I would like to suggest integrating handy-computer/transcribe.cpp as an alternative local transcription backend for noScribe.
noScribes underlying Whisper-based transcription stack has some limitations, particularly regarding newer ASR models and GPU/platform support. transcribe.cpp has just been released and, as a C++ speech-to-text library based on ggml supports multiple model families through GGUF models. It currently supports, among others, Parakeet and Canary models, with CPU, CUDA, Vulkan and Metal backends. This would bring the advantages of more and faster models apart from whisper as well as improved support for GPU acceleration.
The project is also MIT licensed, which should make integration from a licensing perspective straightforward.
I would like to suggest integrating handy-computer/transcribe.cpp as an alternative local transcription backend for noScribe.
noScribes underlying Whisper-based transcription stack has some limitations, particularly regarding newer ASR models and GPU/platform support.
transcribe.cpphas just been released and, as a C++ speech-to-text library based on ggml supports multiple model families through GGUF models. It currently supports, among others, Parakeet and Canary models, with CPU, CUDA, Vulkan and Metal backends. This would bring the advantages of more and faster models apart from whisper as well as improved support for GPU acceleration.The project is also MIT licensed, which should make integration from a licensing perspective straightforward.