What are the most popular open-source speech recognition projects?

The most starred repository in our tracking is huggingface/transformers, with 167,850 stars ¹. It is a general machine learning library, so it is not a speech engine in the strict sense. Among projects built around speech, ggml-org/whisper.cpp leads with 54,482 stars ², followed by cjpais/Handy with 34,130 stars ³ and SYSTRAN/faster-whisper with 25,969 stars ⁴.

The tracked list holds 124 data points ⁵. It covers every repository in our market registry plus the most starred repositories (over 1,500 stars) under the GitHub topics speech-recognition, speech-to-text, whisper, asr and dictation. Here are the top rows, read from the current data.

Showing 8 of 124 rows. See every row in the dataset.

For the full ranking, see the full dataset.

Whisper-based projects fill much of the top of the list

Whisper is OpenAI's speech model, released with open weights, and many of the most followed repositories are tools built around it. whisper.cpp is a C++ port. SYSTRAN/faster-whisper has 25,969 stars ⁴. m-bain/whisperX, a Whisper tool with extra alignment features, has 24,667 stars ¹¹. chidiwilliams/buzz, a desktop transcription app, has 22,078 stars ¹².

You can filter the same table to a single project name, for example the one below.

Older toolkits are still on the list. mozilla/DeepSpeech has 26,770 stars ⁹, yet it is marked archived in the data, so it no longer receives updates. kaldi-asr/kaldi has 15,502 stars ¹³. Stars from the years when these were the standard research tools stay on the page long after development moves on.

Python is the most common language, with C++ and Rust next

Python dominates the rows, because most machine learning research code is written in it. Here are the Python rows.

Showing 8 of 60 rows. See every row in the dataset.

Projects meant to run fast on a laptop or phone tend to use compiled languages. whisper.cpp is written in C++, and cjpais/Handy and screenpipe/screenpipe are written in Rust. This matters if you plan to ship the engine inside an app, because a C++ or Rust project is easier to embed than a Python one.

MIT and Apache-2.0 are the common licenses, and a few are stricter

The license column tells you what you may do with the code. Here are the MIT rows.

Showing 8 of 48 rows. See every row in the dataset.

Apache-2.0 is also common, for example in NVIDIA-NeMo/Speech with 18,604 stars ¹⁷. Some projects use copyleft licenses such as AGPL-3.0 or GPL-3.0, which can require you to publish your own code if you build on theirs. Some rows show NOASSERTION, which means GitHub could not identify a single standard license. Read the repository's own license file before you build on any of them.

Stars show attention over time, not accuracy or fitness for your use

Stars measure how many GitHub users bookmarked a repository. They do not measure word error rate, speed, language coverage or how well a tool handles your accent. A few limits are worth keeping in mind:

• Stars accumulate. They rarely fall, so they reflect a project's whole history. An old project can sit above a better current one.
• Launches spike. A post on a news site can add thousands of stars in a day.
• Topics are self-assigned. Each repository's owners choose its topics, so some entries are only loosely about speech. Evil0ctal/Douyin_TikTok_Download_API, with 20,783 stars ¹⁴, is a video download tool that carries speech topics.
• Libraries are not engines. huggingface/transformers, mozilla-ai/llamafile and openvinotoolkit/openvino are general infrastructure that can run speech models among many others.

The chart below shows the leaders side by side.

To see how fast a project is growing, rather than how large it is, use the second chart.

Each reading is a snapshot, so growth is the change between readings, not a rate we can promise for the future.

What this means if you want to dictate or pick a tool

If you want to talk and have text appear, a star count is a weak guide. Most of the top repositories are engines or developer libraries, not apps you install and use. They need setup, a model download and often a command line.

• For free, general dictation, try the built-in tools first. Our page on free online dictation works in the browser with no setup.
• For a desktop transcription app built on an open engine, look at projects such as Buzz or Handy, and check the last push date in the data. A repository that has not been pushed in a year is a risk.
• For building your own product, check the license, the language and whether the repository is archived before you rely on it.
• For a ready-made app, Yap, which we make, runs dictation on your device on Mac and Windows. See pricing and download for details.

If you are comparing finished products rather than code, see the market overview. For background, read our overview of speech recognition and the guide to open-source speech recognition.