From 7dc382ecbbe3c829f27db302ef6c7febb3ed5e62 Mon Sep 17 00:00:00 2001
From: Maximilian Kohl <maximilian.kohl@amicon-partners.de>
Date: Wed, 5 Aug 2026 23:11:22 +0100
Subject: [PATCH] fix: preserve `logprobs` on audio transcription responses
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

`convert_to_model_response_object` copied only
`language/task/duration/words/segments` from a transcription response, so
`logprobs` was silently dropped — even when the caller explicitly asked for
it via `include[]=logprobs`.

Effect: for `gpt-4o-transcribe` there was no way to read per-token
confidence through the proxy. The only visible workaround was `stream=True`,
which returns the raw SSE body as `text` with `usage: null` — and therefore
silently disables cost tracking and budget enforcement for every
transcription request.

Adding `logprobs` to the copied keys makes the non-streaming path complete,
so callers can have both confidence scores and correct spend accounting.

Tests: two regression tests (with and without `logprobs`) in
`tests/test_litellm/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py`.
---
 .../convert_dict_to_response.py               |  2 +-
 .../test_convert_dict_to_response.py          | 47 +++++++++++++++++++
 2 files changed, 48 insertions(+), 1 deletion(-)

diff --git a/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py b/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py
index a6f10e1ede3..a6819d187e8 100644
--- a/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py
+++ b/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py
@@ -809,7 +809,7 @@ def convert_to_model_response_object(
             if "text" in response_object:
                 model_response_object.text = response_object["text"]
 
-            optional_keys: Final = ["language", "task", "duration", "words", "segments"]
+            optional_keys: Final = ["language", "task", "duration", "words", "segments", "logprobs"]
             for key in optional_keys:  # not guaranteed to be in response
                 if key in response_object:
                     setattr(model_response_object, key, response_object[key])
diff --git a/tests/test_litellm/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py b/tests/test_litellm/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py
index 293e5de304f..6c10b6af444 100644
--- a/tests/test_litellm/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py
+++ b/tests/test_litellm/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py
@@ -14,6 +14,7 @@ from litellm.types.utils import (
     ChatCompletionMessageToolCall,
     Function,
     ModelResponse,
+    TranscriptionResponse,
 )
 
 OPENAI_CUSTOM_TOOL_CALL_RESPONSE = {
@@ -102,3 +103,49 @@ def test_handle_invalid_parallel_tool_calls_skips_custom_tool_calls():
     )
     result = _handle_invalid_parallel_tool_calls([custom_tool_call, function_tool_call])
     assert result == [custom_tool_call, function_tool_call]
+
+
+def test_audio_transcription_preserves_logprobs():
+    """`logprobs` must survive conversion (requested via `include[]=logprobs`).
+
+    Regression: only language/task/duration/words/segments were copied, so
+    `logprobs` was silently dropped for `gpt-4o-transcribe` and callers had no
+    way to read the per-token confidence without falling back to `stream=True`.
+    """
+    response_object = {
+        "text": "Guten Tag.",
+        "logprobs": [
+            {"token": "Guten", "logprob": -0.01, "bytes": [71]},
+            {"token": " Tag", "logprob": -0.02, "bytes": [32]},
+        ],
+        "usage": {
+            "type": "tokens",
+            "input_tokens": 30,
+            "output_tokens": 4,
+            "total_tokens": 34,
+            "input_token_details": {"text_tokens": 0, "audio_tokens": 30},
+        },
+    }
+
+    result = convert_to_model_response_object(
+        response_object=response_object,
+        model_response_object=TranscriptionResponse(),
+        response_type="audio_transcription",
+    )
+
+    assert result.text == "Guten Tag."
+    assert result.logprobs == response_object["logprobs"]
+    # usage must keep working alongside it (cost tracking depends on it)
+    assert result.usage.input_tokens == 30
+
+
+def test_audio_transcription_without_logprobs_is_unchanged():
+    """No `logprobs` in the response → attribute simply absent, no crash."""
+    result = convert_to_model_response_object(
+        response_object={"text": "Guten Tag."},
+        model_response_object=TranscriptionResponse(),
+        response_type="audio_transcription",
+    )
+
+    assert result.text == "Guten Tag."
+    assert getattr(result, "logprobs", None) is None