Latest News

  • Industry Cooperation

  • Publish Date：2026-09-01
NYCU and MediaTek open-source Breeze‑ASR‑26, ushering in a new era of Taiwanese Hokkien AI applications
Concept image illustrating the potential use of Breeze-ASR-26 for Taiwanese Hokkien speech recognition in mobile applications; it does not depict an actual app interface.
Concept image illustrating the potential use of Breeze-ASR-26 for Taiwanese Hokkien speech recognition in mobile applications; it does not depict an actual app interface.
 
Edited by Chance Lai
______
For teams building technology in Taiwanese Hokkien, also known as Taigi, the work has often begun far from the final product: collecting and labeling audio, training a model and securing computing resources before a single app feature could be tested. Breeze-ASR-26 changes that starting point.

Developed by the Speech AI Research Center at National Yang Ming Chiao Tung University's Industry Academia Innovation School and MediaTek Innovation Base, the open-source automatic speech recognition model gives developers a ready-made base for language-learning tools, customer-service systems, health care voice assistants, cultural-preservation projects and other speech-enabled services.

Instead of rebuilding the same foundational capability, downstream teams can focus on the work that turns a model into a useful product: adapting it to a specialist domain, reducing its computing requirements, deploying it on edge devices and designing an experience that works for real users.

A Benchmark Built for a Difficult Language

On the public Breeze Taigi Benchmark, Breeze-ASR-26 recorded an average character error rate, or CER, of 30.13% — the lowest result among the five systems evaluated under the same conditions. The Ministry of Education's Taiwanese Hokkien input method followed at 30.70%, with Yating at 32.11%, Gemini 3 Flash at 32.52% and Breeze ASR 25 at 49.99%. A lower CER means fewer insertions, deletions and substitutions in the output.

The benchmark uses 30 pairs of Mandarin and Taiwanese Hokkien public-service announcements released by Taiwan's Executive Yuan. The researchers normalized the transcripts to make comparisons across systems more consistent — an important step for a language with regional accents, tonal variation and multiple writing conventions.

Breeze-ASR-26 is based on the multilingual Whisper framework and was fine-tuned with about 10,000 hours of synthetic Taiwanese Hokkien speech. Its model weights are available on Hugging Face under the Apache 2.0 license, giving researchers and developers a common technical starting point that they can inspect, adapt and deploy.
 
 
 From Model Weights to Working Products

The project's August 21 snapshot shows how quickly that starting point has spread. MediaTek-Research/Breeze-ASR-26 logged 12,564 downloads in the previous month and 70 likes on Hugging Face. Its model tree listed 21 direct first-generation derivatives, while three Hugging Face Spaces offered web demos, application programming interfaces and other interactive uses.

Community conversions and optimizations now span CTranslate2, GGML, INT8, MLX, CoreML and WhisperKit. Together, those deployment paths bring the same recognition capability to servers, personal computers, smartphones and lower-power edge hardware.

The technology is also appearing in products and learning environments. ByeType Voice has added offline Taiwanese Hokkien recognition on iOS, while Kong Tai-gi offers online and offline recognition on Android. National University of Tainan has integrated Breeze-ASR-26 into its TAIDE Taiwanese-English AI Learning Companion, and FriendliAI provides a dedicated endpoint for developers that want to integrate the model into their own services.

What the Numbers Do — and Do Not — Show

The 30.13% CER is a comparative result on the Breeze Taigi test set, not a universal measure of performance. It should not be converted into a 69.87% “accuracy” figure or compared directly with results from different datasets. The benchmark paper also notes that its Taigi-to-Mandarin mapping is designed to compare systems; because the two languages are distinct, even a strong Taigi recognizer would not necessarily produce a zero CER against Mandarin-character references.

The ecosystem figures require similar care. The total of 12,564 reflects downloads during the previous month, not lifetime downloads, and the 21 direct derivatives do not represent 21 separate adopting organizations. They do, however, document active experimentation across software formats, devices and use cases.

For NYCU and its partners, that activity marks a practical transition. Taiwanese Hokkien speech AI is no longer only a model-building challenge; it is becoming an application platform. By lowering the barriers to data, training, computing and integration, Breeze-ASR-26 turns the next question into a more immediate one: what useful services can be built, and for whom?
 
