赛事直播中的关键瞬间，往往只在几秒之间。泳池里，一次转身就可能改写领先关系；靶场上，一枪成绩就可能决定奖牌归属。解说需要紧跟画面，讲清赛场态势、项目规则、排名变化和比赛走势。比赛结束后，媒体还要第一时间找到精彩瞬间，完成高光剪辑并进入发布流程。
在 2026 年名古屋大型赛事上，央视体育与阿里云 Flink 合作打造大型赛事实时 AI 解说 Agent。它在直播过程中持续完成画面理解、赛事判断、上下文记忆、解说生成和音视频处理，让 AI 跟随比赛“实时看、持续记、及时说”，并支持赛事高光生产。
阿里云 Flink 与千问大模型构建可靠的实时解说流水线
大型赛事实时 AI 解说 Agent 采用事件驱动的流式架构。直播画面、官方赛事数据与赛程信息持续进入系统，有效变化更新比赛状态，并触发解说或剪辑任务。依托 Qwen3.8-Max、Qwen3.8-Flash，以及语音合成模型 qwen-audio-3.0-tts-plus 等理解与生成能力，阿里云 Flink 将信号接入、画面感知、赛事理解、任务策划和执行连接为持续运转的流水线。
大型赛事直播对这条流水线提出了两项要求：成绩、名次和奖牌归属有据可依，数小时的连续运行保持稳定。为此，多源确认、事实校验和故障恢复贯穿各个环节，让生成内容具备可信依据，也让比赛上下文能够持续积累和恢复。
赛事信号接入与画面感知
系统持续接收赛事直播视频流、赛程与参赛名单，以及官方实时赛事事件。视频经过解析、格式标准化、关键帧采样与时间轴归一，形成带时间信息的画面帧；官方成绩、名单和赛程作为结构化数据进入关联处理。Flink 提供持续处理和事件时间组织能力，业务链路据此建立媒体时间与赛事时间的映射，明确当前画面属于哪场比赛、成绩对应哪位运动员和哪个轮次，为音画对齐与高光定位打下基础。
赛事观察插件进一步提取运动员、动作变化、比分字幕和成绩图卡等信息。射击场景关注选手姓名、单枪环数与累计成绩，游泳场景关注道次、转身、交接、触壁和分段计时。感知结果带有时间与来源标记，经由 Flink 持续流转，供后续环节综合判断。
多源理解与持续更新的比赛上下文
从识别画面中的数字到确认一项成绩，还需要结合官方数据、参赛名单、比赛阶段和项目规则。理解插件据此识别关键事件，区分当前成绩与历史回放、分段领先与正式赛果，并更新比分、排名和比赛进程。名次、成绩和奖牌归属以官方确认信息为依据，证据不足的结论暂不进入播报。
Flink 的状态管理承载不断演进的比赛上下文，包括当前阶段、领先选手、逐枪成绩、排名变化和已确认事实。新事件到达后，系统增量更新相关状态；平均值、分差等数值由结构化计算给出，再交给模型组织分析。这既为实时解说提供前情，也让赛后回顾和高光标题能够追溯到具体赛事信息，减少数据错配、重复报分和前后表述不一致。
动态策划与经过校验的内容执行
Agent 根据事件重要性、事实可靠性、比赛节奏与可用播出时长，决定此刻讲什么、做什么。关键成绩和奖牌归属优先播报，比赛间隙补充规则与选手背景；冲刺阶段使用简短赛况，轮次间隙再展开分析。例如，射击比赛中识别到铜牌产生后，Agent 会从项目科普转向奖牌播报。高光策划则确定哪些事件值得成片，以及需要保留的前后过程。
解说任务依次经过文稿生成、事实与数字校验、内容审核、语音合成和播出编排。选手姓名、成绩、排名及结果表述在输出前再次核对，待播队列依据优先级与有效期调整，撤回过时内容。同一份赛事事实可生成三种风格的文稿与音轨。高光任务则完成视频边界确认、裁剪拼接、封面选择、标题生成和成品上传，保持标题与画面事实相符。
持续处理与故障恢复
连续直播中，模型推理和语音生成的耗时波动可能造成等待与积压。链路可结合异步处理、并发控制和超时策略管理外部服务调用，并通过播出队列调整减少解说滞后。
Flink 的检查点机制为受管理的计算状态提供一致性快照，配合可恢复的数据源，在作业重启后恢复比赛上下文并继续处理。已经播出的解说和已经上传的高光，还需要由执行链路记录动作状态，结合幂等或去重机制控制重复输出。
主备切换、监控告警、模型服务防护与人工值班共同构成运行保障。团队持续关注处理积压、模型响应时间、解说输出延迟和作业恢复情况，及时发现并处置异常。Flink 的状态恢复能力与模型、媒体组件和运维协同，支撑 Agent 长时间稳定进入赛事生产流程。
