SpeechConfig.language_code does not constrain input audio transcription language in Live API sessions

[amitanshupanigrahi2704]

Labels: bug, live-api

Description

When using the ADK's run_live() with native audio models (Gemini 2.0 Flash Live), setting language_code="en-US" in SpeechConfig does not constrain the input audio transcription language. The transcribed user input still intermittently appears in other languages (e.g., Hindi, Spanish, Japanese) even though the speaker is speaking English.

Configuration

run_config = RunConfig(
    streaming_mode=StreamingMode.BIDI,
    response_modalities=["AUDIO"],
    speech_config=types.SpeechConfig(
        language_code="en-US",
        voice_config=types.VoiceConfig(
            prebuilt_voice_config=types.PrebuiltVoiceConfig(voice_name="Kore")
        ),
    ),
    input_audio_transcription=types.AudioTranscriptionConfig(),
    output_audio_transcription=types.AudioTranscriptionConfig(),
)

Observed behaviour

• Output speech is correctly in English.
• Input transcription (input_audio_transcription events) frequently returns non-English text even though the user is speaking English exclusively. For example, an English sentence might be transcribed in Devanagari script or mixed with non-Latin characters.

Expected behaviour

Setting language_code="en-US" in SpeechConfig should constrain both the output speech language and the input speech-to-text transcription language. At minimum, the input transcription should respect the configured language and not produce text in unrelated languages.

Root cause analysis

AudioTranscriptionConfig is currently an empty class (pass) with no language_code or similar field:

class AudioTranscriptionConfig(_common.BaseModel):
  """The audio transcription configuration in Setup."""
  pass

There appears to be no mechanism to pass the desired transcription language to the underlying STT model for input audio. The SpeechConfig.language_code seems to only affect TTS output, not STT input.

Suggested fix

Either:

1. Propagate SpeechConfig.language_code to the input transcription pipeline so it constrains the STT language, or
2. Add a language_code field to AudioTranscriptionConfig so developers can explicitly set the transcription language:

   input_audio_transcription=types.AudioTranscriptionConfig(language_code="en-US")

Environment

• ADK version: google-adk 1.x (latest)
• google-genai SDK version: latest
• Model: gemini-2.5-flash-native-audio-preview-12-2025
• Streaming mode: BIDI (WebSocket)

---

[surajksharma07]

@amitanshupanigrahi2704 SpeechConfig.language_code only controls TTS output synthesis — the input STT pipeline is separate and doesn't pick it up.

Could you try setting language_codes=["en-US"] directly on AudioTranscriptionConfig instead:

input_audio_transcription=types.AudioTranscriptionConfig(language_codes=["en-US"])

Note it takes a list rather than a single string.
Give that a try and see if it pins the transcription to English on your end.