Let me introduce to you Kaldi, a speech-to-text engine. I put the link right into models directory, to save you some time.
https://github.com/kaldi-asr/kaldi/tree/master/egs
There are a bunch of shell, Perl and Python scripts there, with some awk in between. These scripts are often copied almost verbatim between models and this, believe me, can lead to all sorts of errors.
The running joke around working with these scripts was "all these .sh should be .hs," i.e., these scripts should be implemented in Haskell.
In A Tour of Hell – Shell scripting Haskell dialectkaldi-asr
kaldi
Shell2015–2024
kaldi-asr / kaldi
kaldi
kaldi-asr/kaldi is the official location of the Kaldi project.
20Distinct people
29Mentions
0Explicit recommenders
15.5KGitHub stars
What to know before exploring
What it is
kaldi-asr/kaldi is the official location of the Kaldi project.
GitHub
Where it is used
What people are saying
No recommendation phrase was found in the indexed comments.
No classified warning is attached to this page. Read the discussions before treating silence as approval.
Also in these discussions
- alphacep/vosk-api · linked in 3 of the same discussions
- espnet/espnet · linked in 3 of the same discussions
- flashlight/wav2letter · linked in 3 of the same discussions
Mentioned in the same Hacker News threads. That is not proof they are substitutes.
This week on Hacker News
Nov 7–Nov 14 UTC · Direct repository links and separately labeled related project stories. Indexed through 11/14/2026, 3:56:07 PM UTC.
No direct repository links were found in the indexed HN discussions over the past seven days. Older discussions are available below.
Mentions over time
HN mentions by year29 dated mentions
Observed mentions: 2015-08-22 18:56:08 UTC to 2024-10-03 06:47:04 UTC.
Counts come from the repository’s dated source evidence. Index cutoff: 2026-11-02 08:16:59 UTC.
View mention counts and date ranges
| Period | Start (UTC) | End (exclusive, UTC) | Mentions | Submissions | Comments |
|---|---|---|---|---|---|
| 2015 | 2015-01-01 00:00:00 UTC | 2016-01-01 00:00:00 UTC | 2 | 2 | 0 |
| 2016 | 2016-01-01 00:00:00 UTC | 2017-01-01 00:00:00 UTC | 5 | 1 | 4 |
| 2017 | 2017-01-01 00:00:00 UTC | 2018-01-01 00:00:00 UTC | 7 | 0 | 7 |
| 2018 | 2018-01-01 00:00:00 UTC | 2019-01-01 00:00:00 UTC | 4 | 1 | 3 |
| 2019 | 2019-01-01 00:00:00 UTC | 2020-01-01 00:00:00 UTC | 1 | 0 | 1 |
| 2020 | 2020-01-01 00:00:00 UTC | 2021-01-01 00:00:00 UTC | 7 | 0 | 7 |
| 2021 | 2021-01-01 00:00:00 UTC | 2022-01-01 00:00:00 UTC | 0 | 0 | 0 |
| 2022 | 2022-01-01 00:00:00 UTC | 2023-01-01 00:00:00 UTC | 2 | 1 | 1 |
| 2023 | 2023-01-01 00:00:00 UTC | 2024-01-01 00:00:00 UTC | 0 | 0 | 0 |
| 2024 | 2024-01-01 00:00:00 UTC | 2025-01-01 00:00:00 UTC | 1 | 0 | 1 |
From the project README
An excerpt from the repository, fetched Nov 12, 2026.
Read the project’s introduction
Kaldi Speech Recognition Toolkit
To build the toolkit: see ./INSTALL. These instructions are valid for UNIX
systems including various flavors of Linux; Darwin; and Cygwin (has not been
tested on more "exotic" varieties of UNIX). For Windows installation
instructions (excluding Cygwin), see windows/INSTALL.
To run the example system builds, see egs/README.txt
If you encounter problems (and you probably will), please do not hesitate to
contact the developers (see below). In addition to specific questions, please
let us know if there are specific aspects of the project that you feel could be
improved, that you find confusing, etc., and which missing features you most
wish it had.
