A privacy-first, local-speech subtitle generator engineered specifically for Hinglish, Indian English dialects, and code-mixed vernacular video workflows that mainstream speech tools mangle.
Dialect Intelligence
Standard speech-to-text models assume speakers stick to one language. Indian creators naturally switch between English and Hindi mid-sentence. SwaLekha is purpose-built to recognize and transcribe code-mixed speech without garbling terms.
Key Capabilities
Tuned with extensive phonetic dictionaries covering common vernacular phrases, slang, and rapid mid-sentence English-Hindi transitions that break conventional engines.
Switch seamlessly between Romanized Latin script ("Aap kaise hain?") and authentic Devanagari ("आप कैसे हैं?"), keeping millisecond timecodes perfectly synchronized.
Runs quantized Whisper models via WebAssembly and WebGPU directly on your machine. Zero audio is ever transmitted to cloud servers or remote APIs.
Interactive timeline block editor with draggable handle boundaries, quick typo corrections, and instant real-time video synchronization preview.
Export standard .srt and .vtt files for YouTube and Premiere, or render hardcoded subtitles directly into video with custom fonts, colors, and outlines.
Acoustic speaker change detection that automatically tags conversational turns between interviewers and guests on podcasts and talk shows.
Acoustic Architecture
| Acoustic Pipeline | Technical Method | Language / Dialect Focus | Privacy & Latency |
|---|---|---|---|
| Speech Model | Quantized Whisper model running in WebAssembly (Wasm + WebGPU) | English, Hindi, and Hinglish hybrid audio | Real-time factor: ~0.35x on Apple Silicon / modern x86 |
| Phonetic Mapping | Custom bilingual vocabulary table with code-mix normalization | Common tech terms, conversational Hindi loanwords | Zero external dictionary calls |
| Dual-Script Rendering | Bi-directional transliteration engine with context-aware casing | Romanized Hinglish & Devanagari Hindi | Instant sub-millisecond script switching |
| Timecode Alignment | Cross-attention weight inspection for word-level timestamps | Universal 10ms frame resolution | Exact start/end sync without manual dragging |
| Data Sovereignty | In-memory audio processing buffer | Completely offline after initial model cache | Zero network footprint |
Workflow
01 · Drop
Drop your MP4, MKV, or audio file directly into the local browser app.
02 · Transcribe
Watch local WebAssembly Whisper transcribe speech with code-mixing awareness.
03 · Review
Inspect Romanized vs Devanagari outputs and adjust any timecode boundaries.
04 · Export
Export standard SRT/VTT files or burn subtitles into your final video clip.