Open-Source Speech Recognition on GitHub: Stars

Developers follow Whisper-based tools more than any other family of open-source speech software, and the newest repositories now rival the old research engines in attention. The table below holds 124 data points ¹, each one a repository and its star count on the day of the latest snapshot. Among projects that clearly do speech recognition, whisper.cpp leads with 54,482 stars ². Handy, a free dictation app, follows at 34,130 stars ³.

Stars measure attention, not quality, and they are a single snapshot. The sections below explain what the numbers show, what they hide, and what they mean if you want to dictate with open-source software. For background on how the technology works, see our overview of speech recognition, and for the commercial side see the market page.

Whisper-based engines draw the most attention

The largest cluster of well-followed speech repositories builds on OpenAI's Whisper model. whisper.cpp, a C++ port that runs on ordinary laptops, has 54,482 stars ². faster-whisper, a Python reimplementation, has 25,969 stars ⁴. WhisperX, which adds word timestamps and speaker labels, has 24,667 stars ⁵. Buzz, a desktop transcription app, has 22,078 stars ⁶.

The pattern is that developers reward projects that make a single released model easy to run: faster, smaller, on a laptop, with a window to click on. Few follow the model itself. Further down the same family sit WhisperLiveKit at 11,155 stars ⁷ for live streaming, and WhisperKit (published as argmax-oss-swift) at 6,406 stars ⁸ for Apple devices.

Older research toolkits keep their followers but are fading

Before Whisper, open-source speech recognition meant research toolkits. Kaldi has 15,502 stars ⁹. Vosk has 15,184 stars ¹⁰ and is still actively pushed. sherpa-onnx has 15,373 stars ¹¹. NVIDIA's NeMo speech repository, which hosts the Parakeet models, has 18,604 stars ¹². FunASR, from the ModelScope community, has 20,750 stars ¹³.

Their fates differ. Mozilla DeepSpeech holds 26,770 stars ¹⁴ but is archived, so it no longer receives updates. Coqui STT, with 2,604 stars ¹⁵, has not been pushed to for well over a year according to its last push date in the table. Stars accumulate and rarely fall, so a large count on a dormant repository mostly records past enthusiasm. When you choose a tool, the last push date and the archived flag say more than the star count does.

Dictation apps are the newest and quickest-rising group

A distinct group of repositories is not a speech engine at all but a finished tool for typing by voice, usually a desktop app that wraps a Whisper-class or Parakeet model. Handy leads, with 34,130 stars ³. FluidVoice has 12,287 stars ²¹, OpenWhispr has 9,573 stars ²², OpenLess has 3,893 stars ²³, and OpenSuperWhisper has 3,064 stars ²⁴. Smaller entries include TypeWhisper for Mac at 1,866 stars ²⁵ and Amical at 1,561 stars ²⁶.

Most of these were created in the last two years, and several were created this year, yet they already outrank toolkits that have existed for a decade. That is the closest the data comes to an answer on speed of growth: the youngest repositories reach large star counts quickly, and the established engines add stars slowly from a high base. Some of these apps are also an interesting split in licensing. Handy and OpenSuperWhisper use the permissive MIT license, FluidVoice and TypeWhisper use GPL-3.0, and OpenLess uses AGPL-3.0, which obliges anyone who offers a modified version as a network service to share their changes. Hex, another Mac dictation app, shows 2,902 stars ¹⁷ but is flagged as archived, a reminder that popular dictation projects can still be shelved.

The largest numbers belong to repositories that are not recognition engines

The highest counts in the table need careful reading. huggingface/transformers has 167,850 stars ²⁷, but it is a general machine learning library in which speech is one small part. debpalash/VoiceStudio has 76,616 stars ²⁸ and jamiepine/voicebox has 57,987 stars ²⁹. Both were created this year. Their names suggest voice generation rather than recognition, and the table does not record what the code does, so treat them as attention to voice technology in general.

Other entries are tools that use transcription as one step. Evil0ctal/Douyin_TikTok_Download_API has 20,783 stars ³⁰, and the video translation tool pyvideotrans has 19,452 stars ³¹. Meetily, a meeting note taker, has 31,927 stars ³², and screenpipe has 21,986 stars ³³. A few rows, such as the learning notes in kmario23/deep-learning-drizzle at 12,963 stars ³⁴, are collections of links rather than software. The dataset's rule excludes repositories whose topics are set to speech words but whose code does something else, yet topics are chosen by each repository's owners and a few such entries remain. Read the table as the field of attention around speech, not as a ranking of recognizers.

Showing 12 of 124 rows. See every row in the dataset.

Where the data disagrees with itself

Three tensions run through the table. First, popularity and recency disagree: the engine with the most followers among the older generation, Kaldi at 15,502 stars ⁹, trails the newest dictation apps in momentum even though its count is large. Second, stars and forks disagree about use. Forks are a rough sign that people build on a project, and some star-heavy repositories have comparatively few forks, so a launch spike can inflate stars without much real use. Third, language is a poor guide to speed: dictation apps appear in Rust, Swift, TypeScript and Python, while engines are mostly Python and C++, so the choice of language reflects the platform each tool targets more than any technical edge.

What the data cannot tell you

The table has real limits, and you should keep them in mind before drawing conclusions.

• No growth rate. Each row is one count on one day. Stars accumulate and rarely fall, so they reflect a project's whole history, and the table has no earlier count to subtract. We can say which repositories are young and large. We cannot say how many stars each added last month.
• No accuracy or speed. A star is a bookmark. It tells you nothing about word error rate, latency or language coverage. For accuracy you need benchmarks, which are covered in our overview of speech recognition.
• Launch spikes. A post on a news site can add thousands of stars in a day, so a young repository's count may be a burst rather than a steady trend.
• Owner-set topics. Repositories choose their own topics, so some rows are only loosely about speech, as the previous section showed.
• Licenses are as declared. Entries marked NOASSERTION, such as Kaldi and PocketSphinx, have a license file GitHub could not classify. That does not mean they are unlicensed, but you should read the license yourself.
• Hosted services are absent. Closed commercial dictation products have no public repository. See the market page for them.

What this means if you want to write by voice

If you want to dictate and are comfortable installing software, the open-source dictation apps are a real option, and the most followed of them, Handy at 34,130 stars ³, is actively pushed. Check three things in the table before you install anything: the last push date, the archived flag, and the license. Prefer a repository pushed within the last few weeks, avoid archived ones for daily use, and read the license if you plan to modify or redistribute.

If you are a developer building voice features, the picture is different. whisper.cpp at 54,482 stars ², faster-whisper at 25,969 stars ⁴ and Vosk at 15,184 stars ¹⁰ are widely used building blocks, and each has an active record of recent pushes. Test them on your own audio, because a star count will not tell you how a model handles your accent, vocabulary or microphone.

If you simply want the least setup, a packaged app with on-device processing and no model to configure may suit you better than a repository. Yap, which we make, is one such app for Mac and Windows, and the market page compares the options. For free, built-in alternatives, our guide to speech recognition explains what your operating system already offers.