文/智能系統研究所

臺語 AI 發展已進入高效能模型可直接支援應用創新的新階段。Breeze-ASR-26 是由陽明交通大學產創學院人工智慧語音研發中心與聯發科創新基地共同合作完成，並由雙方共同推動開源的高效能臺語語音辨認模型。高效能臺語語音辨認能力如今已可直接取得。隨著 Breeze-ASR-26 公開，Breeze-ASR-26 已建立可直接支援下游應用的技術基礎。從學生、教師、研究團隊到新創與產業開發者，都可直接運用既有模型能力，將研發資源投入 App、Edge AI、教育、醫療與智慧服務。臺語 AI 的下一階段，已經進入「誰能做出真正有價值的應用」。

這項轉變的基礎，是臺語已具備一套效能足以支撐後續開發的語音辨認模型。在 Breeze Taigi 公開 Benchmark 中，Breeze-ASR-26 的平均字元錯誤率（Character Error Rate, CER）為 30.13%，在相同測試條件下低於教育部臺灣台語輸入法 30.70%、Yating 32.11%、Gemini 3 Flash 32.52% 與 Breeze ASR 25 的 49.99%，為五套受測系統中最低。這項結果的意義不只在於 Benchmark 領先，更在於臺語語音辨認已有一個經公開測試驗證、可直接取得並持續開發的技術起點。Breeze-ASR-26 已建立可直接使用的高效能臺語語音辨認基礎，讓更多人能把創意直接轉化為 App、Edge AI、教育、醫療與智慧服務。

從「先做模型」到「直接做應用」　臺語 AI 創新起跑線已改變

這項成果由陽明交通大學產創學院人工智慧語音研發中心與聯發科創新基地攜手研發並共同開源。Breeze-ASR-26 讓臺語學習工具、智慧客服、醫療語音助理與文化保存服務等應用，可以直接建立在成熟的臺語語音辨認能力之上。Breeze-ASR-26 以 Whisper multilingual framework 為基礎，使用約 10,000 小時合成臺語語音進行微調。模型公開後，下游團隊可直接運用既有臺語語音辨認能力，將研發重心轉向領域調適、量化、Edge deployment、使用者介面與產品整合。這使臺語 AI 的競爭焦點，已經從重複建置基礎模型轉為應用創新與實際需求。

Benchmark 證明模型「可用」　真正的 Impact 是讓更多人開始創造應用

Breeze Taigi Benchmark 的公開結果提供了關鍵技術佐證。Breeze-ASR-26 的 CER 為 30.13%，在五套受測系統中最低；CER 越低代表辨認錯誤越少。對應用開發而言，這項結果所回答的核心問題是：臺語是否已有一套效能足以讓其他團隊直接往下開發的語音辨認模型？目前的公開測試結果顯示，公開結果已明確證實。當基礎能力可以直接取得，有限的研發資源便能進一步投入產品、服務與場域創新。
 


12,564 次近月下載、21 個衍生模型　開源效應已形成

模型公開後的生態發展，更顯示這項技術已走出實驗室。截至 2026 年 8 月 21 日，MediaTek-Research/Breeze-ASR-26 在 Hugging Face 顯示近月下載 12,564 次、70 Likes；官方 model tree 已形成 21 個第一代直接衍生模型，另有 3 個 Hugging Face Spaces。從 CTranslate2、GGML、INT8、MLX、CoreML 到 WhisperKit，社群正把同一套臺語語音辨認能力帶往伺服器、個人電腦、手機與 Edge 裝置。開發者的核心任務，已經從「重新訓練一套臺語模型」轉為「把臺語 AI 帶進真正的產品與服務」。

從 App 到教育現場　臺語 AI 已已進入真實使用情境

這項轉變已反映在實際應用。ByeType Voice 已加入 Breeze-ASR-26 離線臺語辨認；Kong Tai-gi 提供線上與離線臺語辨認；國立臺南大學「TAIDE 台英語 AI 學伴」正式導入 Breeze ASR 26；FriendliAI 亦提供 Breeze-ASR-26 Dedicated Endpoint。從手機 App、Edge AI 到教育與開發者服務，一條由「開源模型」通往「實際應用」的臺語 AI 生態正已形成。這些案例的重要性不只是增加應用數量，而是顯示臺語語音辨認正成為可重複利用、可持續擴充的 AI 基礎能力。

資料說明：
12,564 為 Hugging Face「Downloads last month」，非歷史累積下載；21 個直接衍生模型不代表 21 個獨立採用組織。30.13% 為 Breeze Taigi 公開測試集的 CER，不應換算為 69.87%「準確率」，亦不可與其他資料集的辨認率直接比較。

  • 主要來源：Breeze Taigi（arXiv:2603.19259）、MediaTek Research Hugging Face、Speech AI Research Center Hugging Face、Apple App Store、Google Play、國立臺南大學、GitHub、FriendliAI。（資料查核日期：2026-08-21）
Related Image(s)：
Back