On October 24, 2026, iFlytek's SparkAudio 1.0, the speech foundation model in its Spark line, went live on the schedule announced at the 1024 Developers Festival on October 22, with updates rolling out to the iFlytek AINOTE X5 and Air3 and the model trained on fully domestic computing power. The most important change in this release is not hearing more clearly, but telling who is speaking.
The change is about attribution, not audibility
Meeting transcription used to solve two problems: keeping the audio, and turning it into text. SparkAudio 1.0's new real-time speaker separation goes one step further: during transcription it first separates voices by person, labels them Speaker 1 and Speaker 2, and once you tap in a name, that person's voiceprint is remembered, so the next time they speak in a meeting, the record shows their name instead of a number. The order is also reversed: there is no need to call every attendee in before the meeting to register their voice. The system recognizes people as the meeting goes on and names can be added afterwards, so last-minute seats and late arrivals are far less likely to be missed.
Two hard problems: overlapping speech and terminology
Real meetings rarely take turns. Two people talking at once, or one person jumping in before another finishes, is normal. Ordinary transcription tends to smear such passages into a single stream, or simply drop the overlapping seconds. Separating by speaker works the other way: first work out how many voices there are and who they belong to, then assign every sentence to the right person, so it is clear who started, who followed, and which half-sentence was an interruption. The other hard problem is mixed language: Chinese and English interleaved, industry terms spoken aloud, where a single wrong character can change the meaning. iFlytek's approach lets terms be corrected by handwriting before the meeting and imported in batches; corrected terms are kept and reused as context next time, because getting the words right is what makes it possible to get the people right.
Who benefits first
The first beneficiaries are meetings whose records need to be usable the moment the meeting ends: formal meetings, review sessions, and discussions where someone later needs to trace who proposed a particular plan. Because attribution is attached at the moment the record is created, the post-meeting line-by-line check is no longer needed. The boundary should be stated plainly, though: the public information so far comes mainly from the festival presentation and demo, and real meeting rooms differ in noise, accents, and how fiercely people talk over each other, so real-world results will have to be judged after upgrading. For users, the test is simple: after the next meeting, open the record and check first whether the overlapping seconds were attributed to the right people.
