Most Downloaded Speech Recognition Models (Hugging Face)

Whisper, made by OpenAI, is the model family developers download most, directly and through copies that other accounts have repackaged for faster or smaller use. The very top of the table, however, is crowded with older community fine-tunes of Meta's wav2vec 2.0 and with audio tools that are not transcribers at all. The table holds 263 data points ¹, each the number of downloads of one model from Hugging Face in the latest month. Read it as a map of what software fetches, not as a vote by people who dictate.

Whisper and its copies are the working default

The most downloaded model that is plainly a general-purpose transcriber from a known lab is openai/whisper-large-v3-turbo, at 6,094,388 downloads ². Its larger sibling openai/whisper-large-v3 follows at 3,755,126 downloads ³, and the mid-sized openai/whisper-small at 3,510,611 downloads ⁴.

Developers do not only take the original files. Systran/faster-whisper-small, a conversion for a faster runtime, reached 4,582,267 downloads ⁵. The tiny conversion of the same runtime reached 2,326,259 downloads ⁶. In the browser world, Xenova/whisper-tiny.en drew 787,091 downloads ⁷, close to the original openai/whisper-tiny at 818,386 downloads ⁸.

The pattern is that small and mid-sized checkpoints are downloaded in bulk. They are cheap to fetch, quick to test and light enough to run in a test suite or on a phone. The family table shows how widely Whisper is spread across makers:

The largest raw counts belong to old fine-tunes that few people choose

The single biggest figure in the table is a single-language wav2vec 2.0 fine-tune by an individual account. Its Japanese version has 15,853,265 downloads ¹³. The Portuguese version has 4,394,855 downloads ¹⁴ and the Russian one 4,277,272 downloads ¹⁵.

The same shape repeats down the list. A Filipino model by Khalsuu has 1,265,235 downloads ¹⁶, and few people have liked it. Many language-specific fine-tunes from individual contributors, mostly created years ago, sit in the table with counts in the hundreds of thousands or millions.

Nobody who needs accurate dictation today picks these because they are the best. The likelier explanation is that they are named in tutorials, example notebooks, test fixtures and continuous-integration jobs that fetch a model on every run. The English version of the same author's model makes the point: its monthly count is only 61,035 downloads ¹⁷, yet its all-time count is the largest of any model in the table. Downloads follow habits and defaults, and habits fade.

Showing 6 of 15 rows. See every row in the dataset.

Several of the top entries do not transcribe speech

The tag is set by each model's author, and some audio tools carry it. The speaker-separation model pyannote/speaker-diarization-3.1 has 6,769,740 downloads ²¹. Its newer sibling pyannote/speaker-diarization-community-1 has 5,143,252 downloads ²². A voice-activity detector from the same group adds 377,102 downloads ²³.

These models decide who spoke when, or whether anyone spoke. They are common companions of transcribers, which explains the volume, but they do not turn speech into words. The same holds for aligners, which match a known transcript to the audio: MahmoudAshraf/mms-300m-1130-forced-aligner has 2,368,747 downloads ²⁴. If you rank only models that produce text, the leaders change.

Conversions for local use carry some of the biggest recent numbers

Some of the biggest recent numbers come from accounts that convert other labs' models into formats made for running on a laptop or phone. argmaxinc/whisperkit-coreml, a package of Whisper models for Apple devices, has 11,924,373 downloads ²⁵. The handy-computer account, which publishes files for a local runtime, has nemotron-3.5-asr-streaming-0.6b-gguf at 1,961,680 downloads ²⁶ and parakeet-unified-en-0.6b-gguf at 1,926,095 downloads ²⁷.

The converted copy can outrun the original. NVIDIA's own parakeet-tdt-0.6b-v3 has 626,738 downloads ²⁸, while the mlx-community version for Apple chips has 2,525,127 downloads ²⁹. The maker column in the table is the account that uploaded the files, so credit for popularity often lands on the converter and not on the lab that trained the model. That matters when you ask who makes the models developers use.

Showing 6 of 18 rows. See every row in the dataset.

New models from NVIDIA, Qwen, Mistral, Cohere and IBM are arriving, but at a smaller scale

Newer entries show where attention is moving. Mistral's Voxtral-Mini-4B-Realtime-2602, built for streaming, has 2,171,152 downloads ³⁴. Alibaba's Qwen team has Qwen3-ASR-1.7B at 1,186,346 downloads ³⁵. NVIDIA's nemotron-3.5-asr-streaming-0.6b has 1,221,712 downloads ³⁶, and Microsoft's VibeVoice-ASR has 686,190 downloads ³⁷.

Further back, Cohere's cohere-transcribe-03-2026 has 118,980 downloads ³⁸ and IBM's granite-speech-4.1-2b has 193,584 downloads ³⁹. Every one of these trails Whisper's main checkpoints, yet they are recent and several draw many likes relative to their downloads. Downloads lag attention: a model announced this year has had less time to be written into tutorials and defaults.

Downloads, likes and all-time counts tell different stories

The columns do not agree, and that is informative. openai/whisper-large-v2 has only 80,446 downloads in the latest month ¹¹, though its all-time total is among the largest in the table. Developers moved on to large-v3 and the turbo variant. In the other direction, a model with many likes can have modest recent use, and a model almost nobody praises can be downloaded millions of times. Likes measure a moment of approval. Monthly downloads measure what scripts still fetch. All-time totals measure history.

Some well-known models show almost nothing, and that is not a verdict

Several famous names read as zero or near zero. ggerganov/whisper.cpp shows 0 downloads ⁴⁰, although the project is widely used, because its users fetch files from other places. Meta's omniASR-CTC-7B also shows 0 downloads ⁴¹, and facebook/seamless-m4t-large shows 0 downloads ⁴². The Allen Institute's OLMoASR shows 3 downloads ⁴³ and Rev's reverb-asr shows 11 downloads ⁴⁴.

A zero can mean the files live in a different repository, are gated, or are served in a way the counter does not catch. It does not show that no one uses the model.

What the numbers cannot tell you

• They count fetches, not users. A server that re-downloads a model on every run counts each time. One busy pipeline can outweigh thousands of people.
• They miss other hosts. Models on GitHub releases, a vendor's own network, or Apple's Core ML packages are not counted, and commercial services that never publish weights are absent.
• They say nothing about accuracy. A popular model can be a poor fit for your language, accent or microphone. Word error rate on your own audio is the test; our overview of how speech recognition works explains how it is measured.
• The maker is an account, not a company. Conversions credit the converter, and individuals' names appear next to labs.
• The tag is author-set. Diarization, voice detection and alignment models appear beside transcribers.

What this means if you are choosing how to write

If you want to dictate, the table tells you less about which app to pick than about which engines the tools underneath are built on. Whisper and its relatives power a large share of local and open-source dictation, and newer streaming models from NVIDIA, Mistral and others are catching up on speed. For English, small and medium models are enough on a modern laptop. For a less common language, check whether a dedicated fine-tune exists, and test it, since many of the most downloaded ones are old.

Choose by your own audio, not by popularity: record a minute of your normal speech, run two or three candidates, and count the corrections you must make. For the wider picture of vendors, prices and products, see our market overview. A model that developers download in the millions may be a test fixture, and a model with few downloads may be the right one for you.