Skip to main content

Aliyun's New Speech Recognition Model Tackles Long Meetings and Industry Jargon

On July 31, Alibaba's Tongyi Qianwen team unveiled its latest speech recognition model, Qwen-Audio-3.0-ASR, now available on the Alibaba Cloud Bailian platform. The model aims to solve two persistent headaches in speech recognition: keeping track of context in long recordings and recognizing niche industry terms without requiring users to train the system.

For anyone who has struggled with transcription tools that lose the thread in a two-hour meeting or mangle medical jargon, this update might be welcome news. The new model comes with several key upgrades that target these pain points directly.

Long Audio, No Lost Context

One of the standout features is its ability to remember context across hours of audio. Instead of treating each segment in isolation, the model references earlier parts of the conversation, ensuring that names and technical concepts stay consistent throughout. This eliminates the frustrating "fragmentation" problem where a speaker's name changes halfway through the transcript or a term is misheard repeatedly.

Built-In Industry Dictionaries

Another major improvement is the inclusion of industry-specific dictionaries. The model boasts a recall rate of 95.36% for medical terminology and 91.87% for IT programming terms. That means it can accurately identify obscure words like "pneumonoultramicroscopicsilicovolcanoconiosis" or "asynchronous JavaScript" without any manual configuration. For businesses in specialized fields, this could save countless hours of editing.

Customizable Hot Words

For enterprise users, the model offers tiered hot word customization. Company-specific vocabulary takes effect immediately, and in most scenarios, the recall rate exceeds 99%. Importantly, adding more hot words doesn't lead to false triggers, a common issue with other systems.

Voice Polishing Built-In

The model also integrates voice polishing, which cleans up filler words, removes repetitions, handles self-corrections, and reorganizes sentences semantically. The output is said to be nearly as readable as the two-step process of using ASR followed by a large language model for polishing. This could be a boon for content creators and journalists who rely on clean transcripts.

Multilingual Support

Covering more than 30 languages, the model performs strongly across the board. In tests, it achieved an average semantic error rate of 17.09% across seven languages, outperforming industry models like Azure and Gemini. This makes it suitable for international meetings and overseas customer service scenarios.

Real-Time Streaming Version

Alongside the main model, Alibaba also released Qwen-Audio-3.0-ASR-Streaming, designed for real-time applications. It offers a theoretical character output delay of just 300 milliseconds, with a character error rate of 7.80% in Chinese industrial scenarios and 11.52% in English. These figures lead the industry, according to the company.

The Qwen-Audio series has previously ranked first globally in the Artificial Analysis evaluation with a 1.7% error rate. It is already being used in areas like meeting minutes, real-time subtitles, educational recordings, and intelligent customer service.

With these enhancements, Alibaba is positioning itself as a strong contender in the speech recognition space, particularly for professional and enterprise use cases. Whether you're a doctor dictating notes, a developer recording a code review, or a journalist transcribing interviews, this model might just make your life a little easier.

Key Points

  • Long audio context memory: Maintains consistency across hours of audio, avoiding fragmentation.
  • Industry-specific dictionaries: High recall rates for medical (95.36%) and IT (91.87%) terminology.
  • Tiered hot word customization: Immediate effect with over 99% recall in most scenarios, no false triggers.
  • Integrated voice polishing: Cleans up filler words and repetitions, producing readable content.
  • Multilingual support: Over 30 languages, outperforming Azure and Gemini in semantic error rate.
  • Streaming version: 300ms delay, leading error rates in Chinese and English.