feat(openai-realtime): allow pinning input transcription language (#557)

The OpenAI realtime factory already reads `language` from the realtime
config, but builds `InputAudioTranscription()` without it, so the language
is silently dropped and the model auto-detects on every utterance.

On 8kHz telephony audio this misfires badly: in our production tests a
Brazilian Portuguese speaker was transcribed as English, French and Chinese
within a single call, which then corrupted downstream extraction.

Every other STT branch in `create_stt_service` already honours
`language` (Deepgram, Google, Cartesia, Dograh, Sarvam), and
`GoogleRealtimeLLMConfiguration` already exposes a `language` field for
Gemini Live. This brings the OpenAI realtime provider in line with both.

- expose `language` on `OpenAIRealtimeLLMConfiguration` (optional,
  defaults to None -> unchanged auto-detect behaviour)
- pass it through to `InputAudioTranscription`, which already accepts it

Verified against pipecat: the session now carries
`{"transcription": {"model": "gpt-realtime-whisper", "language": "pt"}}`.

Co-authored-by: Liberty Card <tecnologialibertycard@gmail.com>

api/services/configuration/registry.py

@@ -584,6 +584,20 @@ class SarvamLLMConfiguration(BaseLLMConfiguration):
 
 
 OPENAI_REALTIME_MODELS = ["gpt-realtime-2"]
+# ISO 639-1 codes accepted by the Realtime API's input_audio_transcription.
+# Not exhaustive — the field allows custom input.
+OPENAI_REALTIME_LANGUAGES = [
+    "en",
+    "es",
+    "pt",
+    "fr",
+    "de",
+    "it",
+    "hi",
+    "ja",
+    "ko",
+    "zh",
+]
 OPENAI_REALTIME_VOICES = [
     "alloy",
     "ash",
@@ -618,6 +632,17 @@ class OpenAIRealtimeLLMConfiguration(BaseLLMConfiguration):
             "allow_custom_input": True,
         },
     )
+    language: str | None = Field(
+        default=None,
+        description=(
+            "ISO 639-1 language code for input audio transcription (e.g. 'pt', 'es'). "
+            "Improves transcription accuracy and latency. Leave unset to auto-detect."
+        ),
+        json_schema_extra={
+            "examples": OPENAI_REALTIME_LANGUAGES,
+            "allow_custom_input": True,
+        },
+    )
 
 
 GROK_REALTIME_MODELS = ["grok-voice-think-fast-1.0"]