Kaldi
Established speech recognition toolkit used in research and production systems.
About
Written in C++, Kaldi has been a foundation of automatic speech recognition research and production systems for well over a decade. The toolkit provides the building blocks for complete recognition pipelines: feature extraction, acoustic modeling spanning GMM and neural network approaches, and decoding based on finite state transducers through OpenFst, with linear algebra handled by LAPACK, OpenBLAS, or ATLAS. Extensive example recipes for standard corpora ship in the repository and have served as starting points for many research papers and commercial systems. Kaldi runs on Linux, macOS, Cygwin, and Windows, and can be cross compiled for Android and WebAssembly. It predates the current wave of end to end speech models and demands more setup and expertise than pip installable alternatives, but remains in active use across academia, industry, and teaching. Released under the Apache 2.0 license.
Should you use Kaldi?
Pick it when
Pick Kaldi when you work with classic hybrid ASR, need FST-based decoding with custom grammars and lexicons, or build on its long recipe history for research, teaching, or an existing production system.
Look elsewhere when
Skip it for new end-to-end work: it predates modern models and demands C++ builds and deep expertise. Sherpa-ONNX, its k2-fsa successor, WeNet, or Whisper reach accurate transcripts with far less effort.
Alternatives to Kaldi
- Sherpa-ONNX STT
Modern successor from the k2-fsa project with prebuilt models for mobile, embedded, and WebAssembly, far less setup than compiling Kaldi.
- Vosk
Packages compact offline models with bindings in many languages and runtime vocabularies, much easier than building Kaldi recipes yourself.
- ESPnet
End-to-end research toolkit that keeps Kaldi-style data conventions, the natural move to Conformer or transducer models from GMM and FST pipelines.
- WeNet
Production end-to-end toolkit with streaming runtimes, replacing Kaldi's multi-stage pipeline with a single Conformer model.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Expert (5/5)
- License
- Apache-2.0
- Added
- Apr 3, 2026