Hey,

I developed VaultASR,a native C++ pipeline that does the entire speech-to-text + speaker diarization stack locally. My major goal has been to effectively utilize the hardware and run end-to-end on the machine locally avoiding any sensitive recordings/data go to cloud

What it does:

• Transcribes audio/video files with OpenAI's Whisper (via whsiper.cpp)
• Detects speech segments using Silero VAD v5 over ONNX Runtime
• Identifies who said what using WeSpeaker speaker embeddings + agglomerative clustering
• Outputs to Text, JSON, SRT, XLSX, Markdown, Docx, or SQLite

Performance on M1:

• Decoded 2 hours of audio in \~10 seconds
• Full transcription + diarization of that same 2h file in minutes
• Runs entirely on Metal GPU with no CPU bottleneck

Stack:

• C++17, CMake
• whisper.cpp (Whisper inference, Metal backend)
• ONNX Runtime (Silero VAD, WeSpeaker) with CoreML acceleration
• FFmpeg for decoding, libxlsxwriter for XLSX, RNNoise for denoising

Roadmap: Goal is to support other execution providers (CUDA (NVIDIA), DirectML (Windows), ROCm (AMD))

GitHub: https://github.com/vamshinr/vaultASR

would love the help extending this project to support other execution providers.