Alibaba Upgrades Real-time Speech Large Model Fun-ASR-Realtime: Initial Latency Reduced to 100ms, Achieving Wenzhounese Speech Recognition Accuracy Over 82%

According to Dynamic Beating monitoring, Alibaba's Tongyi Lab has reduced the initial character delay of the Fun-ASR-Realtime large-scale streaming speech recognition model to the hundred-millisecond level, achieving almost real-time transcription with high accuracy approaching that of offline models. During the recent 100-hour live broadcast on a deserted island after the Movie Storm, the model provided real-time captioning support throughout the outdoor heavy rain and frequent speaker switches, recognizing over 60,000 captions totaling 1.32 million words.

To address the pain point of real-time speech recognition easily misconstruing context, the new model has enhanced its contextual awareness by incorporating historical dialogues and dynamically correcting errors based on real-time hotwords (e.g., automatically correcting "leaf deer" to "night heron" based on subsequent context). The model currently supports 30 languages and 16 dialects, achieving an average character accuracy rate of 88.62% in dialect tests, with Shanghainese at 92.41% and even the notoriously challenging Wenzhounese at 82.74%. The offline version of the model, Fun-ASR-Flash, has claimed the top spot in the global AI evaluation platform Artificial Analysis's word error rate ranking.

Currently, these two new models are only available as commercial API services on Alibaba Cloud Hundred Smeltings, while the underlying FunASR open-source framework and ecosystem models can be accessed on the Moddle community and GitHub.