炒股就看金麒麟分析师研报,权威,专业,及时,全面,助您挖掘潜力主题机会!
(来源:北京商报)
全文共1561字,阅读大约需要3分钟
10月22日下午,科大讯飞1024开发者节主论坛在北京国家会议中心二期举行。在正式场合,一份会议记录能否使用,标准不在设备参数,而在后果:记串一个人的代价,是不是显著高于日常。会上,科大讯飞董事长刘庆峰介绍了新发布的星火语音基座大模型SparkAudio 1.0,随后播放了一段用讯飞办公本拍摄的演示视频。据介绍,这一版本将于10月24日上线,同步向讯飞办公本X5与Air3推送更新,使用者升级后即可感受。
这项工艺给办公本带来的变化,说到底只有一件事:它开始分得清谁在说话。SparkAudio 1.0基于全国产算力训练,新拓展的能力里有一项叫"实时说话人分离"——机器不只要听清说了什么,还要判断这句话是谁说的。
工艺要起作用,得有个载体。讯飞办公本X5就是这个载体:它会跟着人进会议室,摊在桌上,一边听一边把话变成字。这件事本身不难;难的是字落下来之后,还知道那句话是谁说的。
1
一场会里,最难的往往不是听清
过去几年,AI硬件发布会的常见做法是列参数、比跑分、公布准确率。这些指标衡量的都是"听清"——同样的声音,能不能一字不差地转成文字。可一场真实的会议,难处很少在听清。同一句话从不同的人嘴里说出来,分量并不相同;抢话的十几秒里,往往就藏着最关键的分歧。
会议记录过去只做完两件事:把声音留下来,把声音变成字。这两件应对的都是留存;再往前一步是归属——同一句"这个思路我不赞成",出自决策层与出自执行层,含义并不相同。分角色识别要应对的正是这一层。演示里的办公本在三个地方把它往前推了一截。
名字是这样挂上去的。录音转写时,设备先把声音分开,标成"说话人1""说话人2";点开其中一个,把名字写上去。这个动作相当于替这个人做了一次声纹注册——设备记下的不只是三个字,还有这个声音的特征。之后他再开口,记录里浮现的就是名字,不再是编号。
省掉的是会前那一套准备。过去要让设备认人,得提前把参会者挨个叫过来,对着机器说一段话;十几个人的会,光这一项就要占掉不少时间,临时加座、中途进来的人往往就漏掉了。现在顺序倒过来:边开边认,认出来再补名字。名字记一次就够,下一场他推门进来,设备认的还是这个名字。
更难的是抢话。真实会议不是轮流发言,两个人同时开口、一个人没说完另一个就接上,都是常态。普通转写遇到这种片段,会把两路声音糊成一段连续文本,事后谁也分不清哪半句是谁的;有的干脆把重叠的那几秒丢掉,记录上留一个空白。
办公本做的是按人拆开。重叠的那几秒里,先判断有几路声音、各自是谁,再分别归位——谁先起的话、谁接的、哪半句是插进来的,都能落到对的人头上。这一条决定的不是准确率,而是一份记录散会后能不能直接用。会议里最要紧的那句话,往往就发生在抢话的几十秒里。
还有语言上的混。中英文夹着说、行业术语直接念,是专业会议的常态。财务评审里"商誉减值"若被记成"商业减资",这句话就失去了原来的意思;工艺评审里型号少一个字母,指向的就是另一个机型。错的那个字不会报警,它会一直躺在记录里,直到有人拿它当依据。
这些词可以提前准备,也可以边用边学。会前把常被念错的词手写一遍,抬笔之后当前页面的识别结果会重新处理一次;术语兼容批量导入,不必逐个添加。更关键的是记过的词留得住:这场校正过的说法,下一场仍在候选里。术语与说话人一起归入"记住的信息",识别时作为上下文调用——词要先认对,人才能认准。
2
三件事合起来,是一份能直接用的记录
一场真实的会议里,这三件事往往同时发生:有人在抢话,有人在说英文缩写,有人被叫了外号。设备要在同一段时间里,把声音分开、把名字对上、把词认准,最后落成一份谁说的、说了什么、什么时候说的都清楚的记录。
散会后不用回听,不用逐句改错字,也不用猜"说话人3"是谁;三个月后有人问"当初是谁提的这个思路",翻到那一段就有答案。对使用者来说,变化直接体现在动作上:过去散会后要花半小时整理,现在打开就是成稿;过去归档前要人工核对署名,现在署名在记录生成的那一刻就已经在了。
3
这个品类,是讯飞从零做起来的
2019年讯飞推出第一代智能办公本时,切入的正是专业会议记录。据洛图科技统计,在线上主流渠道的办公类电子纸机型中,科大讯飞已连续四年市占率居首,累计销量突破200万台。七年下来,被市面接受的是一套标准——设备能不能进正式场合,看的不是它记了多少,而是它记的能不能被认领。
4
衡量AI的尺子,正在换一把
工艺成熟的标志,向来不在上限有多高,而在下限有多稳:上限决定一场演示能做到多漂亮,下限决定一件事能不能交给它去办。办公情形的AI,前几年主要在应对生成——错了可以改,改完不留痕迹,后果由使用者自己承担。会议记录不属于这一类,它的产出是一份可以被引用、被质疑、被追责的材料,标准不会因为换成机器而降低,反而更严。
标准最终要落到一件具体的事上:浮现争议时,能否翻到那一句、找到那个人。这也是讯飞办公本在这类场合被反复提起的原因——它回答的难题是"记的是谁",不是"记了多少"。
正式场合,更可靠——理由不在于记录更多,而在于记录可用。
图片丨企业供图
本文来源:https://finance.sina.com.cn/stock/t/2026-10-23/doc-iniwctax5107563.shtml
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
For stock trading, refer to Jinqilhu analysts' research reports—authoritative, professional, timely, and comprehensive, helping you uncover potential thematic opportunities!
The full text is 1,561 characters and takes approximately 3 minutes to read.