Kaldi information channels
For HOT news about Kaldi see the project site.
Documentation of Kaldi:
Info about the project, description of techniques, tutorial for C++ coding.
Doxygen reference of the C++ code.
Kaldi forums and mailing lists:
We have two different lists
User list kaldi-help
Developer list kaldi-developers:
To sign up to any of those mailing lists, go to
http://kaldi-asr.org/forums.html:
Development pattern for contributors
Create a personal fork
of the main Kaldi repository in GitHub.
Make your changes in a named branch different from master, e.g. you create
a branch my-awesome-feature.
Generate a pull request
through the Web interface of GitHub.
As a general rule, please follow Google C++ Style Guide.
There are a few exceptions in Kaldi.
You can use the Google's cpplint.py
to verify that your code is free of basic mistakes.
Platform specific notes
Fedora 41 (and later)
In order to build it on Fedora 41 using the libraries that are provided by the distro, you need to install the development libraries and dependencies with
sudo dnf install lapack-devel openfst-devel
then build the package as follows:
cmake -S ./ -Bbuild/Release -DFETCHCONTENT_FULLY_DISCONNECTED=ON -DBuildForFedora=ON
cmake --build /home/gerhard/workspace/kaldi/build/Release
PowerPC 64bits little-endian (ppc64le)
Kaldi is expected to work out of the box in RHEL >= 7 and Ubuntu >= 16.04 with
OpenBLAS, ATLAS, or CUDA.
CUDA drivers for ppc64le can be found at https://developer.nvidia.com/cuda-downloads.
An IBM Redbook is
available as a guide to install and configure CUDA.
Android
Kaldi supports cross compiling for Android using Android NDK, clang++ and
OpenBLAS.
See this blog post
for details.
Web Assembly
Kaldi supports cross compiling for Web Assembly for in-browser execution
using emscripten and OpenBLAS
See this repo
for a step-by-step description of the build process.
Read the full README on GitHubWhat Hacker News says
29 sourcesKaldi Speech Recognition Toolkit
5 points · Open discussionVosk-api isn't an SST engine itself, it is built using the Kaldi speech recognition toolkit (https://github.com/kaldi-asr/kaldi) and nicely implements and packages an API for Kaldi chain/LF-MMI models.
In Nerd-dictation, hackable speech to text on LinuxNot a big problem given so many alternatives around like
E.g. some very active projects are:
* Kaldi (https://github.com/kaldi-asr/kaldi/) obviously, probably the most famous one, and most mature one. For standard hybrid NN-HMM models and also all their more recent lattice-free MMI (LF-MMI) models / training procedure. This is also heavily used in industry (not just research).
* ESPnet (https://github.com/espnet/espnet), for all kind of end-to-end models, like CTC, attention-based encoder-decoder (including Transformer), and transducer models.
* Espresso (https://github.com/freewym/espresso).
* Google Lingvo (https://github.com/tensorflow/lingvo). This is the open source release of Googles internal ASR system, and used by Google in production (their internal version of it, which is not too much different).
* NVIDIA OpenSeq2Seq (https://gi
Read the rest
thub.com/NVIDIA/OpenSeq2Seq).
* Facebook Fairseq (https://github.com/pytorch/fairseq). Attention-based encoder-decoder models mostly.
* Facebook wav2letter (https://github.com/facebookresearch/wav2letter). ASG model/training.
* Vosk (https://github.com/alphacep/vosk-api). Offline lightweight speech recognition API with support for 10 languages.
And there are much more.
There are a lot of open source projects in this space. DeepSpeech is actually one of the outsiders (they are not represented well in the academic community), and also not quite competitive to other software (at least last time I checked).
E.g. some very active projects are:
* Kaldi (https://github.com/kaldi-asr/kaldi/) obviously, probably the most famous one, and most mature one. For standard hybrid NN-HMM models and also all their more recent lattice-free MMI (LF-MMI) models / training procedure. This is also heavily used in industry (not just research).
