DiarizationLM: Clarification on TPST usage for preserving ASR transcript while transferring semantic speaker labels

Hi team,

I’m working with the `diarizationlm.transcript_preserving_speaker_transfer` (TPST) function and need some guidance on the best way to preserve the original ASR transcript while transferring semantic speaker labels (e.g., Doctor/Patient) predicted by an LLM.

### What I’ve tried

**Direct TPST from LLM → ASR:**

```python
transferred_spk_label = diarizationlm.transcript_preserving_speaker_transfer(
words_llm_label, speakers_llm_label,
words_corrected, speakers_corrected
)

# Variable meaning:
# words_llm_label: LLM-generated transcript text
# speakers_llm_label: LLM-generated speaker labels (semantic, e.g., Doctor/Patient)
# words_corrected: ASR transcript (must remain unchanged)
# speakers_corrected: ASR speaker IDs (numeric)
```

**Issue:**
The output contains **ASR numeric speaker IDs**, not the semantic labels.
**Reason:**
TPST inherently outputs labels in the **target speaker space**, so semantic labels get replaced by ASR numeric IDs.

### My Use Case

* Preserve the **exact ASR transcript** (word-for-word).
* Transfer **semantic speaker labels** from the LLM output.
* Support cases where multiple ASR speaker IDs need to be mapped to only semantic roles (eg: Doctor/Patient).

### My Question

1. Is there a recommended method or best practice to use TPST in this scenario while ensuring the ASR transcript stays fully intact **and** semantic labels are preserved/propagated correctly?

Any guidance or suggestions would be greatly appreciated.

Thank you for this great algorithm and all the work behind it!