Descriptions:
Fahd Mirza takes ARK-ASR-3B — a newly released open-source multilingual speech-to-text model — for a hands-on local test, walking through installation, a custom Gradio interface, and live transcription across multiple languages. The 3B parameter model currently tops a major English ASR public leaderboard with roughly 5% word error rate across challenging real-world audio including recorded meetings, earnings calls, and audiobooks.
Running on an Nvidia RTX A6000 with 48 GB VRAM, Mirza demonstrates the model’s three-part architecture: an audio encoder, an MLP adapter, and a Qwen language model backbone that processes audio tokens as if they were text — a design choice that leverages the LM’s prior knowledge to recover from unclear audio. VRAM footprint comes in at just 7.6 GB for the 3B version, with a lighter 0.6B variant available for CPU inference.
The multilingual tests reveal a meaningful gap between the model’s English performance and its coverage of the 19 supported languages. Arabic, German, Hindi, and Spanish produce acceptable results, while Russian causes outright hallucination — a failure mode Mirza flags as a serious problem for voice agent deployments, since the model silently generates plausible-sounding but wrong text rather than refusing unsupported input. The video also touches on the model’s training recipe, Online Policy Distillation (OPD), a student-teacher reinforcement approach behind its strong English benchmark results.
📺 Source: Fahd Mirza · Published August 02, 2026
🏷️ Format: Tutorial Demo