* ESPnet (https://github.com/espnet/espnet), for all kind of end-to-end models, like CTC, attention-based encoder-decoder (including Transformer), and transducer models.
* Espresso (https://github.com/freewym/espresso).
* Google Lingvo (https://github.com/tensorflow/lingvo). This is the
Read the rest
open source release of Googles internal ASR system, and used by Google in production (their internal version of it, which is not too much different).
* NVIDIA OpenSeq2Seq (https://github.com/NVIDIA/OpenSeq2Seq).
* Facebook Fairseq (https://github.com/pytorch/fairseq). Attention-based encoder-decoder models mostly.
* Facebook wav2letter (https://github.com/facebookresearch/wav2letter). ASG model/training.
* (RETURNN (https://github.com/rwth-i6/returnn) and RASR (https://github.com/rwth-i6/rasr), our own, although this is currently free for academic use only. It is used in production as well. Supports hybrid NN-HMM, CTC, end-to-end attention-based encoder-decoder, transducer, etc.)
And there are much more.
You will also find lots of ready-to-use trained models.
Try https://github.com/alphacep/vosk-api. It supports 10 languages, works on Android and RPi and also has big and more accurate server models.
Other good ones are https://github.com/daanzu/kaldi-active-grammar and https://talonvoice.com/
There are toolkits for research like https://github.com/kaldi-asr/kaldi, https://github.com/espnet/espnet, wav2letter, Espresso, Nvidia/Nemo, https://github.com/didi/athena. You can try them too if you want to go deep. Some of them have interesting capabilities.
In Future of DeepSpeech / STT after recent changes at MozillaShameless plug, but I have been working on an open source IDE plugin [1] for the IntelliJ Platform which attempts to do this. Previously, we used an older HMM-based speech toolkit called CMUSphinx [2], but are currently transitioning to a deep speech recognition system. We also tried a number of cloud APIs including Amazon Lex and Google Cloud Speech, but they were too slow -- offline STT is really important for low latency UX applications. For navigation and voice typing, we need something customizable and fairly responsive. Custom grammars would be nice for various contexts and programming languages.
There are a few good OSS offline deep speech libraries including Mozilla DeepSpeech [3], but their resource footprint is too high. We settled on the currently less mature vosk [4], which is based on Kaldi [5] (a more popular deep speech pip
Read the rest
eline), and includes a number of low-footprint, pretrained language models for real-time streaming inference. Research has shown how to deploy efficient deep speech models on CPUs [6], so we're hoping those gains will translate to faster performance on commodity laptops soon. You can follow this issue [7] for updates on our progress. Contributions are welcome!
[1]: https://github.com/OpenASR/idear/
[2]: https://cmusphinx.github.io/
[3]: https://github.com/mozilla/DeepSpeech
[4]: https://github.com/alphacep/vosk-api
[5]: https://github.com/kaldi-asr/kaldi
[6]: https://ai.facebook.com/blog/a-highly-efficient-real-time-te...
[7]: https://github.com/OpenASR/idear/issues/52
IMHO the speech dataset list is missing other interesting free corpora, e.g. the TEDlium dataset, Voxforge, Common Voice. A more comprehensive (but not complete) list can be found here: https://github.com/kaldi-asr/kaldi/tree/master/egs (download links can be found in the scripts)
In Audio Datasets for Machine LearningI was assuming that this is their main competitive moat against Alexa, etc.
I.e. that they avoid sending any audio to some central server.
It is possible:
https://github.com/kaldi-asr/kaldi/issues/3571
In Our first patent trollDid you miss wav2letter achieving a new 2019 SOTA on librispeech? Both projects have a _lot_ of architectures and papers in play, I know it can be hard to keep up.
Aren’t some of the peak Kaldi numbers also with _huge_ language models that aren’t so practical to deploy? The best kaldi results I’ve found here both use “fglarge” LMs, but a quick search didn’t tell me how big that actually is.
