Home
Tencent Hunyuan and Partners Launch MMAE Benchmark, Revealing Under 5% Precision in AI Audio Editing

Artificial intelligence has made significant strides in audio generation, but editing existing audio remains a major challenge. Recently, Tencent Hy, in partnership with leading research institutions including Shanghai Jiao Tong University (SJTU), Nanyang Technological University (NTU), Tianjin University (TJU), Peking University (PKU), and Fudan University (FDU), introduced MMAE (Massive Multitask Audio Editing Benchmark)—the first large-scale, multi-task benchmark designed for general instruction-driven audio editing. This release establishes a systematic evaluation standard for the AI audio editing field, revealing clear gaps in the ability of current technology to perform precise modifications.
From "Generation" to "Editing": The Real Challenge for AI Audio
Traditional audio AI focuses on generating new content from text or prompts. The core of the MMAE benchmark, however, requires models to comprehend existing audio clips and make precise adjustments based on natural language instructions: modify only the parts that need changing, while leaving everything else untouched. This "editing, not reconstruction" approach demands higher audio fidelity, better instruction following, and deeper context understanding—making it far more aligned with real-world applications like podcast post-production, music mixing, and voice personalization.
Tests reveal that current mainstream models achieve an overall Exact Match Rate (EMR) below 5%, highlighting significant gaps in reliable audio editing technology. This means AI tends to over-modify, miss instructions, or degrade original audio quality when tackling practical editing tasks.
MMAE Benchmark Highlights: Multi-Dimensional Evaluation for Real-World Scenarios
The MMAE benchmark is designed with thoroughness and rigor, featuring the following key elements:
2,000 high-fidelity samples: All sourced from real-world scenarios, ensuring practical and diverse evaluation.17,741 fine-grained evaluation metrics: Provide a detailed scoring rubric for objective quantification.7 modal settings: Cover sound, music, speech, and their combinations, supporting testing in complex audio environments.6 levels of task complexity: Range from basic modifications to multi-hop reasoning and multi-round editing, enabling comprehensive assessment of model capabilities.8 types of operations: Support editing tasks at different granularities, both local and global, challenging the model's fine control.AIbase Comment: MMAE is not just a technical evaluation tool—it is a milestone in driving audio AI from "generative" to "editable." It provides a unified standard for researchers and developers, and is expected to accelerate the next generation of audio editing models.
Future Outlook: Audio Editing Could Become a Core Strength of AI Multimodal Systems
As multimodal large models evolve rapidly, precise audio editing will play a vital role in content creation, film post-production, and accessibility tools. The recent collaboration between Tencent Hy and other institutions highlights China's leading position in audio AI research. The industry anticipates more open-source resources and follow-up models to collectively fill this technological gap.
Related article
Apple iOS 27: Siri Gets Major Upgrade, New Rules Block Summary Web Links
As artificial intelligence advances rapidly, Apple has refined its privacy and ecosystem strategies in its latest operating system. The newly released iOS 27 Beta 2 update further tightens Apple Intelligence usage rules, directly restricting Siri's a
Alibaba Cloud Launches Meeo CLI, an Open-Source Tool for One-Click Deployment
Alibaba Cloud has introduced Meoo (Miao Wu), an open-source command-line tool designed to simplify the process of launching local AI programming projects. By bridging the gap between local development and cloud deployment, Meoo enables developers to
Google Unveils Veo 3.1 Lite Video Mode, Granting Ultra Subscribers Unlimited Generation
Google has introduced a new video generation mode, "Veo 3.1 - Lite [Low Priority]," exclusively for Ultra subscribers to boost creative output and cost efficiency. This mode’s key advantage is that it operates without consuming additional subscriptio
Related Special Topic Recommendations
Comments (0)
0/500

Artificial intelligence has made significant strides in audio generation, but editing existing audio remains a major challenge. Recently, Tencent Hy, in partnership with leading research institutions including Shanghai Jiao Tong University (SJTU), Nanyang Technological University (NTU), Tianjin University (TJU), Peking University (PKU), and Fudan University (FDU), introduced MMAE (Massive Multitask Audio Editing Benchmark)—the first large-scale, multi-task benchmark designed for general instruction-driven audio editing. This release establishes a systematic evaluation standard for the AI audio editing field, revealing clear gaps in the ability of current technology to perform precise modifications.
From "Generation" to "Editing": The Real Challenge for AI Audio
Traditional audio AI focuses on generating new content from text or prompts. The core of the MMAE benchmark, however, requires models to comprehend existing audio clips and make precise adjustments based on natural language instructions: modify only the parts that need changing, while leaving everything else untouched. This "editing, not reconstruction" approach demands higher audio fidelity, better instruction following, and deeper context understanding—making it far more aligned with real-world applications like podcast post-production, music mixing, and voice personalization.
Tests reveal that current mainstream models achieve an overall Exact Match Rate (EMR) below 5%, highlighting significant gaps in reliable audio editing technology. This means AI tends to over-modify, miss instructions, or degrade original audio quality when tackling practical editing tasks.
MMAE Benchmark Highlights: Multi-Dimensional Evaluation for Real-World Scenarios
The MMAE benchmark is designed with thoroughness and rigor, featuring the following key elements:
2,000 high-fidelity samples: All sourced from real-world scenarios, ensuring practical and diverse evaluation.17,741 fine-grained evaluation metrics: Provide a detailed scoring rubric for objective quantification.7 modal settings: Cover sound, music, speech, and their combinations, supporting testing in complex audio environments.6 levels of task complexity: Range from basic modifications to multi-hop reasoning and multi-round editing, enabling comprehensive assessment of model capabilities.8 types of operations: Support editing tasks at different granularities, both local and global, challenging the model's fine control.AIbase Comment: MMAE is not just a technical evaluation tool—it is a milestone in driving audio AI from "generative" to "editable." It provides a unified standard for researchers and developers, and is expected to accelerate the next generation of audio editing models.
Future Outlook: Audio Editing Could Become a Core Strength of AI Multimodal Systems
As multimodal large models evolve rapidly, precise audio editing will play a vital role in content creation, film post-production, and accessibility tools. The recent collaboration between Tencent Hy and other institutions highlights China's leading position in audio AI research. The industry anticipates more open-source resources and follow-up models to collectively fill this technological gap.
Apple iOS 27: Siri Gets Major Upgrade, New Rules Block Summary Web Links
As artificial intelligence advances rapidly, Apple has refined its privacy and ecosystem strategies in its latest operating system. The newly released iOS 27 Beta 2 update further tightens Apple Intelligence usage rules, directly restricting Siri's a
Alibaba Cloud Launches Meeo CLI, an Open-Source Tool for One-Click Deployment
Alibaba Cloud has introduced Meoo (Miao Wu), an open-source command-line tool designed to simplify the process of launching local AI programming projects. By bridging the gap between local development and cloud deployment, Meoo enables developers to
Google Unveils Veo 3.1 Lite Video Mode, Granting Ultra Subscribers Unlimited Generation
Google has introduced a new video generation mode, "Veo 3.1 - Lite [Low Priority]," exclusively for Ultra subscribers to boost creative output and cost efficiency. This mode’s key advantage is that it operates without consuming additional subscriptio











