The Qwen large language model has officially upgraded its real-time speech recognition model to Fun-ASR-Realtime.
Fun-ASR-Realtime is a streaming speech recognition system that maintains a first-character latency at the hundred-millisecond level, with accuracy approaching that of offline models. It supports 16 distinct dialects and 30 languages.
Speech recognition models are generally categorized into offline and real-time types. Offline models process pre-recorded audio in one go, achieving higher accuracy. Real-time models transcribe audio as it's recorded, demanding both speed and precision, which makes them more challenging to develop. They are ideal for low-latency applications such as live streaming captions, real-time meeting transcriptions, and customer service calls. Fun-ASR-Realtime targets this latter category, successfully combining speed with accuracy.
Speed Performance: Minimal Latency from Start to Finish
In terms of speed, the model's first-character latency is controlled at the hundred-millisecond level, with text beginning to appear almost as soon as speech ends. For lengthy sentences, the delay for the final character output remains equally low. For instance, during the 100-hour "Return to the Deserted Island" live stream, which featured complex interactions between participants on two islands, frequent speaker switches, and challenging weather conditions, viewers could instantly see captions for speakers like Tim and China Boy, significantly enhancing the viewing experience.
Accuracy: Nearing Offline Performance with Contextual Learning
Regarding accuracy, Fun-ASR-Realtime approaches the performance of its offline counterpart, Fun-ASR-Flash. It has been specifically enhanced for general context understanding, which includes historical dialogue context and real-time hotwords. For example, during a live stream, the term "night heron" was initially misrecognized as "Ye Lu." However, after the context "bird-watching friends should know" was provided, the model immediately corrected itself.
Multilingual and Dialect Capabilities: A Key Strength
Compared to other real-time models on the market, Fun-ASR-Realtime supports a wider variety of speech, capable of recognizing 16 dialects and 30 languages. In tests covering 16 dialects from eight major dialect regions, Fun-ASR-Realtime achieved an average character accuracy rate of 88.62%, leading in 12 dialect categories and outperforming related products from Volcano Engine and Tencent. It performed exceptionally well on challenging Wu dialect varieties: Shanghai dialect at 92.41%, Suzhou dialect at 89.21%, and Wenzhou dialect (often called the "most difficult to understand") at 82.74%. The model also comprehensively led in complex dialects like Hokkien and Hakka.
Complex Scenarios and Multiple Languages: From Industrial Settings to Global Markets
In a comparative evaluation of speech recognition across five types of industrial Chinese and English scenarios—including complex backgrounds, far-field noise, and various accents—Fun-ASR-Realtime performed robustly. It achieved an average character accuracy of 88.42% for Chinese and 91.58% for English, leading among related products. Furthermore, the model has been specially optimized for multilingual scenarios in East and Southeast Asia, such as Thai, boosting recognition accuracy by 20%, making it suitable for overseas customer service and international conferences.
Offline Model Fun-ASR-Flash: Tops Global Authority Rankings
The offline model for Fun-ASR-Realtime, named Fun-ASR-Flash, was recently launched. On the globally authoritative AI evaluation platform Artificial Analysis, Fun-ASR-Flash (listed under its former name "Fun-realtime-ASR-preview") achieved a character error rate of 1.7%, ranking first in the world.
Comments