gemini-2.5-flash-native-audio Live API: Transcription Fails with "Invalid Argument" (1007) or No Messages in Node.js SDK (@google/genai@^1.34.0`) #1212

Product/Service: Vertex AI Gemini Live API

SDK: Node.js (@google/genai)

SDK Version: ^1.34.0 (as reported by package.json)

Model: gemini-2.5-flash-native-audio-preview-09-2025

Operating System: Linux

Description:

I am attempting to establish a real-time speech-to-speech session with transcription using the
gemini-2.5-flash-native-audio-preview-09-2025 model via the @google/genai Node.js SDK. The primary issue is the
inability to receive either inputAudioTranscription or outputAudioTranscription messages, or to even stabilize a
connection when configuring the API as per documentation.

There appear to be two related problems:

1. Connection Failure with Modality.TEXT in responseModalities:
   When Modality.TEXT is included in the responseModalities configuration (as indicated by documentation for enabling
   transcriptions), the WebSocket connection immediately terminates with Code: 1007 Reason: Request contains an invalid
   argument..

2. Missing Transcription Messages (Even with Transcription Flags Enabled):
   When Modality.TEXT is removed from responseModalities (to achieve a stable connection) but
   input_audio_transcription: true and output_audio_transcription: true are enabled, the connection is stable and audio
   streams successfully. However, no message.inputAudioTranscription or message.outputAudioTranscription objects are ever
   received in the onmessage callback. This prevents any transcription from being obtained.

This behavior contradicts the official documentation (e.g., GEMINI.md in our project, and conceptual code examples
online) which suggests including TEXT in responseModalities is necessary for transcription. This issue is similar to one
reported for the Python SDK where a version update was required to fix missing transcription fields
(googleapis/python-genai#1279).

Steps to Reproduce:

1. Initialize GoogleGenAI client with apiVersion: "v1alpha".
2. Connect to ai.live.connect using model: 'gemini-2.5-flash-native-audio-preview-09-2025'.
3. Configure the session with:
   • Scenario 1 (Connection Failure):

1 config = {
2 responseModalities: [Modality.AUDIO, Modality.TEXT], // As per documentation for transcription
3 input_audio_transcription: true,
4 output_audio_transcription: true,
5 // ... (other config like systemInstruction, tools)
6 };
* Scenario 2 (Missing Transcription Messages):

1 config = {
2 responseModalities: [Modality.AUDIO], // To achieve stable connection
3 input_audio_transcription: true,
4 output_audio_transcription: true,
5 // ... (other config like systemInstruction, tools)
6 };
4. Observe the WebSocket connection status and onmessage callbacks.

Expected Behavior:

• For Scenario 1: The WebSocket connection should establish successfully, and transcription messages
  (inputAudioTranscription, outputAudioTranscription) should be received in the onmessage callback when audio input is
  provided, along with audio output.
• For Scenario 2: The WebSocket connection should establish successfully, and transcription messages
  (inputAudioTranscription, outputAudioTranscription) should be received in the onmessage callback when audio input is
  provided, along with audio output.

Actual Behavior:

• For Scenario 1: The WebSocket connection immediately closes with Code: 1007 Reason: Request contains an invalid
  argument..
• For Scenario 2: The WebSocket connection establishes and audio streams successfully. However,
  message.inputAudioTranscription and message.outputAudioTranscription are never populated in the onmessage callback.

Relevant Logs/Context:

Scenario 1 (Failure with Modality.TEXT) Logs:

1 Gemini Session Config: {
2   responseModalities: [ 'AUDIO', 'TEXT' ],
3   input_audio_transcription: true,
4   output_audio_transcription: true,
5   systemInstruction: 'You are Lisa, the highly professional, warm, and hyper-efficient AI receptionist...', // (Your
  full instruction text)
6   tools: [ { functionDeclarations: [Array] } ]
7 }
8 Gemini Session Model: gemini-2.5-flash-native-audio-preview-09-2025
9 Gemini Session Opened

10 Gemini Session Closed. Code: 1007 Reason: Request contains an invalid argument.

Scenario 2 (Stable, but no transcription messages) Logs:

(These logs would contain Gemini Session Opened, setupComplete, modelTurn with audio inlineData, generationComplete, but
crucially, no inputAudioTranscription or outputAudioTranscription messages after sending user audio.)

Additional Information:

• We've confirmed saveAudioChunk is working, writing raw pcm data from both user and AI.
• We've confirmed Deepgram can transcribe these saved audio files post-call.
• The issue appears to be specific to the Gemini Live API's handling of transcription with the Node.js SDK at version
  ^1.34.0.