A customer and a colleague are discussing a budget. Your colleague says it is approved. The customer says it still needs a decision on Friday.
Put both sentences under the customer's name and the record becomes misleading, even if every word is correct. Later, an assistant might retrieve the supposed approval without showing you the disagreement.
This is why speaker recognition matters in conversation memory. You need to recover what a particular person said, not just the words that were spoken in the room.
Speaker 1 is not a contact
Three separate jobs can be hidden behind a transcript:
| Job | What it gives you |
|---|---|
| Transcription | The spoken words as text |
| Speaker diarization | Sections labeled by different speakers within a recording |
| Recurring-speaker recognition | A link between speech and an existing speaker profile across conversations |
A transcript labeled Speaker 1 and Speaker 2 has separated the voices. It has not necessarily identified either person. Speaker 1 on Monday can be someone entirely different from Speaker 1 on Thursday.
The current public Draki app provides OpenAI cloud transcription after your permission. It does not automatically separate speakers, enroll voices or recognize recurring people. You can associate saved people with a conversation manually. The voice-ID discussion here explains development goals and evaluation concepts, not a feature shipped in the public app.
The difference becomes useful before a repeat meeting. Instead of trying to remember which recording contains the customer's earlier concern, you can look through the conversations associated with that person. Dates and source passages still matter, especially if their position has changed. We explore that use in remembering people and conversations.
Check a disagreement, not just a monologue
A quiet introduction is an easy place to start evaluating speaker recognition. It tells you less about a meeting where people interrupt, correct each other or speak from opposite ends of a table.
Try a short conversation whose participants you know. Have one person propose Tuesday for a deadline and another explain that Thursday is the earliest possible date. Then agree who will do the work. Check that the saved record preserves both the final date and the person who accepted the task.
This is a test you can run, not a Draki accuracy benchmark. Background noise, overlapping speech and microphone placement can affect attribution. Before relying on an important amount or commitment, open the passage behind the answer and check the speaker. If the association looks wrong, resolve it before treating the statement as that person's history.
A pause controls what happens next
You can use Draki's app without a bracelet. The optional bracelet makes it convenient to keep capture available as you go about your day, and you can turn recording on or off whenever you choose. If a work discussion turns personal, or someone asks to speak off the record, stop capture at that point.
The earlier part of the conversation remains a separate question. Pausing does not erase a transcript already saved, remove a voice profile or withdraw text already sent to another assistant. To remove existing information, use the data-control and deletion instructions.
Explain recording before it starts. Recurring voice recognition is a separate use to discuss with the people involved: being comfortable with one recorded meeting does not automatically settle how their voice will be used to recognize them later.
What the current app does
New public recordings go through Draki to OpenAI over Wi-Fi after your permission. Text is stored for cloud summaries and search. Public Draki does not create voice profiles. The privacy policy describes these data flows and the optional MCP connections for sharing text and context.
Earlier private test versions may have created voice profiles or reference samples. Those are separate from a transcript, and switching to the public app does not erase them. Ask support to include them in a deletion request if applicable. The storage guide explains current storage.
Draki offers a custom MCP connection for AI assistants that support remote MCP, subject to their account requirements and settings. The connection provides optional, read-only access to text and context you authorize for sharing. It does not edit or delete your records or export audio, voice samples or voice profiles. The external assistant processes shared text in its own service. Disconnecting stops future access; it does not remove text the provider has already received.
The purpose of a voice ID is straightforward: when you return to a conversation, the words should still belong to the person who said them. That is the foundation for asking a more useful question next time you meet.
FAQ
- Can I pause or stop recording in Draki?
- Yes. You can turn recording on or off whenever you want, using the mobile app on its own or with the optional Draki bracelet. Stopping controls further capture; it does not erase a conversation already saved.
- Does a voice ID tell Draki a stranger's real name?
- No. A voice ID helps associate recurring speech with a speaker profile. A reliable name or contact association is a separate part of identifying the person.
- Is speaker diarization the same as recognizing someone next week?
- No. Diarization separates different speakers within a recording. Recognizing a recurring person links speech to an existing speaker profile across conversations.
- Where does Draki process recordings and voice information?
- The current public app uses OpenAI cloud transcription after your permission. It does not automatically separate or recognize speakers. Audio is sent through Draki to OpenAI over Wi-Fi after permission; text is stored for cloud summaries and search. With your authorization, the optional MCP connection can share text and context with a compatible assistant.
Written by Darijan Ducic
Draki for iPhone
