Is Kaldi still used in 2026? An honest look

Short answer: yes — but "Kaldi" now means two different things, and only one of them is under active development. Classic Kaldi (the C++/bash toolkit with HMM-DNN models and WFST decoding) has slowed to a near-maintenance pace, yet it still runs in a huge number of production systems. Next-gen Kaldi (the k2 / lhotse / icefall / sherpa stack) is very much alive, shipping releases through 2026, and is one of the leading choices for streaming and on-device speech recognition.

If you're deciding whether to learn Kaldi, evaluate it for a project, or you just saw "Kaldi" in a job description and wondered if it's a red flag — here's the honest picture.

"Kaldi" is now two projects

Classic Kaldi (kaldi-asr/kaldi)

This is what most people mean by "Kaldi": the toolkit Daniel Povey and collaborators released in 2011, built around Gaussian mixture models, then DNN acoustic models (nnet3), the chain (LF-MMI) recipes, and weighted finite-state transducer (WFST) decoding. For most of the 2010s it was the default research and production ASR toolkit.

In 2026 the classic repository is still online and still builds, but active development has largely stopped. There is no modern end-to-end recipe, no transformer acoustic model in the core recipes, and the community's energy has moved elsewhere. It is not where you start a new project — but it is far from dead in the field. Telephony platforms, government and defense contractors, broadcast monitoring, and plenty of enterprises still run nnet3/chain models that were tuned years ago and work fine.

Next-gen Kaldi (the k2-fsa ecosystem)

This is the part that's alive. Led by Povey, "Next-gen Kaldi" (sometimes called Kaldi 2.0) is a set of PyTorch-native libraries:

• k2 — differentiable finite-state automata / FST operations that plug into PyTorch, so you can train with lattice-based losses.
• lhotse — data preparation and audio dataloading for modern deep-learning workflows.
• icefall — training recipes (transducer / RNN-T, CTC, Zipformer, etc.).
• sherpa, sherpa-onnx, sherpa-ncnn — deployment runtimes for servers, browsers, Android, iOS, Raspberry Pi, RISC-V and NPU accelerators, with no cloud dependency.

As of mid-2026 these repositories were seeing commits within the same week you might check them — icefall and k2 both had updates in July 2026. sherpa-onnx in particular has become a popular way to run offline speech-to-text, text-to-speech, VAD and speaker diarization on-device.

Where Kaldi still beats Whisper in 2026

Whisper (and models like wav2vec 2.0, MMS, and NVIDIA's Parakeet/Canary) won the "paste in an audio file, get a good transcript" use case. But there are jobs Whisper is simply not built for, and that's where the Kaldi lineage still wins:

Requirement	Why Kaldi / next-gen Kaldi fits
Real-time streaming, low latency	Whisper is a chunk-based encoder-decoder; true streaming is awkward. icefall's transducer models and sherpa are designed for it.
On-device / embedded	sherpa-onnx and sherpa-ncnn run small models on phones, Pi, and microcontrollers offline. Whisper large needs a GPU.
Lattices, word-level confidence, keyword spotting	WFST decoding gives you a lattice and calibrated confidence scores. Whisper gives you text (and sometimes made-up text on silence).
Custom vocabulary / pronunciation lexicons	Add domain terms, product names, or drug names via the lexicon and language model without retraining the acoustic model.
Forced alignment	Kaldi remains a go-to for aligning known transcripts to audio (subtitling, dataset creation, linguistics).
Telephony / 8 kHz narrowband	Decades of recipes and models tuned for call-center audio.
Strict on-prem, no cloud, auditable pipeline	Everything runs locally and is inspectable end to end.
Cost at very high volume	A tuned small model on your own hardware can be far cheaper per hour than GPU-hosted Whisper large.

Where Whisper (and friends) win

• Out-of-the-box multilingual accuracy with zero training.
• Speed to a prototype — pip install and you have a transcript in minutes.
• Punctuation, casing, and formatting for free.
• Ecosystem — faster-whisper, WhisperX, countless wrappers, hosted APIs, and fine-tuning tutorials.
• Hiring pool — far more engineers have Whisper experience than Kaldi experience.

For a deeper side-by-side, see our Kaldi vs Whisper vs Wav2Vec comparison.

A note on Vosk

If you've used Vosk for offline speech-to-text, you've used Kaldi — Vosk wraps Kaldi models behind a simple API in a dozen languages. It's still widely downloaded and is a common reason "Kaldi" quietly sits inside products whose developers never touched a Kaldi recipe.

What it means if "Kaldi" is in a job description

Seeing Kaldi in a 2026 JD usually signals one of a few things:

• Telephony, contact-center, or broadcast ASR with an established Kaldi pipeline to maintain and extend.
• Government, defense, or regulated industries that require fully on-prem, auditable systems.
• Embedded / edge voice teams — increasingly this means next-gen Kaldi (sherpa), not the classic toolkit.
• A legacy system someone needs to modernize — sometimes the real job is migrating off classic Kaldi.

Knowing WFST decoding, lattices, lexicons, and LF-MMI is still a genuine differentiator in interviews for these roles, because few newer engineers have it. If you want the fundamentals, start with our complete guide to Kaldi.

So should you learn it?

• Yes, learn the concepts (WFST, lattices, lexicons, forced alignment) — they transfer and they're rare.
• Yes, learn next-gen Kaldi if you're targeting streaming or on-device ASR roles — k2/icefall/sherpa is where that work happens.
• Don't sink weeks into classic Kaldi's nnet3/chain recipes unless a specific job requires maintaining them.
• Best combination for the job market: Whisper for breadth, next-gen Kaldi for streaming/edge, plus the classic-Kaldi vocabulary to work on legacy systems.

Bottom line: Kaldi in 2026 is not a museum piece. The classic toolkit is coasting but still deployed everywhere; the ecosystem's future is next-gen Kaldi, and it's a serious, actively developed option for the problems Whisper doesn't solve.