OpenAI Audio: JSON transcription responses discard non-text metadata

Bug description
The AudioTranscriptionResponse or AudioTranscription implementation discards valuable JSON metadata (language, segments, duration, etc.) from OpenAI's audio transcription API responses. When responseFormat.isJsonType() is true, only the text field is preserved , despite the API returning structured JSON containing additional fields. This violates the principle of least surprise since users explicitly requesting JSON format expect full structured data.

Environment

• Spring AI version: 1.1.0-SNAPSHOT (current main branch)
• