Alibaba upgrades real-time speech model Fun-ASR-Realtime: first-word latency of 100 milliseconds, Wenzhou dialect recognition accuracy exceeds 82%

According to Beating's monitoring, Alibaba Tongyi Lab has reduced the first-character latency of the streaming speech recognition model Fun-ASR-Realtime to the millisecond level, achieving character output the moment the speech ends, with accuracy approaching that of offline models. During the recent 100-hour island live broadcast by Filmstorm, the model provided real-time subtitle support throughout, even in harsh conditions with outdoor rainstorms and frequent speaker switches, recognizing over 60,000 entries totaling 1.32 million characters.

To address the common issue of misinterpretation in real-time speech recognition, the new model enhances contextual awareness, dynamically correcting errors by combining historical dialogue and real-time hot words (for example, automatically correcting "Ye Lu" to "Ye Lu" based on subsequent context). The model currently supports 30 languages and 16 dialects, achieving an average character accuracy rate of 88.62% in dialect testing, with Shanghai dialect at 92.41% and the notoriously difficult Wenzhou dialect reaching 82.74%. The offline version, Fun-ASR-Flash, ranked first on the word error rate leaderboard of the global AI evaluation platform Artificial Analysis.

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