Best open-source speech recognition (speech-to-text) models in 2026
Open speech recognition splits into four jobs: fast general transcription, live streaming with a small delay, long recordings with speaker labels, and tiny models for phones and edge devices. Whisper large-v3 turbo is still the most downloaded model here, but newer open models add features it does not offer natively, such as streaming, hour-long input and speaker labels. Pick by the job, then by license.
-
Whisper large-v3-turbo Most downloaded
OpenAI's MIT-licensed model is still the most downloaded speech-to-text model here, with about 5.1 million downloads in the 30 days to 11 November 2026. The turbo version is about 8x faster than large and runs in roughly 6 GB of GPU memory.
-
Qwen3-ASR
Alibaba's Apache 2.0 pair (1.7B and 0.6B): 30 languages plus 22 Chinese dialects, singing, streaming and timestamps.
-
Cohere Transcribe
Cohere's Apache 2.0 2B model that only transcribes: 14 languages, no timestamps or speaker labels, and you must name the language. Downloading it needs an access form on Hugging Face.
-
Jais-ASR
IFM's 5.7B Arabic recognizer under Apache 2.0, covering Modern Standard Arabic and dialects from 19 countries, with code-switching, optional diacritics and dialect labels. It has the best average on the Open Universal Arabic ASR Leaderboard in IFM's own table, but trails the top systems on clean read speech.
-
Voxtral Realtime
Mistral's Apache 2.0 4B streaming model: 13 languages and a delay you set between 80 ms and 2.4 s, so you can trade speed for accuracy.
-
Nemotron 3.5 ASR
NVIDIA's 0.6B streaming model: 40 language-locales in one model, with the delay set per request from 80 ms to 1.12 s. Eight of the locales need fine-tuning first. The license is permissive.
-
Confucius4-R2T2
NetEase Youdao's streaming model, adapted from Qwen3-ASR-1.7B: it commits words permanently and never revises them, with chunks from 80 ms to 2 s. Its own tests give 5.76% average error at 160 ms against 21.01% for the base model, and the license needs a separate agreement above 100 million monthly users or RMB 1 billion of revenue.
-
VibeVoice-ASR
Microsoft's MIT-licensed 8.7B model takes an hour of audio in one pass and adds speaker labels, timestamps and 50+ languages.
-
MOSS-Transcribe-Diarize
OpenMOSS's Apache 2.0 0.9B model that writes who spoke and when in one pass, for recordings up to 90 minutes in 50+ languages, with hotwords. In its own tests on Chinese-language meetings, podcasts and movies it scores better than VibeVoice-ASR and several commercial systems, and it had about 128,000 downloads in the 30 days to 11 November 2026.
-
Nemotron 3.5 Transcribe Newest
NVIDIA's open speech-to-text model: 40 languages, speaker labels, translation and prompt-controlled transcript style. The weights are about 68 GB in BF16, with FP8 and NVFP4 versions, and the license is permissive.
-
Granite Speech 5.0
IBM's Apache 2.0 pair: a 2B model that transcribes six languages and translates, plus a 470M CTC model built for speed and edge devices.
-
Parakeet TDT 0.6B v3
NVIDIA's 0.6B model for 25 European languages, with automatic language detection and word timestamps. CC BY 4.0 allows commercial use with credit.
-
GigaAM Multilingual
Sber's MIT-licensed 220M and 600M recognizers built for Russian, Kazakh, Kyrgyz and Uzbek, where its own tests put the large model far ahead of Whisper large v3 on the three Central Asian languages. English is only moderate and the output has no punctuation.
-
GigaAM-v3
Sber's 220M to 240M Russian recognizer under MIT. Its e2e versions write punctuated, normalized text, and Sber's tests show about 8.4% average word error rate against 25.1% for Whisper, with call-center and atypical speech covered.
-
Whistle
Cactus Compute's Apache 2.0 recognizer in a 16.9 MB file: seven European languages, word timestamps and, by the developer's own measurement, a first token in about 11 ms on a laptop CPU.
-
LFM3-Audio-2B
Liquid AI's 2B on-device voice model: transcription, synthesis and translation in 18 languages from one model. The LFM Open License allows commercial use only below US$10 million in yearly revenue.
How we chose
- Open weights you can download today from the developer's official repository.
- Word error rates in model cards are self-reported on different test sets, so we compare features instead of quoting a ranking.
- Streaming delay, maximum audio length and languages are listed, because they decide whether a model fits your job.
- Run the model on your own recordings before you decide, since accents, noise and domain words change results a lot.