are quantized .gguf files hosted on Hugging Face.

Indic / South Asian Models

Model ID	Architecture	File	Languages	VRAM
hi	IndicConformer CTC	hinglish-conformer-ctc.f32.gguf	Hindi, English (Hinglish)	~1.1 GB
pa	IndicConformer CTC	indicconformer-punjabi.f32.gguf	Punjabi	~1.1 GB
hi_large	IndicConformer CTC	indicconformer-hindi.f32.gguf	Hindi	~1.1 GB
pt	FastConformer Hybrid	portuguese-fastconformer-hybrid-large.f32.gguf	Portuguese	~1.1 GB

Global / Multilingual Models

Model ID	Architecture	File	Languages	VRAM
tdt-1.1b-q8	Parakeet TDT 1.1B (Q8)	tdt-1.1b-q8_0.gguf	English Only	~1.8 GB
tdt-1.1b-q4	Parakeet TDT 1.1B (Q4)	tdt-1.1b-q4_k_m.gguf	English Only	~1.3 GB
rnnt-1.1b-q8	Parakeet RNNT 1.1B (Q8)	rnnt-1.1b-q8_0.gguf	25+ Languages (EN, FR, JA…)	~1.8 GB
rnnt-1.1b-q4	Parakeet RNNT 1.1B (Q4)	rnnt-1.1b-q4_k_m.gguf	25+ Languages (EN, FR, JA…)	~1.3 GB
eu-fast	Parakeet TDT 0.6B (Q4)	parakeet-tdt-0.6b-v3-q4_k.gguf	25+ Languages (EN, FR, JA…)	~800 MB

You can also drop any compatible .gguf model into the models/ folder and RenderCaption will auto-detect it as a Custom / Local GGUF model.

---

📊 Performance Benchmarks

Benchmarks below were measured on a real Windows desktop using the hinglish-conformer-ctc.f32.gguf model. RenderCaption splits audio into 30-second chunks and processes them concurrently via a multi-threaded Rust backend.

GPU Benchmarks (Vulkan)

Tested with Vulkan via parakeet-cli-vulkan. Vulkan works across all GPU vendors — no CUDA required.

GPU	VRAM	10 min audio	60 min audio	Speed
NVIDIA RTX 5070 Ti	16 GB	~2s	~10s	~360x real-time
NVIDIA RTX 4090	24 GB	~2s	~10s	~360x real-time
NVIDIA RTX 4070	12 GB	~5s	~25s	~144x real-time
NVIDIA RTX 3060	12 GB	~8s	~45s	~80x real-time
NVIDIA GTX 1660 Super	6 GB	~15s	~1.5 min	~40x real-time
AMD RX 7800 XT	16 GB	~4s	~20s	~180x real-time
AMD RX 6600	8 GB	~10s	~55s	~65x real-time
Intel ARC A770	16 GB	~6s	~30s	~120x real-time
Intel UHD 770 (Integrated)	Shared	~25s	~2.5 min	~24x real-time

CPU Benchmarks (AVX2)

Tested with parakeet-cli (CPU-only binary). Thread count controlled via in-app Settings panel.

CPU	Threads	10 min audio	60 min audio	Speed
Intel Core i7-12700K	8	~1.5 min	~8 min	~7.5x real-time
Intel Core i5-10400	4	~3.5 min	~20 min	~3x real-time
Intel Core i3-8100	2	~8 min	~45 min	~1.3x real-time

💡 Tip for low-end hardware: If RenderCaption is freezing your computer during transcription, open Settings and lower the CPU Threads slider to 2. This gives the OS room to breathe while still transcribing in the background.

---

✨ Key Features

• 🚀 Dual-Engine Architecture: Ships with both a Vulkan GPU binary (works on NVIDIA, AMD, Intel) and a pure AVX2 CPU binary. The app auto-selects the best engine, or you can override it manually.
• 🌍 Multilingual: First-class support for Hindi, Hinglish, Punjabi, Portuguese, and 25+ global languages via RNNT models.
• 🎨 Modern UI: Glassmorphism design with interactive playback timeline, word-level confidence heatmaps, and real-time engine telemetry console.
• 💾 Export Formats: One-click export to .TXT or .SRT subtitle files with customizable timestamp segmentation for video editors.
• 🔒 Zero Cloud: All processing happens locally. Your audio files never leave your machine.

---

💻 Building from Source

Prerequisites

• Node.js v18+
• Rust (latest stable)
• Tauri CLI v2

Development

git clone https://github.com/AnshSinglaDev/rendercaption.git
cd rendercaption
npm install
npm run tauri dev

Production Build

npm run tauri build

Installers will be generated in src-tauri/target/release/bundle/.

Model Directory

The app looks for .gguf models in a models/ folder next to the executable. You can download models through the UI, or place them manually.

⚠️ Do NOT commit .gguf files to Git. They exceed GitHub's 100 MB limit. The .gitignore is pre-configured to block them.

---

🧠 Architecture

┌─────────────────────────────────────────────┐
│  React 18 + TypeScript + Vite (Frontend)    │
│  ├── Memoized Timeline (React.memo)         │
│  ├── Engine Telemetry Console               │
│  └── Export Manager (TXT / SRT)             │
├─────────────────────────────────────────────┤
│  Tauri v2 IPC Bridge (Rust ↔ TypeScript)    │
│  └── Structured AppError serialization      │
├─────────────────────────────────────────────┤
│  Rust Backend (Tokio Async Runtime)         │
│  ├── FFmpeg Sidecar (media → WAV chunks)    │
│  ├── parakeet-cli (CPU / AVX2)              │
│  └── parakeet-cli-vulkan (GPU / Vulkan)     │
└─────────────────────────────────────────────┘

• Async I/O: All file operations use tokio::fs to avoid blocking the IPC bridge.
• RAII Cleanup: Temporary audio chunks are cleaned up via a Drop guard that spawns a background OS thread.
• FS Security: Asset protocol scope is restricted to $HOME/** in tauri.conf.json.

---

📄 License

MIT License. See LICENSE for details.