{"id":"edge0-ark-asr-3b","nom":"ARK-ASR-3B","editeur":"Edge0","pays":null,"poids_ouverts":true,"poids_selon":null,"licence":"Apache 2.0","licence_selon":null,"liens":{"page":"https://edge0.ai/models/asr-3b"},"fiche_auto":true,"page":"https://quelleia.com/voix/modeles/edge0-ark-asr-3b/","releve":"2026-10-19","en_bref":"15e sur 53 en transcription en anglais (Open ASR) (ex æquo à valeur égale).","presentation":"Modèle de transcription multilingue qui associe un encodeur audio de type Whisper à un décodeur Qwen d'environ 3 milliards de paramètres. Edge0 le présente comme le plus précis de sa gamme Audio8 ASR, destiné aux enregistrements difficiles, longs ou multilingues traités sur GPU.","sortie":"2026-09-10","taille":"environ 3 milliards de paramètres (décodeur), plus un encodeur audio dédié","langues":{"annoncees":{"annoncees":"19 langues, dont le français","nombre":19,"francais":true,"liste":["zh","en","de","ja","fr","ko","es","pl","it","ro","hu","cs","nl","fi","hr","sk","sl","et","lt"]},"prises_en_charge_selon_open_tts":null,"evaluees":["en"]},"preuves":[{"appuie":["presentation","taille"],"lu":"https://huggingface.co/Edge0/ARK-ASR-3B/raw/main/README.md (« ARK-ASR-3B is a 3B-scale audio-capable autoregressive Transformers model for automatic speech recognition. »), 17/10/2026"},{"appuie":["presentation","taille"],"lu":"https://huggingface.co/Edge0/ARK-ASR-3B/raw/main/README.md (« It combines a Whisper-style audio encoder, an MLP adapter, and a Qwen decoder with custom »), 17/10/2026"},{"appuie":["langues"],"lu":"https://huggingface.co/Edge0/ARK-ASR-3B/raw/main/README.md (« ARK-ASR currently supports Chinese, English, German, Japanese, French, Korean, Spanish, Polish, Italian, Romanian, Hungarian, Czech, Dutch, Finnish, Croatian, Slovak, Slovene, Estonian, and Lithuanian ASR. »), 17/10/2026"},{"appuie":["sortie"],"lu":"https://edge0.ai/models/asr-3b (« Audio8 · ASR · 3B September 10, 2026 »), 17/10/2026"},{"appuie":["presentation"],"lu":"https://edge0.ai/models/asr-3b (« Audio8 ASR 3B reserves the largest parameter budget in the family for the hardest audio: meeting records, interviews, media subtitles and multilingual speech »), 17/10/2026"}],"classements":[{"id":"transcription/open-asr-anglais","nom":"Transcription en anglais, Open ASR","rang":15,"sur":53,"valeur":5.92,"unite":"%","chiffres":{"rtfx":479.11,"jeux":{"AMI-Cleaned":7.89,"Earnings22-Cleaned-AA-chunked":8.07,"Gigaspeech-Cleaned":6.94,"LS Clean":1.04,"LS Other":2.37,"Private (conversational)":12.9,"Private (scripted)":3.11,"SPGISpeech":2.46,"URGENT24":13.29,"Voice Arena Monsoon":3.48,"Voxpopuli-AA-Cleaned":3.55}},"voisins":["thestageai-thewhisper-large-v3-turbo","boson-ai-higgs-audio-v3-stt"],"source":{"nom":"Open ASR Leaderboard","mention":"Source : Open ASR Leaderboard, Hugging Face","url":"https://huggingface.co/spaces/hf-audio/open_asr_leaderboard","releve":"2026-10-19"}}],"noms_dans_les_sources":{"hf_audio":"Edge0/ARK-ASR-3B"}}