Alibaba Tongyi Fun-ASR Speech Model Major Upgrade: First-word Latency Only 100 Milliseconds, Wenzhou Dialect Recognition Exceeds 82%

Alibaba Tongyi Lab has reduced the first-character latency of its streaming speech recognition model, Fun-ASR-Realtime, to the millisecond level, achieving instant character output as speech concludes, with accuracy approaching that of offline models. The new model supports 30 languages and 16 dialects, with Wenzhou dialect recognition accuracy reaching 82.74%.

(Previous context: xAI officially launches Grok voice API, TTS at $4.2 per million characters, recognition rate beats ElevenLabs)
(Background supplement: OpenRouter analyzes 100 trillion tokens: What humans use AI for, rise of Chinese models, and user retention secrets)

Table of Contents

Toggle

• Real-world validation: 100-hour desert island live stream
• Dialect recognition performance
• Offline version dominates simultaneously

On July 6, Alibaba Tongyi Lab announced it has reduced the first-character latency of its streaming speech recognition model, Fun-ASR-Realtime, to the millisecond level, achieving "instant character output as speech concludes," with accuracy approaching that of offline models.

The core of this upgrade lies in the enhancement of context-aware capabilities. The new model can perform dynamic error correction by combining historical conversations and real-time hot words, for example, automatically correcting "Ye Lu" to "Night Heron" based on subsequent text, effectively addressing the pain point of real-time speech recognition easily misinterpreting content out of context.

Real-world validation: 100-hour desert island live stream

During the recently concluded 100-hour desert island live stream by Film Hurricane, Fun-ASR-Realtime provided real-time subtitle support throughout, under harsh conditions of outdoor rainstorms and frequent speaker switching, recognizing a total of over 60k entries, amounting to 1.32 million characters.

This data indicates that the model maintains stable recognition quality in real-world scenarios of mono audio, high ambient noise, and overlapping speakers—a critical benchmark for real-time subtitle applications.

Dialect recognition performance

The model currently supports 30 languages and 16 dialects, achieving an average character accuracy of 88.62% in dialect tests. Performance for major dialects is as follows:

• Shanghai dialect: 92.41%
• Wenzhou dialect: 82.74% (known as the "hardest to understand" dialect)
• Average (16 dialects): 88.62%

With Wenzhou dialect, a high-difficulty item within Wu dialects, reaching 82.74% accuracy, this demonstrates a significant improvement in the model's generalization capability for low-resource dialects.

Offline version dominates simultaneously

The offline model, Fun-ASR-Flash, has claimed the top spot on the Word Error Rate leaderboard of the global AI evaluation platform Artificial Analysis, further validating the technical leadership of the Fun-ASR series in the field of speech recognition.

API services for both models have been launched on Alibaba Cloud Bailian, with the underlying open-source toolkit and models available on ModelScope and GitHub.