实时看、持续记、及时说，央视体育携手阿里云 Flink 打造大型赛事实时AI解说Agent
解说跟随直播节奏，三种风格随时切换
系统利用直播流安全审核产生的时间窗口，优化画面理解、文稿生成与语音合成的全流程耗时，使 AI 解说与 CCTV5 等直播流基本同步，让线上观众几乎感受不到额外延迟。观众还可在央视体育客户端中随时切换“低音沉稳”“高能爆发”“清亮温和”三种解说风格。
实时播报成绩，跟进关键结果
射击比赛中，Agent 将选手、枪次、单枪成绩与累计成绩准确关联，成绩有效变化后再触发播报，避免重复报分。游泳比赛中，Agent 围绕出发、转身、交接和触壁跟踪分段成绩与位次变化，讲清谁在领先、谁在追赶，并区分阶段排名与最终赛果。
记住比赛前情，讲清多轮表现变化
射击前十枪结束后，Agent 可以比较最近五枪与此前五枪的平均成绩，结合累计成绩、排名和分差分析选手如何追赶。比分接近时，还可结合赛制提示淘汰风险，让解说从单次报分延伸到连续比赛分析。
在比赛间隙穿插规则与知识
Agent 根据项目和比赛阶段选择知识内容：赛前介绍轮次安排，击发间隙讲解呼吸控制与持枪稳定性，临近淘汰时说明场上局势。奖牌产生后及时转向关键赛况，让科普自然融入直播节奏，帮助首次观看该项目的观众理解比赛。
结合历史资料与实时表现讲述运动员
以徐嘉余的男子 50 米仰泳决赛为例，Agent 在出场时介绍道次、预赛成绩和历史表现，赛中跟踪成绩与名次变化，赛后及颁奖阶段结合本场结果回顾比赛经历。知识检索补充人物资料，实时上下文连接当场表现，形成连贯的人物讲述。
关键事件驱动高光拆条
关键得分、反超、淘汰或夺冠等事件确认后，高光 Agent 启动剪辑任务，保留事件的完整过程，完成裁剪拼接、封面选择、标题生成和上传。高光链路可在精彩事件发生后一分钟内完成片段合成与上传，产出供编辑审核发布的成品。
AI 融入赛事，实时计算连接赛事现场与内容生产
大型赛事项目多，赛程密集，阿里云Flink团队与政企事业部文化传媒行业线团队通力合作，让央视体育解说覆盖海量赛事，为技术团队带来新解法，大幅拓宽赛事解说的覆盖范围。对观众而言，规则科普、运动员故事和多种声音风格丰富了观赛选择，也为冷门项目和多路直播获得完整解说提供了可能。
对媒体而言，实时解说与事件驱动的高光生产让内容从“赛后加工”走向“赛中即产”。本次实践中，两条链路按业务要求分别接入视频源，编辑团队负责审核把关、深度策划和创意表达，形成 AI 生产与专业团队协同的工作方式。
这次实践也呈现了实时计算与大模型协作的路径：大模型提供理解与表达能力，Flink 持续组织实时数据、维护随事件演进的上下文，并提供可恢复的计算状态。Agent 因而能够从单次问答走向持续感知、实时决策和任务执行。
未来，实时 AI 解说能力可进一步延伸到更多运动项目、语种和赛事场景。央视体育与阿里云也将继续探索实时计算与大模型的融合，让赛场瞬间被及时看见、准确讲述和长久留存。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The key moments in the live broadcast are often only in a few seconds. In the pool, the lead relationship may be rewritten at a turn, and on the range, the score may determine where the medal is. The story needs to be closely followed by a picture of the race scene, the rules of the project, changes in rankings and trends in the game. After the competition, the media also needed to find a good moment to finish the high-light clip and enter the publish process.
