Transcribe audio in your browser — offline, private, no API keys needed.

🎙️ Try it now → https://naklitechie.github.io/VoiceVault/

VoiceVault is part of the NakliTechie series of browser-native tools: single-file apps that run entirely on your device. No server. No API keys. No data leaving your machine.

• Multiple Input Methods — Drag & drop audio files, record directly from your microphone, or paste a URL to remote audio.
• Offline Transcription — Uses OpenAI's Whisper model running entirely in your browser via Transformers.js. Once cached, works completely offline.
• Multiple Export Formats — Export transcriptions as TXT, SRT (subtitles), VTT (web captions), or JSON (with timestamps).
• Audio Playback — Play back your original recordings directly from the transcription history.
• Retranscribe — Change language or model settings and re-run transcription on any saved recording.
• Transcription History — All transcriptions are stored locally in your browser with datestamped organization.
• Timestamped Segments — Every word is timestamped — useful for subtitles, meeting notes, or interview transcripts.
• Confidence Scores — Optional highlighting of low-confidence words so you know which parts to double-check.
• Batch Processing — Queue multiple audio files and transcribe them sequentially.
• Privacy First — No audio ever leaves your device. Toggle auto-delete to avoid storing recordings.

VoiceVault supports any audio format your browser can decode:

• MP3 — Most common audio format
• WAV — Uncompressed audio
• WebM — Default recording format for microphone input
• OGG — Open source audio format
• M4A — AAC audio (Safari/iOS)
• FLAC — Lossless audio

Both models are quantized (8-bit) for faster browser inference. The model is downloaded from Hugging Face on first use and cached indefinitely.

VoiceVault uses Transformers.js to run OpenAI's Whisper speech recognition model entirely in your browser:

1. First Load — The Whisper model (~250 MB for whisper-base) is downloaded from Hugging Face CDN and cached in your browser's IndexedDB.
2. Audio Processing — Your audio is decoded and resampled to 16kHz (Whisper's native sample rate) using the Web Audio API.
3. Transcription — The model processes the audio in 30-second chunks with overlapping windows to handle long recordings.
4. Storage — Transcriptions and optional audio recordings are stored in IndexedDB for offline access.
5. Export — Transcriptions can be exported in multiple formats, all generated client-side.

No audio data is ever sent to any server — not even to Hugging Face. The model runs entirely via WebAssembly in your browser.

No build step. No dependencies. Just open the file:

# Option 1: Open directly in browser
open voicevault.html

# Option 2: Serve with Python
python3 -m http.server 8000
# Then visit: http://localhost:8000/voicevault.html

# Option 3: Serve with Node.js (npx)
npx serve .

Coloured with memphis-03 · BACTERIO — Sottsass's 1978 squiggle pattern (the shape that visually anticipated audio waveforms). Cyan recording action, hot-pink mic, signal-yellow ready badge. The 80s mixing-desk energy translated to a transcription tool.

Palette pulled from Rangrez, the global colour-palette library that backs all NakliTechie projects.

Host on any static file server:

• GitHub Pages — Push to a GitHub repo and enable Pages
• Netlify — Drag and drop the HTML file
• Vercel — Connect your repo
• Local server — python3 -m http.server

The app is a single HTML file with no build process.

VoiceVault works in modern browsers that support:

• Web Audio API
• IndexedDB
• WebAssembly
• MediaRecorder API (for microphone recording)
• Transformers.js

Tested on:

• ✅ Chrome 120+
• ✅ Firefox 120+
• ✅ Safari 17+
• ✅ Edge 120+

• No server calls — All processing happens locally
• No analytics — No tracking, no telemetry
• No data collection — Your audio and transcriptions stay on your device
• Optional storage — Toggle auto-delete to prevent audio from being stored
• Clear anytime — Delete individual transcriptions or clear all history

• Audio length — Very long recordings (>1 hour) may take significant time to transcribe
• Speaker diarization — Cannot distinguish between different speakers (not yet supported in browser Whisper)
• Background noise — Whisper performs best with clear speech; heavy background noise may reduce accuracy
• First load time — Initial model download (~250 MB) takes time depending on your connection
• Memory usage — Large audio files require significant RAM for processing

"Why are you uploading that to a server?"

Meeting transcriptions, medical dictation, voice notes, interview recordings — these are exactly the audio files you don't want processed by a cloud API. OpenAI's Whisper API charges per minute and logs every request.

VoiceVault is the answer: world-class transcription with nothing leaving your device.

Built by Chirag Patnaik

MIT

Browser-native audio transcription using OpenAI's Whisper model via Transformers.js. No server-side components — everything runs locally in your browser.