We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·17d ago·5 min read
Transcribe.cpp
TL;DR
Transcribe.cpp is a new transcription library built on ggml that supports the latest models from the handy-computerHF organization. It addresses the challenges of distributing cross-platform speech-to-text applications, which often rely on limited tools like whisper.cpp and ONNX.
✦ Why It Matters
Engineers can contribute to the development of Transcribe.cpp by reporting issues and suggesting improvements.
Key Takeaways
- Explore transcribe.cpp for efficient local speech recognition.
- Utilize the library's support for over 60 ASR models.
- Leverage acceleration options like Vulkan and CUDA for performance.
- Contribute to the library by reporting issues or adding bindings.
Full Summary
Transcribe.cpp is a transcription library designed to address the challenges of distributing cross-platform speech-to-text applications. Built on the ggml framework, it supports over 60 automatic speech recognition (ASR) models, all of which have been numerically validated and tested for word error rate (WER) accuracy.
The library offers acceleration through technologies like Vulkan, Metal, and CUDA, ensuring efficient performance on various hardware. It also features support for both streaming and batch transcription, making it versatile for different use cases.
The library is designed to be a drop-in replacement for whisper.cpp, maintaining compatibility with existing models. With bindings available in Python, JavaScript/TypeScript, Rust, and Objective-C/Swift, it aims to facilitate easy integration into applications.