In 2026, when sport and Alibaba Cloud Flink worked together to create the fact of the big game, AI explained Agent. During the live broadcast, it continues to complete visual understanding, game judgement, context memory, narrative generation and audio and video processing, allowing AI to follow the competition “real-time, continuous, timely” and to support the high-profile production of events.
Alibaba Cloud Flink constructs reliable real-time discourses with LLM Line
AI explains that Agent uses an event-driven fluid structure. Live video, official race data and course information are kept in the system, effectively changing the state of the game and triggering a speech or editing mission. Alibaba Cloud Flink connects signals, visual perception, race understanding, mission planning and execution of water lines connected to continuous operation, based on the ability to understand and generate Qwen3.8-Max, Qwen3.8-Flash, and qwen-audio-3.0-tts-plus.
The live broadcast of the major competitions made two demands on the line: results, titles and medals are well-founded, and several hours of continuous operation remain stable. To this end, multiple sources of identification, fact-checking and failure recovery cut across the chain, providing a credible basis for generating content and allowing the continuous accumulation and recovery of context.
Race signal access and video perception
The system continues to receive live video streaming of events, lists of events and entries, and official real-time events. Videos are analyzed, format standardized, key frame sampling combined with time axis, creating a frame with time information; official scores, lists and games are processed as structured data. The Flink provides continuous processing and event time organization capability, whereby business links create a mapping of media time and event time, identifying which competition the current picture belongs to, which athletes and which rotations the current picture corresponds to, and laying the foundation for the alignment of sound and light.
The game observation plugin further extracts information on athletes, movement changes, score subtitle and achievement cards. The shooting range is focused on the names of the contestants, the number of single-gun rings and cumulative results, and the swimming area is focused on paths, turns, encounters, bumps and fractions. Perception results are marked with time and source, and flow through the Flink on a continuous basis for subsequent comprehensive diagnosis.
Multisource Understanding and Continuous Update of context
From the identification of the figures in the picture to the identification of an achievement, there is also a need to combine official data, lists of contestants, competition stages and project rules. It is understood that the plugin identifies key events, distinguishes between current achievements and historical replays, takes the lead from official results, and updates the score, ranking and competition process. The attribution of titles, scores and medals is based on officially confirmed information, and the conclusion that there is insufficient evidence is not yet available.
Flink's status management carries the evolving context race, which includes the current phase, leading players, gun-by-gun scores, changes in ranking and confirmed facts. Upon the arrival of a new event, the system updates the relevant state; averages, differentials, etc., are given by a structured calculation and are then presented to the model organization analysis. This both provides for real-time interpretation and allows post-race review and high-light headings to be traced back to specific event information, reducing data misalignment, repetition and inconsistent presentation.
Dynamic planning and verified content execution
Agent decides what to say and what to do at this moment, based on the importance of the event, the reliability of the facts, the rhythm of the game and the time available for broadcast. Key achievements and medals are given priority coverage, complementary rules for the margin of the competition and the background of the player; short races are used during the sprint phase, and analyses are carried out in the secondary gap of the cycle. For example, after the copper medals were identified in the shooting competition, Agent would move from the project to the medal show. High-light planning determines which incidents are worth making, and the processes that need to be preserved.
The task of interpretation is generated by scripts, facts and figures verified, content verified, voice synthesis and broadcast. The names, achievements, rankings and presentation of the results of the contesters are checked again before the output, and the waiting list is adjusted according to the priority and the validity period to withdraw the obsolete content. The same race fact produces three styles of script and tracks. The high-light mission completes video boundary confirmation, tailoring, cover selection, title generation and uploading of the finished product, keeping the title consistent with the facts of the image.
Ongoing disposal and failure recovery
Time-consuming fluctuations in model reasoning and voice generation may result in waiting and backlogs in continuous live broadcasts. The link combines walk-by processing, co-processing control and timeout strategy to manage external service calls and to reduce narration lags through broadcast queue adjustments.
