en.Wedoany.com Reported - On July 6, China's Alibaba Qwen announced the upgrade of its real-time speech recognition large model, Fun-ASR-Realtime. After the upgrade, a single model can support 16 dialects and 30 languages. The API has been launched on the Alibaba Cloud Bailian platform in China, opening up for applications in scenarios such as real-time subtitles, meeting transcription, voice assistants, cross-border customer service, and multilingual communication.

The focus of this upgrade is on two directions: "real-time" and "complex speech coverage." As a streaming speech recognition model, Fun-ASR-Realtime emphasizes outputting text synchronously while the user is speaking. The latency for the first character is controlled at the millisecond level, and the output latency for the last character of long sentences is low. Compared to offline recognition, which transcribes uniformly after the recording ends, real-time recognition tests the model's continuous judgment ability on speech segments, context, pauses, noise, and accent variations. In scenarios like meetings, live streaming, customer service, and smart hardware, users do not wait for a sentence to fully end before needing results; the system must recognize while listening and minimize misjudgments caused by contextual changes.

In terms of dialect capabilities, Fun-ASR-Realtime covers 16 dialects from eight major dialect regions. In relevant tests, the model achieved an average character accuracy rate of 88.62% and outperformed China's Volcano-related products and China's Tencent-related products in recognizing 12 types of dialects. Dialect recognition has long been a challenge in speech technology due to significant differences in initials, finals, tones, speaking speed, liaison, and vocabulary across regions. The same Mandarin vocabulary may be altered by different accents in real spoken language. If the model is only adapted to standard Mandarin, it is prone to transcription errors when applied to scenarios like customer service, live streaming, short videos, government hotlines, in-vehicle voice, and smart terminals. By integrating 16 dialects into the same real-time model, Alibaba Qwen means that enterprises do not need to deploy multiple recognition capabilities separately for different regions.

Multilingual capabilities point to broader commercial use scenarios. Fun-ASR-Realtime supports 30 languages and has undergone specialized optimization for East Asian and Southeast Asian multilingual scenarios, such as Thai, improving recognition accuracy by 20%. Such capabilities are more suitable for cross-border customer service, international conferences, overseas live streaming, multinational corporate training, travel services, international e-commerce, and localized operations. For companies going global, speech recognition is not just about converting sound to text; it also relates to customer service quality inspection, ticket generation, meeting minutes, subtitle translation, sales lead organization, and compliance record-keeping. If a real-time model can simultaneously handle multilingual and accent issues, enterprises can reduce manual transcription and repetitive organization costs in their overseas operations.

This upgrade also places greater emphasis on contextual capabilities. Real-time speech recognition is easily affected by environmental noise, multiple speakers, specialized vocabulary, and temporary hot words, especially in scenarios like live streaming, industrial sites, medical consultations, financial customer service, and technical meetings, where names, brand names, product models, and device names are often not common words. Fun-ASR-Realtime has been enhanced with broad contextual capabilities, leveraging historical dialogue context and real-time hot words to improve recognition results. For example, in a live streaming scenario, the model can recalibrate easily misheard words based on preceding and following semantics, making subtitles closer to the actual expression. This capability is closer to actual business needs than simply improving Mandarin dictation scores.

After the API was launched on the Alibaba Cloud Bailian platform in China, developers and enterprises can integrate Fun-ASR-Realtime into their own business systems. The integration method will affect the speed of model deployment: if speech recognition capabilities can only be used in a single product, it is more like a tool function; if it enters customer service systems, meeting systems, smart hardware, live streaming platforms, in-vehicle systems, and enterprise office platforms via API, it becomes a foundational speech infrastructure. Fun-ASR-Realtime forms a combination with the recently launched offline mode model Fun-ASR-Flash, where the former is suitable for real-time subtitles and voice interaction, and the latter is more suitable for recording files, long audio, interviews, and batch transcription tasks. The simultaneous availability of both real-time and offline models on the Bailian platform allows enterprises to choose different recognition paths based on the scenario.

The competition in speech recognition models is shifting from "being able to understand Mandarin" to "being able to understand complex sounds in the real world." Users may have accents, pause, repeat, insert English words, switch languages, or be in environments like meeting rooms, live streaming rooms, factories, cars, or outdoors. By integrating dialects, multilingual capabilities, low latency, contextual hot words, and API services into a single upgrade, Fun-ASR-Realtime is targeting a more complex enterprise-level voice entry point. How many industry-specific vocabularies it can cover, whether it can maintain low latency under high concurrency calls, and whether it can adapt to more noisy scenarios will directly determine the depth of use of such real-time speech models in commercial systems.