Skip to main content

Aliyun's New Speech Recognition Model Tackles Professional Jargon

On July 31, Alibaba's Tongyi Qwen team officially released its latest speech recognition model, Qwen-Audio-3.0-ASR. This new model is designed to tackle two persistent headaches in speech recognition: keeping track of context in long recordings and correctly identifying specialized jargon without requiring users to train it on their own vocabulary. It's now live on Alibaba Cloud's Bailian platform, ready for developers and businesses to try.

Five Key Upgrades

The model brings five major improvements that set it apart from its predecessors.

1. Long Audio Context Memory

One of the biggest challenges in transcribing hours-long meetings is that the model can lose track of what was said earlier, leading to inconsistencies in names or technical terms. Qwen-Audio-3.0-ASR addresses this by referencing previous context throughout the transcription. This means that if someone mentions a project name early in a meeting, the model will keep using that same name later, avoiding the 'fragmentation' problem where different segments of the audio seem disconnected.

2. Built-in Industry Dictionaries

Gone are the days of manually adding every obscure term. The model comes with pre-loaded dictionaries for specific fields. For medical terminology, it achieves a recall rate of 95.36%, and for IT programming, it hits 91.87%. That means it can accurately recognize complex, niche words without any extra setup—a real time-saver for professionals in these industries.

3. Tiered Hot Word Customization

For companies that have their own unique vocabulary, the model allows for hot word customization. These enterprise-specific terms take effect immediately, and in most scenarios, the recall rate exceeds 99%. Importantly, adding more hot words doesn't trigger false positives, so you don't have to worry about the model mishearing common words just because you've added a bunch of custom terms.

4. Integrated Voice Polishing

Transcriptions often come out messy, full of filler words like 'um' and 'uh', repetitions, and self-corrections. This model cleans all that up in one go. It removes fillers, tidies up repetitions, handles self-corrections, and even reorganizes the content semantically. The result is a polished transcript that reads almost as well as if you'd run it through a separate large language model for editing—but without the extra step.

5. Multilingual Support

Covering more than 30 languages, this single model is built for international use. In tests, it achieved an average semantic error rate of 17.09% across seven languages, outperforming industry rivals like Azure and Gemini. That makes it a solid choice for international meetings and overseas customer service scenarios.

Real-Time Streaming Version

Alongside the main model, Alibaba also introduced Qwen-Audio-3.0-ASR-Streaming, designed for real-time applications. It has a theoretical character output delay of just 300 milliseconds, which is nearly instantaneous. In Chinese industrial scenarios, it posts a character error rate of 7.80%, and for English, it's 11.52%—both leading the industry. This series previously ranked first globally in the Artificial Analysis evaluation with a 1.7% error rate, and it's already being used in areas like meeting minutes, real-time subtitles, educational recordings, and intelligent customer service.

Key Points

  • Long audio context memory ensures consistent transcription across hours of content.
  • Built-in dictionaries for medical and IT fields achieve high recall rates without manual configuration.
  • Hot word customization delivers over 99% recall in most scenarios, with no false triggers.
  • Integrated voice polishing cleans up transcripts in one step, rivaling two-step ASR+LLM approaches.
  • Multilingual support covers 30+ languages, outperforming Azure and Gemini in semantic error rate.
  • Real-time streaming version offers 300ms delay and industry-leading error rates.

With these capabilities, Qwen-Audio-3.0-ASR is poised to make professional transcription more accurate and efficient, saving time and reducing frustration for users across various industries.