The Flink checkpoint mechanism provides consistent snapshots of the managed computational state, aligns with the recoverable data source and resumes the context competition after restarting operations and continues to process it. The narratives that have been aired and the high light that has been uploaded also need to record the operational status of the implementation link, combining the thiram, etc., or weight mechanisms to control the duplicate output.
Operating security is provided by the master switch, surveillance alerts, model service protection and manual duty. The team continued to focus on the disposal of backlogs, model response times, late interpretation and operation recovery, and timely detection and disposal of anomalies. Flink ' s state-rehabilitation capability, in synergy with models, media components and transport, supports Agent ' s long and steady entry into the race production process.
Look real time, keep recording, and say in time, AI explains Agent when sport works with Alibaba Cloud Flink to build the reality of the big game.
It follows the rhythm of live broadcasts. Three styles change at any time.
The system uses the time window generated by the live stream security audit to optimize the full flow of image understanding, script generation and speech synthesis, and to keep AI ' s interpretation basically in line with live streaming such as CCTV5, leaving the online audience almost without any additional delay. Audiences can also switch the three typologies of “low-silent” “high-energy burst” “light-temperature” from the viewing sports client at any time.
Report on achievements in real time and follow up on key results
During the shooting competition, Agent linked the players, the number of shots, the individual scores and the cumulative scores accurately, and the results were effectively changed before triggering the broadcast and avoiding repetition of the scores. In the swimming competition, Agent tracks the scores and rankings around departures, turnovers, intersections and bumps, telling who's leading, who's chasing, and distinguishing between the stages and the final results.
Remember before the game. Tell me about multiple rounds of change.
After the first 10 shots, Agent can compare the average of the last five with the previous five, and how to catch up with the cumulative scores, rankings and differential analyses. When the score is close, it can also be combined with a game-based hint of phase-out risk, allowing explanations to be extended from a single score to a continuous competition analysis.
Put rules and knowledge in the middle of the game.
Agent selects the content of the knowledge according to the project and competition phases: the rotational arrangement before the race, the interpretation of respiratory control and gun stability in the break-up, and the situation on the ground as it approaches phase-out. The medals were created in time for a shift to key events, allowing Cope to naturally integrate into the live rhythm and helping viewers to understand the competition for the first time.
I'll tell you about athletes with history and real-time performance.
For example, in the 50-metre-long swim finals of Xu Kazu, Agent presented the troupe, the results and the historical performance, followed the results and the changes in titles during the competition, and reviewed the course of the competition in the context of the results at the post-election and award stages. Knowledge retrieval complements personal data, and real-time context links on-the-spot, forming coherent personality narratives.
Critical event driven high light strips
Upon confirmation of events such as key scores, counter-supers, phase-out or win, Highlight Agent initiates the editing mission, preserves the event in its entirety and completes the tailoring, cover selection, title generation and upload. The high-light link can be synthesized and uploaded within one minute of the dramatic event, with the output for editing and validation of the publish product.
AI Integration, real-time calculation of the connection to the event site and content production
The Alibaba CloudFlink team works in partnership with the Ministry of Political Affairs and Enterprise ' s cultural and media line team in order to allow sports to cover big events, bring new solutions to the technical team and significantly broaden the coverage. For the audience, rules of procedure, athletes' stories and a variety of sound styles enrich the game options and make it possible to get a full explanation for cold-door projects and multiple-road live broadcasts.
For the media, real-time interpretation and event-driven high-light production allows content to move from “post-game processing” to “immediate production”. In this exercise, two links are connected to the video source, as required by the business, and the editorial team is responsible for reviewing clearances, in-depth planning and creative expression to form the AI production and professional team's working methods.
The exercise also presented the path of real-time calculation of collaboration with LLM: LLM provides the ability to understand and express, Flink organises real-time data on a continuous basis, maintains context as events evolve and provides a state of calculation that can be restored. Agent has thus been able to move from a single question and answer session to continuous perception, real-time decision-making and mandate implementation.
In the future, real-time AI interpretation can further extend to more sports, languages and games. ZO and Alibaba Cloud will also continue to explore the integration of real-time calculations with LLM, so that the arena is immediately visible, accurately described and permanently preserved.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.