Which speech-to-text APIs do developers build on most?

Developers build on OpenAI most, judged by SDK downloads. The PyPI package openai was downloaded 295,856,631 times in the month to October 19, 2026 ¹. On npm, the same package reached 175,958,798 downloads ². Google is second on both registries, and every other provider is far smaller.

One caution comes first. The biggest packages are general SDKs for a whole AI platform, so a download does not mean someone transcribed audio. Read the ranking as a map of which vendors developers already have installed, not as a count of speech-to-text projects.

OpenAI and Google lead both registries

On PyPI, google-genai (Gemini API transcription) was downloaded 137,607,856 times ³. On npm, @google/genai reached 102,468,836 ⁴. Together with OpenAI, these two vendors account for the top four rows of the data.

The table below lists the leading Python packages in the dataset.

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

The same view for JavaScript:

Showing 8 of 42 rows. See every row in the dataset.

Below the leaders, the field is a long tail

After OpenAI and Google, the next group is made up of other model platforms and infrastructure providers. On npm, @mistralai/mistralai (Mistral Voxtral Transcribe) was downloaded 32,317,341 times ¹¹. On PyPI, groq (GroqCloud Speech to Text) reached 16,795,477 ⁵, and mistralai reached 16,185,864 ⁶.

Cloud-vendor speech services sit in the same range. The Python package google-cloud-speech was downloaded 15,131,032 times ⁷. On npm, Cloudflare's cloudflare package reached 13,493,182 ¹².

Open models are installed through their own Python packages

Open models do not need an API key, so their packages show a different pattern. faster-whisper was downloaded 9,614,222 times on PyPI ⁸, and the older SpeechRecognition library reached 9,042,851 ¹⁰. The original openai-whisper package was installed 4,986,176 times ¹⁷.

On npm, open models are much smaller. The leading one in the data is sherpa-onnx-node at 2,109,793 downloads ¹⁸.

Speech specialists are used, but at a smaller scale

The companies that sell only speech to text appear well below the general platforms. Deepgram's @deepgram/sdk was downloaded 4,092,784 times on npm ¹⁴, and its Python package deepgram-sdk reached 3,166,523 ¹⁹. AssemblyAI's Python package was downloaded 2,797,588 times ²⁰. ElevenLabs Scribe, through the elevenlabs Python package, reached 9,073,441 ⁹.

A smaller number does not mean a weaker product. A specialist's SDK is installed only by people who use that one service, while a general AI SDK is installed by everyone who touches the platform.

The trend over time

The dataset tracks these packages weekly, so the history of each row shows whether a package is growing. The chart shows the downloads over time.

The two charts below split the same rows by registry.

What these numbers do and do not measure

The dataset holds 102 ²¹ data points from 4 ²² distinct sources. Each is a published figure, and each row covers the official SDK or library packages for a speech-to-text API or open model.

Be clear about the limits:

• Automated builds dominate. Continuous-integration systems that install packages on every build dominate the counts, and one large customer can move them.
• Private mirrors are missing. Installs from private mirrors and caches are not counted, so companies that cache packages internally are undercounted.
• A download is not a user. One developer can cause many downloads, and one download can serve many end users.
• General SDKs overstate transcription. The openai, google-genai and mistralai packages cover far more than speech to text. The same is true of cloudflare, replicate and together. The count says nothing about how much of that use is transcription.
• Some packages mix uses. The gradium package on PyPI also covers text to speech, so its count is not only speech to text.
• One rolling month. Each figure is a month to a single date, so a short spike can distort a reading.

The count also leaves out hosted products with no developer package, and anyone who calls an API with plain HTTP requests and no SDK.

What this means for choosing a speech-to-text tool

If you are picking an API, downloads tell you about ecosystem size: how many developers already use the vendor's SDK, and how likely you are to find examples and answers. They do not tell you about accuracy, latency, price or language coverage. For those, compare products directly in our market overview and read the guide to speech-to-text APIs.

If you want to run a model yourself, the open-model rows show where developers go: faster-whisper on Python is the clear starting point in this data, and sherpa-onnx on JavaScript.

If you only want to dictate, you do not need any of these packages. A finished app is simpler, and Yap, our own on-device dictation app for Mac and Windows, is one option. See the full overview of speech recognition for how the pieces fit together, and download to try it.

For every row, its history and its source, open the full dataset.