Poor and inconsistent recognition of speakers within a conversation. Incorrect identification of my voice. This breaks any use cases: analyze dialogs, calls, insight of the day etc

The current voice-identification feature isn’t viable in real-world use because it only works when users manually label speakers each time, which creates friction and undermines trust. We should prioritize rock-solid detection of the owner’s voice first: if the system can’t reliably identify the owner, there’s little value in interpreting other voices—especially for Daily Insight—because all downstream analytics hinge on a stable, correctly tagged anchor. By making owner-voice recognition automatic, accurate, and persistent across sessions, we reduce labeling burden, prevent error propagation, and immediately improve insight quality and adoption. Once the owner’s voice is consistently recognized, we can incrementally expand to other speakers with far better precision and a far better user experience.

Voice id has been pretty useless for me. It still doesn’t identify people I have labeled multiple times. Only does it for the day I update them. And it is struggling to differentiate me from everyone else so my “suggestions for improvement” 50% of the time are commentary about a conversation that wasn’t me at all.

I disagree with Daniel, the UI/UX is really bad.

Its working WAAAAY better than before and gets things right most of the time (sometimes).

But training the voices, knowing how many seconds we have record, choosing which sample of voice they have been trained etc... Is a must have since EVERYTHING else depends on this feature.

The tech my be working but without a good UX for identifying voice its a pain to use.

It’s an amazing UI over time I hope the recognition gets better. And add back auto speaker creation, but then you’d need the ability to merge saved voices with the auto tagged ones etc etc and anyways I understand why we don’t do that yet

When will this be available? It says "shipped," but I don't have it. It is a critical feature to make this app usable.

Would love to see this feature improved. Two biggest issues currently:

1. Poor and inconsistent recognition of speakers within a conversation. For example, I can have a meeting with 2 people, and it will identify 5 different speakers, and fail to even identify those "5" consistently.
2. Incorrect identification of my voice. It will incorrectly identify other speakers as me, which corrupts any conclusions or insights I may receive from Ask AI feature.

I would love to be able to train it on more than one voice. i.e., it constantly thinks my wife is my daughter. If I could get them both to read a phrase like I did when training my app, then it could keep them separated more easily.

It forgets people's voice, also mine at the next conversation

This needs to be improved :) but its a great "start" seems like it forgets who people are the next day or even next section.

Can we store mappings of "name -> speaker id" automatically?

At a minimum, within the same recording session. Ideally across any recording.

There are a few people that I'd be interested in automatically tagging, such as myself, family members, or important coworkers. Those names help improve searching and readability, along with confirming "wait, what was that action item, and who needed it?"

Right now, I need to manually tag each speaker per-conversation, which may be inadvertently split into multiple despite being part of the same recording session. As I can't listen to the recording, I also can't just go "yeah, that's Kalani" and instead need to infer based on other context clues said in the transcript.

Using other tools, you can do this via embedding and clustering via pyannote.audio, SpeechBrain, and Resemblyzer to create embeddings profiles per-speaker. Then, leave it up to the user to confirm "yes, this is Kalani", "no, this is not Kalani". This approach allows you to expand and improve the embeddings and clustering over time as you hear more data, which also goes well for people who will speak differently in different situations.

However, you could also do an enrollment approach, whereby common speakers will give you a known voice sample, similar to how Google Assistant or Apple Siri handle initial setup.

Introduction of visual display picture identification is difficult as they are all the same colour.

Would be nice to set colours per contact and maybe add additional tags or add them to groups.

Also editing a contact would be greatly appreciated.

I am running Version 1.897.1 of the Mac app, and I don't have voice identification. Why?

Will there be any capability in here to name the speakers after and then have them tagged the same way moving forward? If not should I set that as another Feature Idea if I wish to see it?

Great news, and essential feature!

Huzzah! Thank you!

Great feature, but it's not working so well when it miscategorizes. In a meeting I'm in at this moment, it thinks my boss's voice is mine. When I try to change the speaker for a given paragraph, it changes all of them. Is there a way to reassign specific spoken sections to a different speaker without redoing all of them? And maybe it could learn from its mistakes going forward.

This is an improvement, but please let me use freeform naming of people that aren't in my contacts, e.g. a client meeting or interview.

Noticed that this has gone live - looks like it's working really well so far! Don't see it integrated into any AI features yet but it looks great in the transcript.