传音TEX AI语音算法团队斩获MLC-SLM 2026国际挑战赛Task 1亚军

This article shares the key takeaways and technical insights from Transsion's TEX AI Center speech algorithm team, which won second place in Task 1 of the MLC-SLM 2026 International Challenge, a top authoritative competition in the global speech technology field.

1. A total of 110 teams from around the world participated in the challenge. The task required participants to process unsegmented, unannotated long-form real multi-speaker conversation audio, and output transcription results with speaker information. Evaluation covers accuracy, timestamp alignment and speaker attribution, making it an extremely challenging task. Transsion secured second place globally with a tcpMER of 15.41%.

2. To address the challenge's difficulties, the team did not just optimize a single model. Instead, it built a three-stage cascaded system with two separate pipelines to handle speaker diarization and high-accuracy transcription respectively, before fusing the outputs to generate the final result. This approach enabled full closed-loop optimization across the entire processing workflow.

3. This technology is already deployed in multiple scenarios across Transsion's smart devices, and will be expanded to more application areas in the future.

This article offers valuable actionable insights for consumer electronics brands targeting emerging overseas markets, particularly for product R&D and building technological competitiveness.

1. Product R&D guidance: The language environment in emerging markets is highly complex, and multi-lingual voice interaction is a core pain point for user experience. Proactive investment in multi-lingual speech recognition and interaction technology allows brands to build a differentiated experience advantage. Transsion has already built technical capabilities covering multiple languages, local accents and dialect variants, and has validated its value through deployment across multiple device scenarios.

2. R&D strategy guidance: Brands do not need to limit themselves to a single technical route. Transsion adopted a parallel exploration approach combining cascaded systems and end-to-end solutions, iteratively optimizing performance through incremental experiments, reducing tcpMER from 22.78% to 15.41%. This steady, methodical R&D approach is highly replicable.

3. Brand marketing guidance: Top results at leading international competitions serve as strong third-party validation of a brand's technical strength, and can be leveraged in brand marketing to reinforce professional positioning.

This article provides useful reference insights for sellers building consumer electronics businesses in overseas markets, covering opportunity assessment and capability building.

1. Opportunity outlook: A clear breakout point for differentiated competition in emerging overseas markets has emerged: high-quality voice interaction experiences in complex multi-lingual environments remains an unmet user need. Proactive investment in related capabilities can deliver significant competitive advantages.

2. Actionable lessons: For multi-lingual speech technology R&D, avoid limiting work to single-model optimization. You can follow Transsion's approach of parallel exploration of dual technical routes, full closed-loop workflow refinement, and incremental iterative optimization to reduce R&D trial-and-error costs while steadily improving performance.

3. Risk warning: As smart devices evolve, voice interaction capability has become a core competitive differentiator. Sellers without relevant technical investment will likely fall behind in future product competition. Brands should either invest in independent R&D proactively, or partner with supply chain players with mature technology to fill this capability gap.

This article offers valuable insights for consumer electronics contract manufacturers serving overseas markets, and factories building their own end-device businesses.

1. Product design and manufacturing guidance: For smart devices exported to emerging markets today, mature multi-lingual voice interaction capability has become a clear required specification. Factories need to include multi-lingual speech recognition and interaction capabilities as hard requirements during the product design and manufacturing preparation stage to meet brand clients' demands.

2. Business opportunity outlook: Deploying multi-lingual speech technology requires hardware integration and adaptation. Factories with existing hardware integration capabilities can partner with AI technology companies and brands looking for technology output partnerships to secure more overseas-related orders and expand their business scope.

3. Digital R&D inspiration: For factories advancing technological upgrading and digital R&D, Transsion's approach of starting from a solid baseline and incremental iterative optimization is a useful reference. This method allows teams to steadily refine capabilities, reduce R&D trial-and-error costs, and steadily improve technical standards.

This article clarifies industry direction and provides deployment reference for AI voice technology service providers and smart device technology supporting service providers.

1. Industry development trend: Multi-lingual real-scenario speech processing has become the core R&D direction for the industry. The international community also prioritizes full-pipeline speech technology for real multi-speaker conversation scenarios, and related demand will continue to grow.

2. Core client pain points: Overseas brands targeting emerging markets need full-pipeline voice solutions that can handle unannotated long-form audio, speaker diarization, and multi-lingual transcription. Technical advantages in a single link cannot meet real-world deployment needs, and full-pipeline collaborative capability is what clients actually require.

3. Reference solution: You can adopt Transsion's cascaded system approach, which uses two separate pipelines to handle speaker diarization and high-accuracy transcription respectively before fusing outputs; meanwhile, pursuing parallel R&D for both cascaded and end-to-end architectures reserves space for future technical iteration, and allows adaptation to different business scenario requirements.

This article offers useful insights for AI technology service platforms and overseas consumer electronics recruitment platforms on platform operation and strategic layout.

1. Merchant demand insight: Currently, the majority of overseas brands on these platforms have demand for full-pipeline multi-lingual speech technology. Platforms need to proactively build a supply of relevant technology merchants to meet client demand and improve platform competitiveness.

2. Recruitment direction guidance: Platforms can prioritize onboarding enterprises like Transsion that have mature full-pipeline deployment capability for multi-lingual speech technology. This enriches the platform's technology service categories, attracts more overseas-focused brand clients, and expands platform scale.

3. Risk mitigation guidance: Deploying multi-lingual speech technology relies heavily on full-pipeline collaborative capability, and advantages in a single technical point cannot meet client deployment needs. When onboarding technology providers, platforms should prioritize evaluating vendors' full-pipeline engineering deployment capability to avoid onboarding immature technology that harms platform reputation and creates operational risks.

This article presents the latest industry and academic developments in multi-lingual speech technology, and offers high reference value for both industrial and academic researchers in the AI speech field.

1. New field trends: In the multi-lingual conversational speech model space, the core research challenge has shifted to real-world deployment problems. This international challenge took processing unsegmented, unannotated real long-form conversations as a core track, with full-pipeline algorithm collaboration as the core evaluation dimension. The track drew 110 teams globally, reflecting its high research momentum.

2. Technical route reference: Transsion's three-stage cascaded system architecture uses separate pipelines to recover speaker structure and complete multi-lingual transcription, before fusing outputs to generate final results. This approach delivered strong performance, while parallel exploration of both cascaded and end-to-end routes reserves room for future technical evolution. This R&D approach is valuable for researchers to reference.

3. Industrialization direction reference: Multi-lingual speech technology has already achieved large-scale deployment in smart device scenarios targeting emerging markets, where it solves real user pain points and delivers clear commercial value. This clarifies the industrialization direction for related technologies.