Kaldi vs Whisper

Kaldi is a free, open-source C++ toolkit for speech recognition aimed at researchers and professionals. Its code and scripts cover methods including linear transforms, MMI, boosted MMI, MCE discriminative training and feature-space discriminative training. The toolkit integrates finite-state transducers through OpenFst and includes a matrix library that wraps standard BLAS and LAPACK routines. Example recipes support building speech recognition systems, and downloadable model archives cover speech recognition, speaker identification, speech activity detection, diarization and language modeling. The models may be used for any purpose. Kaldi is released under Apache License 2.0, which permits commercial use and redistribution. It runs on Linux distributions and Mac OS X, supports Android cross-compilation, and can be compiled to WebAssembly for browser execution. Windows compatibility is not actively maintained, and example scripts do not work there. Users obtain the current development version by cloning the GitHub repository and following its INSTALL file. Documentation is expert-oriented; the project directs help requests to its mailing list and bug reports or feature requests to GitHub issues.

Whisper is an open-source speech recognition model for transcription, speech translation, language identification, and voice activity detection. Its documented command-line and Python interfaces process audio files, and it can translate speech from multiple languages into English. The repository lists six model sizes—tiny, base, small, medium, large, and turbo—with different speed and memory tradeoffs; stated VRAM needs range from about 1 GB to 10 GB, depending on size. Installation is documented through pip and requires the ffmpeg command-line tool. Rust may be needed if a prebuilt tiktoken wheel is unavailable. The code and model weights are released under the MIT License, with local installation using Python and ffmpeg. Whisper is listed for Linux, macOS, Windows, and self-hosted deployment. OpenAI says it was trained on 680,000 hours of multilingual and multitask supervised data collected from the web. Performance varies widely by language, and it does not outperform models specialized for LibriSpeech. Turbo is not trained for translation; multilingual models are recommended for translating non-English speech into English.