fix(gemini): expose realtime transcription capabilities

This commit is contained in:
Marcin Juraszek 2026-09-23 00:19:23 +02:00
parent 0848080432
commit 4f55bee6ab
7 changed files with 49 additions and 8 deletions

View file

@ -0,0 +1,8 @@
from litellm.types.llms.openai import OpenAIAudioTranscriptionOptionalParams
from .transformation import GeminiAudioTranscriptionConfig
class GeminiRealtimeAudioTranscriptionConfig(GeminiAudioTranscriptionConfig):
def get_supported_openai_params(self, model: str) -> list[OpenAIAudioTranscriptionOptionalParams]:
return ["language", "keywords"]

View file

@ -62738,6 +62738,7 @@
"supports_function_calling": false
},
"gemini/gemini-3.5-transcribe-live": {
"audio_transcription_config": "gemini_realtime_transcribe",
"input_cost_per_audio_token": 3.5e-06,
"input_cost_per_token": 3.5e-06,
"litellm_provider": "gemini",

View file

@ -8951,6 +8951,15 @@ class ProviderConfigManager:
)
return VertexAIAudioTranscriptionConfig()
elif (
litellm.LlmProviders.GEMINI == provider
and model_cost_entry.get("audio_transcription_config") == "gemini_realtime_transcribe"
):
from litellm.llms.gemini.audio_transcription.realtime_transformation import (
GeminiRealtimeAudioTranscriptionConfig,
)
return GeminiRealtimeAudioTranscriptionConfig()
elif (
litellm.LlmProviders.GEMINI == provider
and model_cost_entry.get("audio_transcription_config") == "gemini_transcribe"

View file

@ -62738,6 +62738,7 @@
"supports_function_calling": false
},
"gemini/gemini-3.5-transcribe-live": {
"audio_transcription_config": "gemini_realtime_transcribe",
"input_cost_per_audio_token": 3.5e-06,
"input_cost_per_token": 3.5e-06,
"litellm_provider": "gemini",

View file

@ -2148,18 +2148,28 @@ def test_gemini_model_info_reports_transcription_params(monkeypatch):
assert "keywords" in model_info["supported_openai_params"]
def test_gemini_live_transcription_dispatch_has_no_batch_params():
def test_gemini_live_transcription_dispatch_returns_realtime_params(monkeypatch):
from litellm.litellm_core_utils.get_supported_openai_params import (
get_supported_openai_params,
)
monkeypatch.setitem(
litellm.model_cost,
"gemini/gemini-3.5-transcribe-live",
{
"audio_transcription_config": "gemini_realtime_transcribe",
"mode": "audio_transcription",
"supported_endpoints": ["/v1/realtime"],
},
)
params = get_supported_openai_params(
model="gemini-3.5-transcribe-live",
custom_llm_provider="gemini",
request_type="transcription",
)
assert params is None
assert params == ["language", "keywords"]
@pytest.mark.parametrize(

View file

@ -27,14 +27,24 @@ def test_gemini_transcription_dispatch_returns_batch_params(monkeypatch):
assert "timestamp_granularities" in params
def test_gemini_live_transcription_dispatch_does_not_use_batch_params():
def test_gemini_live_transcription_dispatch_returns_realtime_params(monkeypatch):
monkeypatch.setitem(
litellm.model_cost,
"gemini/gemini-3.5-transcribe-live",
{
"audio_transcription_config": "gemini_realtime_transcribe",
"mode": "audio_transcription",
"supported_endpoints": ["/v1/realtime"],
},
)
params = get_supported_openai_params(
model="gemini-3.5-transcribe-live",
custom_llm_provider="gemini",
request_type="transcription",
)
assert params is None
assert params == ["language", "keywords"]
def test_base_model_label_does_not_strip_bedrock_tools():

View file

@ -5,9 +5,10 @@ import httpx
import pytest
import litellm
from litellm.llms.gemini.audio_transcription.transformation import (
GeminiAudioTranscriptionConfig,
from litellm.llms.gemini.audio_transcription.realtime_transformation import (
GeminiRealtimeAudioTranscriptionConfig,
)
from litellm.llms.gemini.audio_transcription.transformation import GeminiAudioTranscriptionConfig
from litellm.llms.gemini.common_utils import GeminiError
from litellm.types.utils import LlmProviders
from litellm.utils import ProviderConfigManager, get_optional_params_transcription
@ -72,11 +73,12 @@ def test_provider_config_manager_returns_gemini_config():
assert isinstance(provider_config, GeminiAudioTranscriptionConfig)
def test_provider_config_manager_does_not_use_batch_config_for_gemini_live():
def test_provider_config_manager_uses_realtime_config_for_gemini_live():
provider_config = ProviderConfigManager.get_provider_audio_transcription_config(
model="gemini-3.5-transcribe-live", provider=LlmProviders.GEMINI
)
assert provider_config is None
assert isinstance(provider_config, GeminiRealtimeAudioTranscriptionConfig)
assert provider_config.get_supported_openai_params("gemini-3.5-transcribe-live") == ["language", "keywords"]
class TestSupportedParams: