Speech-to-Text SDK Downloads: PyPI and npm

Developers build on the large AI platforms first, and on speech-specific tools second. The OpenAI package on PyPI was downloaded 295,856,631 downloads ¹ in the last month, Google's google-genai 137,607,856 downloads ², and on npm the OpenAI package reached 175,958,798 downloads ³. Those packages are not speech tools. They carry chat and image calls too, and transcription is one endpoint among many. Read the top of this table as a map of where AI development happens, not as a ranking of speech recognition.

The dataset holds 102 data points ⁴: the official SDK and library packages for each speech-to-text API and open model, read from PyPI and npm. For background on how the technology works, see our overview of speech recognition. For the companies behind these packages, see the market overview.

General AI platforms dwarf every speech-only provider

The top of both registries is occupied by companies whose main product is a language model. On PyPI, the next tier after OpenAI and Google is Groq at 16,795,477 downloads ⁵, Mistral at 16,185,864 downloads ⁶ and Google Cloud Speech-to-Text at 15,131,032 downloads ⁷. On npm, Mistral's package leads after OpenAI and Google with 32,317,341 downloads ⁸, and Cloudflare follows with 13,493,182 downloads ⁹.

The Cloudflare and Oracle rows show the same effect. The oci package on PyPI (5,164,205 downloads ¹⁰) is Oracle's whole cloud client, while its dedicated npm speech package, oci-aispeech, has 473,238 downloads ¹¹. The speech-only package is a small fraction of the umbrella one. The same holds for Telnyx, Replicate, fal, Together AI and Fireworks AI, whose SDKs serve many models.

Among dedicated speech providers, ElevenLabs and Deepgram lead

Packages built mainly for transcription give a cleaner signal. On PyPI the ElevenLabs package has 9,073,441 downloads ¹², Deepgram's deepgram-sdk 3,166,523 downloads ¹³ and AssemblyAI 2,797,588 downloads ¹⁴. On npm the order of the first two flips slightly: Deepgram has 4,092,784 downloads ¹⁵, ElevenLabs 3,970,853 downloads ¹⁶, and AssemblyAI 2,280,177 downloads ¹⁷.

ElevenLabs is best known for voice synthesis, so part of its count is probably text-to-speech users who never transcribe anything. The row is listed because the same package exposes its Scribe transcription. That caveat applies to every provider with several products.

The established cloud vendors sit in the same range. Amazon Transcribe has 2,186,454 downloads ¹⁸ for its batch npm client and 1,277,058 downloads ¹⁹ for streaming. Azure's Python speech package has 2,520,663 downloads ²⁰. Smaller specialists follow far behind: Speechmatics' real-time npm client has 192,130 downloads ²¹, Soniox's PyPI package 588,965 downloads ²², and Rev AI's 563,673 downloads ²³.

Showing 8 of 37 rows. See every row in the dataset.

Open models are used through a handful of runtimes

Open models split into two groups: runtimes that many people install, and research code that few do. On PyPI, faster-whisper leads with 9,614,222 downloads ²⁷. It is followed closely by the classic SpeechRecognition wrapper at 9,042,851 downloads ²⁸, and by the original openai-whisper package at 4,986,176 downloads ²⁹.

Below them come the engines aimed at devices and servers: sherpa-onnx with 1,614,806 downloads ³⁰, SpeechBrain with 1,015,647 downloads ³¹, Vosk with 633,645 downloads ³², NVIDIA NeMo with 445,752 downloads ³³, and the Apple-silicon port mlx-whisper with 340,121 downloads ³⁴. The JavaScript side tells a similar story in a different shape: sherpa-onnx-node has 2,109,793 downloads ³⁵, more than any other open-model npm package here, while Vosk on npm has only 8,270 downloads ³⁶.

Two older projects show what a stalled library looks like. Mozilla DeepSpeech has 2,375 downloads ³⁷ on PyPI and 784 downloads ³⁸ on npm, and Coqui STT has 1,211 downloads ³⁹. PocketSphinx, from an even earlier generation, still has 511,134 downloads ⁴⁰, probably from old tutorials and pinned dependencies.

Showing 8 of 28 rows. See every row in the dataset.

The same provider can look large or tiny depending on which package you count

A provider often ships several packages, and they rarely move together. Speechmatics has a real-time and a batch package on each registry, and the counts differ widely: 61,951 downloads ⁴¹ for the Python batch package, 237,868 downloads ⁴² for the Python real-time one, and only 13 downloads ⁴³ for speechmatics-sdk. A reader who picked one package would reach a very different conclusion from one who picked another.

Picovoice is the extreme case. Its Cheetah and Leopard packages each stay in the hundreds or low thousands, for example 2,012 downloads ⁴⁴ for pvcheetah and 843 downloads ⁴⁵ for @picovoice/leopard-node. Its model runs on-device from a licensed binary, so low downloads say little about how many devices run it.

Language and region also change the picture. The Chinese cloud vendors have modest counts on these Western registries: Baidu's PyPI package has 38,841 downloads ⁴⁶ and Tencent Cloud's 8,787 downloads ⁴⁷. Alibaba's dashscope is higher at 2,487,885 downloads ⁴⁸. Developers in China often install from domestic mirrors, which this count excludes, so these figures likely understate real use.

What the numbers cannot tell you

Download counts measure installs, not people and not minutes of audio transcribed. Automated builds dominate the counts: a continuous-integration system that installs a package on every build adds a download each time. A single large customer can move a package's total, and installs from private mirrors and caches are not counted at all. The numbers therefore show how much software depends on a package, not how many developers chose it or how happy they are with it.

Four further limits matter when you read the table.

• Umbrella packages inflate the count. Nothing in a download tells you whether the code calls the transcription endpoint. The OpenAI, Google, Mistral, Groq and Cloudflare rows are mostly chat and other uses.
• Quality is invisible. A package with many downloads may be popular because it is the default in a framework, not because its recognition is accurate. For accuracy, see the evidence in our overview of speech recognition.
• Only official packages are counted. Unofficial wrappers are excluded by design, so a provider with a strong community wrapper is understated.
• Hosted use is missed. Many people use these models through a web interface, a REST call or a finished app and never install an SDK.

Different registries also disagree. Deepgram and ElevenLabs trade places between PyPI and npm, and Gladia appears here only through npm, with 18,223 downloads ⁴⁹. These are readings of a single window, so a ranking of close neighbors should not be treated as settled. The page tracks the monthly figure every week, which lets you see whether a gap is stable.

What it means if you are choosing how to write

If you are a developer picking a transcription service, downloads show which SDKs are well maintained, widely integrated and likely to have answers on forums. They do not show which engine is most accurate or cheapest for your language and audio. Test two or three candidates on your own recordings.

If you want recognition that runs on your own machine, the open-model rows show where the community has settled: faster-whisper and openai-whisper on PyPI, and sherpa-onnx for device and JavaScript use. On a Mac with Apple silicon, mlx-whisper has a solid base of 340,121 downloads ³⁴.

If you simply want to dictate and are not writing software, none of this requires you to install anything. The built-in tools on your phone or computer are often enough, and a dedicated app can add speed and formatting. The package counts describe the infrastructure underneath such apps, and the market overview covers the apps themselves.