On the afternoon of October 22, the main forum of iFlytek's 1024 Developer Festival was held at the Beijing National Convention Center Phase II. In formal settings, whether a meeting record is usable depends not on device specifications but on consequences: Is the cost of misattributing one person significantly higher than in everyday situations? At the event, iFlytek Chairman Liu Qingfeng introduced the newly released SparkAudio 1.0, a speech foundation large model, followed by a demonstration video recorded with an iFlytek Office Notebook. This version will go live on October 24, with updates simultaneously pushed to the iFlytek Office Notebook X5 and Air3; users can experience it after upgrading.
Ultimately, the change this technology brings to the notebook amounts to one thing: it can tell who is speaking. SparkAudio 1.0 is trained on fully domestic computing power, and among its newly expanded capabilities is one called 'real-time speaker separation'—the machine must not only hear clearly what was said but also determine who said it.
Technology needs a carrier to take effect. The iFlytek Office Notebook X5 is that carrier: it goes into the conference room with people, sits on the table, and transcribes speech into text as it listens. This task itself is not difficult; the challenge lies in knowing who said each sentence after the words are captured.

In recent years, the common approach at AI hardware launch events has been to list specifications, compare benchmark scores, and announce accuracy rates. These metrics all measure 'hearing clearly'—whether the same audio can be transcribed word for word. But in a real meeting, the difficulty rarely lies in hearing clearly. The same sentence carries different weight when spoken by different people; within thosedozen-odd seconds of people talking over one another, the most critical disagreements are often hidden.
Meeting records have traditionally only done two things: preserve the audio and convert it into text. Both address preservation; the next step is attribution—the phrase 'I disagree with this approach' means something different when it comes from leadership versus from execution-level staff. Role-based identification addresses precisely this layer. The demonstration pushed this capability forward in three ways.
Names are attached this way. During recording transcription, the device first separates the voices, labeling them as 'Speaker 1,' 'Speaker 2,' and so on; you then tap one and write in the actual name. This action is equivalent to enrolling that person's voiceprint—the device records not just three characters but the acoustic features of that voice. The next time they speak, the record displays their name instead of a number.
