systran/faster-whisper v1.2.1 is a speech recognition library that reimplements OpenAI Whisper models using CTranslate2. The latest release as listed on the GitHub repository is v1.2.1, and it is distributed as a Python package. The provided information does not state what changed in this release, so the specific modifications cannot be confirmed.
The project handles transcribing speech into text with Whisper models on top of CTranslate2. The repository description states that it supports large-batch processing and 8-bit computation. It is released under the MIT license. The README claims that under the same accuracy conditions it can be up to four times faster than the original Whisper while also reducing memory usage, and explains that quantization can further improve efficiency. The performance table lists processing time and memory usage by model and precision, batch size, and test hardware.
These performance figures should be read as the results shown for each condition in the README and the repository's claims; the same results are not guaranteed on other hardware or configurations. The documentation notes that using the GPU requires a separate NVIDIA library. Since audio processing uses PyAV, installing FFmpeg on the system is not required. Therefore, the speed and memory benefits should be judged by comparing the model, precision, and batch settings in the actual deployment environment.