Hey HN, I built Ellis, an AI notetaker for individuals, not orgs. And since my personal schedule includes meetings outside of Zoom, it was important that it works well for in-person meetings.
In fact to me, in-person conversations have always been a greater source of learning and creative ideas than virtual. On Zoom, the fixed screen and talking points tend to bring out a more transactional nature in me.
Yet building a notetaker for in-person isn't trivial. Online, each speaker has their own audio stream. In a room (or during a walk outside) everyone shares the same mic.
Models like AssemblyAI give you transcriptions with Speaker A and Speaker B, but you still have to mark yourself as the one who said X and not Y.
This is especially true if I want to get coached from my conversations. A useful use case for me as an introvert.
Here's how Ellis does this:
1) You create a voice profile of yourself (Deepgram diarization + pyannote's wespeaker model)
2) It splits the recording into speaker segments using AssemblyAI
3) It surfaces the most distinctive line from each unnamed speaker — "Moana is the best movie ever!"
4) You tap to confirm.
Remember to ask for consent! Ellis prompts you with ways to ask at the start of a recording.
Still not solved: two near-identical voices in the same room (e.g. me and my dad). Curious if anyone found a diarization approach that separates near-identical voices without per-speaker enrollment?
If you have an opinion about the tool, I'm all ears! You can try it for free.
Hey HN, I built Ellis, an AI notetaker for individuals, not orgs. And since my personal schedule includes meetings outside of Zoom, it was important that it works well for in-person meetings.
In fact to me, in-person conversations have always been a greater source of learning and creative ideas than virtual. On Zoom, the fixed screen and talking points tend to bring out a more transactional nature in me.
Yet building a notetaker for in-person isn't trivial. Online, each speaker has their own audio stream. In a room (or during a walk outside) everyone shares the same mic.
Models like AssemblyAI give you transcriptions with Speaker A and Speaker B, but you still have to mark yourself as the one who said X and not Y.
This is especially true if I want to get coached from my conversations. A useful use case for me as an introvert.
Here's how Ellis does this: 1) You create a voice profile of yourself (Deepgram diarization + pyannote's wespeaker model) 2) It splits the recording into speaker segments using AssemblyAI 3) It surfaces the most distinctive line from each unnamed speaker — "Moana is the best movie ever!" 4) You tap to confirm.
Remember to ask for consent! Ellis prompts you with ways to ask at the start of a recording.
Still not solved: two near-identical voices in the same room (e.g. me and my dad). Curious if anyone found a diarization approach that separates near-identical voices without per-speaker enrollment?
If you have an opinion about the tool, I'm all ears! You can try it for free.