What are the most popular speech recognition models?

By Hugging Face downloads, the most popular speech recognition model is a single-language wav2vec2 model for Japanese. jonatasgrosman/wav2vec2-large-xlsr-53-japanese was downloaded 15,853,265 times in the 30 days to October 19, 2026 ¹. Behind it, the picture is more familiar: Whisper models and tools built on Whisper fill much of the top of the list.

This page reads a ranking of the most downloaded models tagged for automatic speech recognition on the Hugging Face Hub. The dataset holds 263 data points ². Counts cover the 30 days to October 19, 2026.

A Japanese wav2vec2 model tops the list, ahead of WhisperKit

The two leading rows are not what most people would guess. The Japanese wav2vec2 model above has 15,853,265 downloads ¹. Second is argmaxinc/whisperkit-coreml, a package of Whisper models converted for Apple devices, with 11,924,373 downloads ³.

Third and fourth show how mixed the list is. pyannote/speaker-diarization-3.1 has 6,769,740 downloads ⁴. That model separates who spoke when. It does not turn speech into words, but it carries the speech recognition tag. Next is OpenAI's openai/whisper-large-v3-turbo with 6,094,388 downloads ⁵.

Showing 8 of 263 rows. See every row in the dataset.

Whisper is the most downloaded general-purpose model family

Among models built to transcribe many languages, Whisper dominates. openai/whisper-large-v3 has 3,755,126 downloads ¹⁰, and the smaller openai/whisper-small has 3,510,611 downloads ¹¹. Re-packaged versions add to this: Systran/faster-whisper-small has 4,582,267 downloads ⁷. Our question on how widely Whisper is used looks at the family in detail.

Newer models are already in the top 20

Several models first published in 2025 and 2026 sit among the most downloaded:

• mlx-community/parakeet-tdt-0.6b-v3, a Parakeet build for Apple's MLX framework: 2,525,127 downloads ¹².
• mistralai/Voxtral-Mini-4B-Realtime-2602, from Mistral: 2,171,152 downloads ¹³.
• nvidia/nemotron-3.5-asr-streaming-0.6b, from NVIDIA: 1,221,712 downloads ¹⁴.
• Qwen/Qwen3-ASR-1.7B, from Qwen: 1,186,346 downloads ¹⁵.

Many of these are published in formats made for running on a laptop or phone, such as MLX and GGUF builds. That points to developers looking for models that run locally, which is also the approach behind on-device dictation.

What these numbers do and do not measure

A download is a developer or a program fetching model files. It is not a user, a transcription, or a measure of quality.

• Automated pipelines inflate counts. Systems that re-download a model on every run count each time. A model used in many build or test jobs can outrank a model with more real users.
• Other hosts are invisible. Models served from GitHub releases, a vendor's CDN or Apple's Core ML packages are not counted here.
• The tag is set by authors. A few audio models that do not transcribe, such as the speaker diarization models above, carry the speech recognition tag and appear in the ranking.
• Older models stay high. Single-language wav2vec2 models from 2022 keep large counts, likely from pipelines that still pull them. The all-time column on each row shows how long a model has been accumulating downloads.
• Only open models on one hub. Cloud services such as Google, Amazon or OpenAI's hosted transcription do not appear.

What this means when you choose how to dictate

Downloads tell you what developers build with, not which tool is best for you. Use the ranking to learn which engines are widely supported: a popular model usually has more apps, guides and fixes around it. For accuracy, look at benchmark results on your language and audio instead, such as in our comparison of Whisper and Parakeet.

If you just want to dictate, you do not need to download a model at all. Your operating system's built-in dictation or a ready-made app handles that, and you can compare products on our market page. If you want to understand how these models work, start with our overview of speech recognition and the guide to speech recognition models.

You can see every row, with its maker, library, creation year, likes and all-time downloads, in the full dataset.