This saves the preparation that used to be required before meetings. Previously, to make the device recognize participants, you had to gather everyone individually and have them speak into the machine for a segment—a meeting of over a dozen people could consume considerable time just on this step, and walk-ins or last-minute additions were often missed. Now the order is reversed: recognize while the meeting runs, and fill in names afterward. A name needs to be registered only once; the next time they walk in, the device still recognizes that same name.
The harder challenge is overlapping speech. Real meetings are not turn-based; two people speaking simultaneously or one cutting in before another finishes are the norm. When ordinary transcription encounters such segments, it merges the multiple audio streams into one continuous text, making it impossible to tell afterward which half came from whom; some systems simply discard the overlapping seconds entirely, leaving a gap in the record.
The notebook handles this by splitting overlapping speech by person. Within those few overlapping seconds, it first determines how many voices are present and who they belong to, then assigns each segment accordingly—so whoever started, whoever responded, and whoever interjects can all be correctly attributed. This isn't about accuracy per se; it's about whether a record can be used directly after the meeting ends. The most important sentence in a meeting often occurs during those few dozen seconds of overlapping speech.

There's also linguistic mixing—Chinese-English code-switching and jargon pronounced as-is are standard in professional meetings. If goodwill impairment (goodwill impairment) is transcribed as commercial capital reduction (commercial capital reduction) during a financial review, the sentence loses its original meaning; in an engineering review, missing a single letter in a model designation points to a completely different product. The wrong character won't trigger an alert—it will sit in the record until someone relies on it as evidence.
These terms can be prepared in advance or learned as they're used. Before the meeting, hand-write words that are commonly mispronounced; after lifting your pen, the current page's transcription results will be reprocessed. Terminology can be imported in batches rather than added one by one. More importantly, words that have been noted are retained: a corrected term from one session remains available as a candidate in the next. Terminology and speaker identity are both stored as 'remembered information' and invoked as context during recognition—the word must be identified correctly before the speaker can be.
Together, these three capabilities produce a record that is ready to use directly.
In a real meeting, all three things often happen simultaneously: people are talking over one another, someone is using an English acronym, and someone has gone by a nickname. The device must, within the same time span, separate the voices, match names to speakers, and identify terms correctly, ultimately producing a record that clearly states who said what and when.
After the meeting, there's no need to re-listen, correct errors sentence by sentence, or guess who 'Speaker 3' is; months later, when someone asks 'Who proposed that approach originally?', flipping to that section provides the answer. For users, the change is reflected directly in their actions: where it previously took half an hour to organize notes after a meeting, now a draft is ready upon opening; where manual verification of attributions was needed before archiving, the attribution is already present at the moment the record is generated.
When iFlytek launched its first-generation smart office notebook in 2019, it targeted professional meeting recording. According to Leto Technology, iFlytek has ranked first in market share for four consecutive years among office e-paper devices on major online channels, with cumulative sales exceeding 2 million units. Over seven years, what the market has come to accept is a standard: whether a device can enter formal settings is judged not by how much it records but by whether what it records can be claimed and verified.
A hallmark of mature technology is never how high the ceiling is but how stable the floor is: the ceiling determines how impressive a demonstration can be; the floor determines whether a task can be handed off to it. AI for office scenarios has, in recent years, mainly dealt with generation—errors can be corrected, changes leave no trace, and consequences fall on the user. Meeting records are not in that category; their output is a document that can be cited, questioned, and held accountable. The standard does not relax because a machine produces it—in fact, it becomes stricter.
Standards must ultimately land on a concrete test: when a dispute arises, can you flip to that exact sentence and identify the speaker? This is also why iFlytek's office notebook is repeatedly mentioned in such settings—it addresses the question of 'who was recorded,' not 'how much was recorded.'
More reliable in formal occasions — the reason lies not in recording more, but in recording what is usable.
Source: https://finance.sina.com.cn/stock/t/2026-10-23/doc-iniwctax5107563.shtml If you have any copyright concerns, please contact us at: 13951431913#139.com or submit a comment on the right side.

