Which companies make the most used open speech recognition models?

The most used open speech recognition models do not come mainly from big companies. An independent developer, jonatasgrosman, makes the most downloaded model, with 15,853,265 downloads in the 30 days to October 19, 2026 ¹. Among companies, OpenAI leads with Whisper: whisper-large-v3-turbo has 6,094,388 downloads ². NVIDIA, Mistral, Microsoft and Qwen appear with newer models.

The data is a ranking of the most downloaded speech recognition models on the Hugging Face Hub, with 263 data points ³. Each row names the maker as shown on the model's page, which can be a person, a lab or a company.

Downloads by maker favor individuals and tool builders

The chart groups downloads by maker. Two things stand out. First, many leading makers are not the creators of the underlying technology. jonatasgrosman fine-tuned Facebook's wav2vec2 for many languages, and several of those models sit in the top 20: Japanese, Portuguese, Russian, Polish, Dutch, Greek, Arabic and Hungarian, among others. Second, tool builders such as Systran (faster-whisper) and argmaxinc (WhisperKit) rank high by repackaging Whisper.

OpenAI is the leading company by Whisper

OpenAI's rows are all Whisper models, and several are in the top 10.

Showing 8 of 11 rows. See every row in the dataset.

Beyond the turbo model, whisper-large-v3 has 3,755,126 downloads ⁴ and whisper-small has 3,510,611 downloads ⁵. Whisper's reach is bigger still when you add the people who repackage it. argmaxinc's WhisperKit has 11,924,373 downloads ¹¹ and Systran's faster-whisper-small has 4,582,267 downloads ¹².

pyannote ranks high with models that do not transcribe

pyannote makes two of the top five rows. pyannote/speaker-diarization-3.1 has 6,769,740 downloads ¹³, and pyannote/speaker-diarization-community-1 has 5,143,252 downloads ¹⁴. These models label who is speaking. They are often used alongside a transcription model, which explains the volume, but they carry the speech recognition tag without producing text.

Large labs publish newer models with fast-growing numbers

Several large organizations show up with models published in 2025 and 2026:

• Mistral: Voxtral-Mini-4B-Realtime-2602 has 2,171,152 downloads ¹⁵.
• NVIDIA: nemotron-3.5-asr-streaming-0.6b has 1,221,712 downloads ¹⁶, and parakeet-tdt-0.6b-v3 has 626,738 downloads ¹⁷.
• Qwen: Qwen3-ASR-1.7B has 1,186,346 downloads ¹⁸.
• Microsoft: VibeVoice-ASR has 686,190 downloads ¹⁹.
• Facebook: wav2vec2-base-960h has 1,720,472 downloads ²⁰.

A growing group of makers publish builds of other people's models: mlx-community and handy-computer, for example, convert Parakeet, Whisper, Voxtral and Cohere models for local use. Their rows are real demand, but credit for the underlying model belongs to the original lab.

What the numbers do and do not measure

• Maker means publisher. The maker is the account that uploaded the model, not always the team that invented it. A converted model counts for the converter.
• Downloads are not users. Automated pipelines that re-download a model on every run inflate counts.
• Only the Hugging Face Hub. Models hosted on GitHub releases, a vendor's CDN or Apple's Core ML packages are not counted, and closed cloud services such as Google, Amazon or Deepgram do not appear at all.
• Tags are author-set. A few non-transcription audio models carry the tag.
• Models are counted separately. A maker with many models, like jonatasgrosman, gets its strength from breadth, not from one popular release.

What this means when you choose a tool

For most people the maker of the model matters less than the app around it. Open models from OpenAI, NVIDIA and Mistral run in many apps, and independent builders are often the reason they run fast on your hardware. Check the product's own accuracy results and privacy terms rather than the model's download count. You can compare dictation products on our market page.

For background, read our overview of speech recognition and the guide to speech recognition models. See also how Whisper is used. Every row, with its maker and library, is in the full dataset.