Transform Audio & Video to Text. 100% Locally. 100% Private.
Convert your lectures, podcasts, tutorials, and meetings into timestamped transcripts, SRT subtitles, and structured JSON. Powered by faster-whisper for blazing-fast inference on CPU or GPU.
What's New in Voxcribe v2
Voxcribe started as a Bengali-only transcription script. Version 2 transforms it into a full-featured, lightning-fast multilingual transcription suite while keeping its privacy-first simplicity.
Voxcribe v1
The Original Prototype
- ✕OpenAI Whisper (Standard PyTorch inference, high memory overhead)
- ✕Bengali only (
language="bn"hardcoded) - ✕Plain text output only (single TXT file)
- ✕No timestamps for segments
- ✕Defaulted to `base` model (lower accuracy on complex dialect)
- ✕No Voice Activity Detection (hallucinations during silence)
Voxcribe v2.0
Production Ready & Faster
- faster-whisper (CTranslate2): 4x faster with up to 70% lower memory footprint
- Multilingual + Auto-detection: Bengali, English, Hindi, and 90+ languages
- 4 Output Formats: Timestamped TXT, Clean Plain TXT, Subtitle SRT, Structured JSON
- Per-Segment Timestamps: Precise start and end timestamps down to milliseconds
- `small` Model Default: High-accuracy balance optimized for Bangla + English mixed speech
- Silero VAD Built-in: Automatic voice filtering prevents silence hallucinations
Key Features Built for Real Workflows
Whether you are a student transcribing Bengali lectures, a developer creating subtitles, or a researcher building AI pipelines.
90+ Languages & Auto-Detect
Automatically detects the spoken language with high confidence scores or allows forcing specific languages like Bengali (bn), English (en), or Hindi (hi).
100% Privacy-First & Local
Zero telemetry, zero audio uploads, zero external API keys. All inference happens strictly on your local hardware.
CTranslate2 Fast Inference
Powered by faster-whisper with 8-bit quantization on CPU and float16 on NVIDIA CUDA GPUs for maximum efficiency.
Per-Segment Timestamps
Tracks precise start and end times for every spoken sentence, making it easy to seek to relevant sections in video lectures.
4 Rich Export Formats
Export into timestamped TXT, clean reading TXT, subtitle SRT for VLC/YouTube, and JSON for developer pipelines.
Voice Activity Detection (VAD)
Uses built-in Silero VAD to strip silent gaps, speeding up transcription and preventing repetitive hallucinations.
Mixed Language & "Banglish"
Seamlessly understands code-switching — preserving English technical terms (e.g. "React component", "API endpoint") inside Bengali speech naturally.
Broad Video & Audio Formats
Transcribes MP4, MKV, AVI, MOV, WebM, MP3, WAV, M4A, AAC, and more. FFmpeg automatically extracts high-quality audio.
Complete Offline Capability
After the initial one-time model download, Voxcribe works 100% offline without any internet connection.
Choose the Right Model for Your PC
Voxcribe supports all faster-whisper model tiers. Start with small (recommended) for optimal Bengali accuracy on standard laptops.
tiny
~75 MBQuick tests, voice memos
base
~145 MBLightweight CPU processing
small
~460 MBRecommended for Bengali & mixed English
medium
~1.5 GBComplex lectures, heavy background noise
large-v3
~3.0 GBStudio quality, subtle dialects, GPU
Installation & Interactive CLI Builder
Get up and running in 3 simple commands, or configure your exact command with our visual builder.
Clone the open-source repository to your local computer.
Create an isolated environment for dependencies.
Installs the global `voxcribe` command directly on your machine.
Visual Command Builder
Output Tailored to Every Need
Inspect the exact files Voxcribe generates. Pick timestamped text for revision, plain text for AI chats, SRT for video subtitles, or JSON for developers.
[00:00:03] আসসালামু আলাইকুম সবাইকে। [00:00:08] আজ আমরা React-এর component architecture নিয়ে কথা বলব। [00:00:15] প্রথম গুরুত্বপূর্ণ বিষয়টি হলো local state management...
Your Voice and Media Never Leave Your Machine
Unlike cloud-based speech APIs that require monthly subscriptions and upload your private voice recordings to foreign servers, Voxcribe executes every mathematical tensor locally on your CPU or GPU.
Built for the Global Community
Voxcribe is a free, MIT-licensed open-source project dedicated to making speech recognition accessible, private, and fast for everyone.
Join & Support the Project
Whether you want to contribute test audio samples, report bugs, improve documentation, or star the repository on GitHub — every contribution helps make AI more accessible.