> ## Documentation Index
> Fetch the complete documentation index at: https://docs.corti.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Diarized Transcript Handling

> Best practices for ordering and rendering speaker-separated transcripts from /streams

When **diarization** is enabled (`diarize: true`) on [`/streams`](/stt/streams), the server attributes speech to distinct speakers. Because segments for different speakers are finalized independently, `type: "transcript"` messages are **not guaranteed to arrive in chronological order** over the WebSocket.

<Note>
  This applies to multi-speaker `/streams` sessions only when [diarization](/stt/diarization) is enabled. The `/transcribe` endpoint is single-speaker only.
</Note>

***

## Diarized Transcript Segment Ordering

Each `transcript` message carries a `data` **array**, and multiple segments for different speakers can arrive in the same message. Ordering of transcript text should be mindful of both speaker and speech time: A later-spoken segment for one speaker can also arrive before an earlier-spoken segment from another.

<Check>
  Only **final** segments are returned on `/streams` (`final: true`); there are no interim results to reconcile. The handling concern is purely **ordering**, not interim-vs-final dedup as is required with [dictation transcript handling](/stt/best-practices-transcribe)).
</Check>

Each segment in `data` includes a `time` object you can use to recover the true order:

| Field                 | Use                                                                                              |
| :-------------------- | :----------------------------------------------------------------------------------------------- |
| `time.start`          | **Primary ordering key** — when the speech began (seconds)                                       |
| `time.end`            | Secondary key for tie-breaks and overlap checks (seconds)                                        |
| `speakerId`           | A distinct integer per detected speaker (up to four). Returns `-1` when diarization is disabled. |
| `participant.channel` | Audio channel the segment was attributed to                                                      |

<Note>
  `speakerId` and `participant.channel` are independent concepts. Diarization separates speakers *within* a transcribed stream; the channel reflects audio routing. Do not assume a fixed mapping between the two.
</Note>

<Warning>
  Rendering segments in socket-arrival order may interleave one speaker's sentence into the middle of another's. In such a case, no words are lost, but the sequence may be unreliable. Always order segments by `time.start`.
</Warning>

***

## Recommendation

* **Position each segment by `time.start`**, not by arrival order. Insert into the running transcript at the position determined by `time.start` rather than appending to the end.
* Use `time.end` as a secondary key for tie-breaks.
* Iterate the full `data` array on every message and order before rendering — do not assume a single segment per message.

```js theme={null}
case "transcript":
  // message.data is an array — order by start time before committing to the UI
  const ordered = [...message.data].sort(
    (a, b) => a.time.start - b.time.start || a.time.end - b.time.end
  );
  ordered.forEach((seg) => {
    insertByStartTime(seg); // position by seg.time.start, do not blindly append
  });
  break;
```

<Note>
  Please [contact us](https://help.corti.app) if you need further assistance working with diarized transcripts.
</Note>