Woofun AI reports that Alibaba's Tongyi Lab has optimized the Fun-ASR-Realtime large-scale streaming speech recognition model, reducing initial character delay to the hundred-millisecond level. This update enables near real-time transcription with accuracy comparable to offline models. During a recent 100-hour live broadcast in adverse weather conditions, the system processed over 60,000 captions totaling 1.32 million words, maintaining performance despite frequent speaker switches and heavy rain.

The model enhances contextual awareness by integrating historical dialogue and dynamically correcting errors using real-time hotwords, such as adjusting "leaf deer" to "night heron" based on subsequent context. It supports 30 languages and 16 dialects, achieving an average dialect accuracy of 88.62%, with specific rates of 92.41% for Shanghainese and 82.74% for Wenzhounese. While the offline variant Fun-ASR-Flash leads global word error rate rankings, these new models are currently available only as commercial APIs on Alibaba Cloud Hundred Smeltings, with open-source frameworks accessible via Moddle and GitHub.