What is Kaldi speech recognition?

Kaldi is a free, open-source toolkit for building speech recognition systems. It is not a model that you speak into and get text back from. You train or download a recognizer, then run it with Kaldi's tools. Its code is licensed under Apache 2.0 ¹, and the project says it is meant for researchers ². If you want to dictate, a finished model or app will serve you better. If you want to build your own recognizer, start with one of its recipes.

Our guide to open-source speech recognition places Kaldi among the older toolkits, next to the newer neural models. For the wider subject, see our overview of speech recognition.

Kaldi is a toolkit for building a recognizer, not a model to dictate with

The Kaldi project describes its toolkit as "a toolkit for speech recognition written in C++ and licensed under the Apache License v2.0" ¹. The word that matters is toolkit. Kaldi gives you the parts to train and run a recognizer, including the code that decodes audio into words. It does not give you a single trained model that works out of the box for your voice.

The project is also clear about who it is for. Kaldi is "intended for use by speech recognition researchers" ². It also says the toolkit will let you do many kinds of operations that do not make sense ³. That matters if you are looking for a simple dictation tool, and less if you have the time and skill to build one.

Kaldi is organized around recipes, and it dates from 2011

Kaldi's stated goal is to provide "complete recipes" for building a speech recognition system from widely available databases, such as those from the Linguistic Data Consortium ⁴. A recipe is a set of scripts that takes you from raw training data to a working recognizer.

The code was first released on 14 May 2011 and presented at ICASSP 2011 in Prague ⁵. That makes Kaldi older than the neural models that now dominate open-source speech recognition. Its age is not a reason to dismiss it, but it does mean that you should check whether a recipe still runs on your current system before you rely on it.

Kaldi ships few ready-made models, and other tools embed it

The project's home page says the model builds it offers are "not many" ⁶. So if you want to try Kaldi quickly, you will probably train or adapt a model yourself.

You will meet Kaldi more often inside other tools. A 2017 paper on the Montreal Forced Aligner, which times the words and sounds in a recording against its transcript, says it uses Kaldi "instead of HTK" ⁷. Home Assistant's Speech-to-Phrase tool was Kaldi-based up to version 1.4.3, and its developers say the rewrite now happens in another repository ⁸. When you meet a Kaldi-based tool, check its version, because the engine underneath can change between releases.

The license is permissive, but it covers the code only

The Apache 2.0 license lets you use and modify the Kaldi code, including in a product ¹. It does not vouch for the accuracy of any model you train with Kaldi, or for the data you train on. Check the license of each dataset and model separately before you ship anything.

This is general information, not legal advice. For a commercial product, have a lawyer confirm the terms.

Kaldi suits researchers and small, fixed jobs, not everyday dictation

Choose Kaldi when you need to control every stage of a recognizer, when you want to reproduce a published recipe, or when you are researching speech technology. Kaldi is also a reasonable choice for a narrow job, such as a small set of commands on a device you control, where you can build and test a recognizer around a short word list.

Choose a finished model or app when you want to dictate, transcribe a recording or caption a video. Those jobs depend on accuracy across many words and speakers, and a ready-made model, tested on your audio, is usually faster to a good result. The open-source speech recognition guide compares those models by license, language and speed.

If you are choosing between Kaldi-era toolkits and neural models, read what Vosk is and what NVIDIA Parakeet is. For the alternatives to Whisper, see whether there are better alternatives to Whisper.

What to do next

1. Decide whether you are building or using. If you only want to dictate or transcribe, skip Kaldi and start with a model from the open-source speech recognition guide.
2. Start from a recipe that matches your language and data. Check that the recipe runs on your system, and check the license of the data as well as the code.
3. Measure your results. Score each system by word error rate (WER) on the same hand-checked recordings. Use the same test set for every system you compare.
4. Check the version of any Kaldi-based tool before you depend on it, since the engine may have changed.
5. Plan the parts around the recognizer. A recognizer needs voice activity detection to know when someone is speaking, and a path from its output to the text you need. The guide to speech recognition in Python covers the code side.