From f22b17a925a2ecc6141e03108a185c8cbc0b7c27 Mon Sep 17 00:00:00 2001
From: Ben Langfeld <blangfeld@powerhrg.com>
Date: Mon, 26 Oct 2026 16:01:43 -0300
Subject: [PATCH] fix(elevenlabs): keep the flat word list on diarized
 responses

---
 .../audio_transcription/transformation.py     | 29 ++++++++++++-------
 .../test_transformation.py                    |  8 +++--
 2 files changed, 24 insertions(+), 13 deletions(-)

diff --git a/litellm/llms/elevenlabs/audio_transcription/transformation.py b/litellm/llms/elevenlabs/audio_transcription/transformation.py
index c5cb3a91ac7..6de3f07c39d 100644
--- a/litellm/llms/elevenlabs/audio_transcription/transformation.py
+++ b/litellm/llms/elevenlabs/audio_transcription/transformation.py
@@ -3,7 +3,7 @@ Translates from OpenAI's `/v1/audio/transcriptions` to ElevenLabs's `/v1/speech-
 """
 
 import io
-from collections.abc import Iterable, Iterator, Mapping
+from collections.abc import Iterable, Iterator, Mapping, Sequence
 from itertools import chain, groupby
 from types import MappingProxyType
 from typing import Final, Protocol
@@ -245,7 +245,10 @@ def _transcription_response(response_object: Mapping[str, object]) -> Transcript
 
     words: Final = tuple(_JSON_OBJECTS.validate_python(response_object.get("words", ())))
     if any(word.get("speaker_id") is not None for word in words):
-        return _diarized_response(ElevenLabsSTTChunk.model_validate(response_object))
+        response: Final = _diarized_response(ElevenLabsSTTChunk.model_validate(response_object))
+        # Callers already passing diarize=True read the flat word list, so it stays alongside the segments
+        response["words"] = _openai_words(words)
+        return response
     return _flat_response(response_object, words)
 
 
@@ -260,19 +263,23 @@ def _flat_response(
 
     # Map ElevenLabs words to OpenAI format
     if "words" in response_object:
-        response["words"] = [
-            {
-                "word": word_data.get("text", ""),
-                "start": word_data.get("start", 0),
-                "end": word_data.get("end", 0),
-            }
-            for word_data in words
-            if word_data.get("type") == "word"
-        ]
+        response["words"] = _openai_words(words)
 
     return response
 
 
+def _openai_words(words: tuple[Mapping[str, object], ...]) -> Sequence[Mapping[str, object]]:
+    return [
+        {
+            "word": word_data.get("text", ""),
+            "start": word_data.get("start", 0),
+            "end": word_data.get("end", 0),
+        }
+        for word_data in words
+        if word_data.get("type") == "word"
+    ]
+
+
 def _diarized_response(transcript: ElevenLabsSTTChunk) -> TranscriptionResponse:
     response: Final = TranscriptionResponse(text=transcript.text)
     response["task"] = "transcribe"
diff --git a/tests/unit/llms/elevenlabs/audio_transcription/test_transformation.py b/tests/unit/llms/elevenlabs/audio_transcription/test_transformation.py
index b90da5683af..feffbaf4556 100644
--- a/tests/unit/llms/elevenlabs/audio_transcription/test_transformation.py
+++ b/tests/unit/llms/elevenlabs/audio_transcription/test_transformation.py
@@ -48,7 +48,7 @@ def test_transform_audio_transcription_response_keeps_only_spoken_words():
         "language_code": "en",
         "text": "Hello world",
         "words": [
-            {"type": "word", "text": "Hello", "start": 0.0, "end": 0.4},
+            {"type": "word", "text": "Hello", "start": 0.0, "end": 0.4, "speaker_id": "speaker_0"},
             {"type": "spacing", "text": " ", "start": 0.4, "end": 0.5},
             {"type": "audio_event", "text": "(laughter)", "start": 0.5, "end": 0.9},
             {"type": "word", "text": "world", "start": 0.9, "end": 1.3},
@@ -247,7 +247,11 @@ def test_diarized_response_builds_speaker_segments():
             "type": "transcript.text.segment",
         },
     )
-    assert "words" not in response
+    assert response["words"] == [
+        {"word": "Bom", "start": 3.0, "end": 3.4},
+        {"word": "dia", "start": 3.4, "end": 3.8},
+        {"word": "Oi", "start": 5.0, "end": 5.3},
+    ]
 
 
 def test_diarized_response_reports_fractional_duration_as_typed_usage():