The best I can find for the test set are (from main RESULTS):
test-clean: 4.31% WER
test-other: 10.62% WER
From here (not sure why buried): https://github.com/kaldi-asr/kaldi/blob/master/egs/librispee...
test-clean: 3.80% WER
test-other: 8.76% WER
This is the 2019 wav2letter SOTA:
https://github.com/facebookresearch/wav2letter/blob/master/r...
Sure, one of their techniques is to semi supervised train on 60k hours of audio, so I’ll pos
Read the rest
t their best numbers both with and without that step (as you said “if kaldi is trained on the same data”)
Without 60k unsupervised audio:
test-clean: 2.31% WER
test-other: 5.18% WER
With 60k audio:
test-clean: 2.03% WER
test-other: 4.11% WER
That’s so far ahead of kaldi’s librispeech RESULTS, let’s go up the list and look at their acoustic model tests that don’t even bother with a language model:
Without 60k or language model:
(test-clean / test-other WER)
3.05 / 7.01
With 60k without language model:
2.30 / 5.29
Am I missing a kaldi whitepaper here?
I found “State-of-the-Art Speech Recognition Using Multi-Stream Self-Attention With Dilated 1D Convolutions” on wer_are_we claiming 2.2 / 5.8 using Kaldi as a base, which is impressive, but they changed the network architecture and I didn’t find this in any of the open-source kaldi librispeech subdirectories.
(If arbitrary model changes count, at some point you might as well treat both wav2letter and kaldi as “tensorflow with some speech helpers” because I’m sure you can implement either framework’s best neural architecture in the other framework.)
If there’s a better paper on kaldi than 2.2 / 5.8 you should comment with it and PR it to wer_are_we, because wav2letter is currently beating every other LS result posted there.
I think its on purpose to not give too much advantage to downloaders, there was a discussion long time ago, but they still continue with mp3. There are also other not so nice points: https://github.com/kaldi-asr/kaldi/issues/2141
In Mozilla releases the largest to-date public domain transcribed voice datasetEven more so if they compared it to Kaldi TDNN-LSTM with RNNLM lattice rescoring (test-clean 3.22%, apparently): https://github.com/kaldi-asr/kaldi/blob/master/egs/librispee...
In Facebook open-sources a speech-recognition system and a machine learning libraryKaldi-asr/kaldi: This is now the official location of the Kaldi project
1 points · Open discussionLooks like Kaldi gets 4.27% WER on CV:
https://github.com/kaldi-asr/kaldi/blob/master/egs/commonvoi...
But that's likely trained exclusively on CV audio and using a language model derived from the CV training data so that's also not a fair comparison unless the other engines were trained the same way.
Comparing systems trained on different datasets (with different language models) like this is like comparing apples to oranges. Mixing wildly different CPU and memory requirements into the benchmarks just makes it worse.
It should be fairly straightforward to do an unbiased comparison by training and evaluating with the standard Librispeech split and language model. It might be interesting to see how accuracy improves as the models scale up until they match the resource requirements of the other engines.
That said, the speed and memory usa
Read the rest
ge are impressive and I like the focus on very low resource environments. Seems like it has a lot of potential even if it may not be SOTA.
Sadly, Mozilla's effort is somewhat flawed in its current iteration. There are only very few different sentences (tokens) used as prompts, leading to a lot of repetitions of the same spoken utterance instead of greater coverage of possible utterances, and there is near complete overlap in the sentences spoken in the training set and those in the development and test sets, making it nearly impossible to generalize results from performance evaluations to utterances not contained in the training set. See https://discourse.mozilla.org/t/common-voice-v1-corpus-desig...
Some suggestions for improvements were made in a discussion on the Kaldi Github (https://github.com/kaldi-asr/kaldi/issues/2141), but the bottom line is that it's not clear what the exact use-case of the corpus is. It's not command recognition, there's no particular domain, just
Read the rest
more or less random English sentences collected from CC0 sources. Use cases should drive corpus collection.
Showing the 15 most recent of 29 sources. Enable JavaScript to search and browse every indexed mention.