MediaTek Research, the advanced research division of the MediaTek Group, has released MR Breeze ASR 25, an open-source automatic speech recognition (ASR) model optimized for Taiwanese Mandarin and local speech patterns. Based on OpenAI's Whisper architecture, the model was designed to better understand the linguistic nuances of Taiwan, including Mandarin-English code-switching, a common feature in everyday speech.
Released under the permissive Apache 2.0 license, MR Breeze ASR 25 is part of MediaTek's broader effort to democratize AI tools across industries and promote the development of creative applications in voice-enabled technologies.
While open-source ASR models have made impressive progress, they often fall short in recognizing local accents and regional expressions, said Dr. Da-Shan Shiu, head of MediaTek Research. Simple phrases could be confusing due to the similar pronunciation in Mandarin. We built MR Breeze ASR 25 to better understand Taiwan's linguistic reality.
Compared with OpenAI's Whisper, MR Breeze ASR 25 improves speech recognition accuracy by nearly 10% and boosts code-switching performance by 56%, MediaTek said.
The model is the second-generation release in the MR Breeze series and is designed to run efficiently on edge devices and local systems, crucial for privacy-sensitive or latency-critical applications in areas like customer service, accessibility tech, and smart devices.
The launch of MR Breeze ASR 25 continues MediaTek Research's initiative to open-source a comprehensive suite of traditional Chinese AI tools. Since 2025, the group has released: a visual-language model (V-LLM) with integrated image capabilities; a Taiwanese-accent-tuned text-to-speech (TTS) system; and a function-calling model that connects language processing to external tools and data.
With MR Breeze ASR 25, MediaTek hopes to set a new benchmark in localized voice AI and accelerate innovation across the Mandarin-speaking world.
Article edited by Jack Wu
