From aa52fd59d96521a577f055ea4e81a24be37701e5 Mon Sep 17 00:00:00 2001
From: Tech Carrement <tech@carrement.ai>
Date: Tue, 30 Jun 2026 17:01:24 +0200
Subject: [PATCH] fix(elevenlabs): support multichannel transcription in native
 /audio/transcriptions
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Two changes to litellm/llms/elevenlabs/audio_transcription/transformation.py so
ElevenLabs Scribe's multichannel speech-to-text works through the native
transcription path:

1. Request — serialize boolean form-data values as lowercase `true`/`false`
   instead of Python's `str(True)` -> `"True"`. ElevenLabs (and its SDK, via
   httpx) expects lowercase, so flags like `use_multi_channel` / `diarize` were
   being sent as `"True"` and silently ignored.

2. Response — when `use_multi_channel=true`, ElevenLabs returns a per-channel
   `transcripts` array (each entry has a `channel_index` and its own `words`)
   instead of a flat top-level `text`. Surface that `transcripts` array (and
   `audio_duration_secs`) on the TranscriptionResponse so per-channel/speaker
   data isn't dropped. These ride through to the client as extra fields because
   TranscriptionResponse allows extras; single-channel responses are unchanged.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---
 .../audio_transcription/transformation.py     | 33 +++++++++++++++++--
 1 file changed, 31 insertions(+), 2 deletions(-)

diff --git a/litellm/llms/elevenlabs/audio_transcription/transformation.py b/litellm/llms/elevenlabs/audio_transcription/transformation.py
index 68d1b5e16dd..c31e5df1b7f 100644
--- a/litellm/llms/elevenlabs/audio_transcription/transformation.py
+++ b/litellm/llms/elevenlabs/audio_transcription/transformation.py
@@ -23,6 +23,20 @@ from ...base_llm.audio_transcription.transformation import (
 from ..common_utils import ElevenLabsException
 
 
+def _to_form_value(value: object) -> str:
+    """Serialize a multipart form-field value the way ElevenLabs expects.
+
+    httpx (which the ElevenLabs SDK uses) encodes booleans as lowercase
+    ``true``/``false``. Python's ``str(True)`` yields ``"True"``, which the
+    ElevenLabs API does not recognize — so a boolean flag such as
+    ``use_multi_channel`` or ``diarize`` would be silently ignored. Normalize
+    bools to lowercase; everything else is stringified as before.
+    """
+    if isinstance(value, bool):
+        return "true" if value else "false"
+    return str(value)
+
+
 class ElevenLabsAudioTranscriptionConfig(BaseAudioTranscriptionConfig):
     @property
     def custom_llm_provider(self) -> str:
@@ -79,7 +93,7 @@ class ElevenLabsAudioTranscriptionConfig(BaseAudioTranscriptionConfig):
         for key, value in optional_params.items():
             if key in self.get_supported_openai_params(model) and value is not None:
                 # Convert values to strings for form data, but skip None values
-                form_data[key] = str(value)
+                form_data[key] = _to_form_value(value)
 
         #########################################################
         # Add Provider Specific Parameters
@@ -91,7 +105,7 @@ class ElevenLabsAudioTranscriptionConfig(BaseAudioTranscriptionConfig):
         )
 
         for key, value in provider_specific_params.items():
-            form_data[key] = str(value)
+            form_data[key] = _to_form_value(value)
         #########################################################
         #########################################################
 
@@ -140,6 +154,21 @@ class ElevenLabsAudioTranscriptionConfig(BaseAudioTranscriptionConfig):
                             }
                         )
 
+            # Surface ElevenLabs multichannel output. With `use_multi_channel=true`
+            # the response carries a per-channel `transcripts` array (each entry has a
+            # `channel_index` plus its own `words`) instead of a flat top-level `text`.
+            # Pass it through verbatim so callers can attribute words to a speaker by
+            # channel; without this the per-channel data is dropped.
+            transcripts = response_json.get("transcripts")
+            if isinstance(transcripts, list):
+                response["transcripts"] = transcripts
+
+            # Carry the billed audio duration (present on both single- and
+            # multi-channel responses) so callers can attribute cost.
+            duration = response_json.get("audio_duration_secs")
+            if duration is not None:
+                response["audio_duration_secs"] = duration
+
             # Store full response in hidden params
             response._hidden_params = response_json