Mozilla.ai has officially released transcribe.cpp, an open-source C/C++ speech-to-text (STT) inference library. This new tool promises portable, GPU-accelerated support for a range of STT models, simplifying the integration of fast, local transcription into applications.
Developed under Mozilla.ai's Builders in Residence (BiR) program, transcribe.cpp leverages the ggml runtime. This allows it to support diverse STT model families via GGUF, with inference accelerated by Metal, Vulkan, and CUDA backends.
Solving Portability and Performance
The library addresses common weaknesses in existing STT models: poor portability and suboptimal performance. Many models are developed in isolation, leading to platform-specific limitations (e.g., MLX models on Macs only) and inconsistent acceleration.
transcribe.cpp provides a uniform interface, bringing GPU acceleration to these disparate models. This makes the library valuable not just for its creator CJ Pais's desktop application, Handy, but for any developer seeking to integrate robust STT functionalities.
This release marks a significant milestone as the first independent open-source project supported by Mozilla.ai's BiR program. The program aims to advance applied, cutting-edge research while aligning with Mozilla's broader roadmap.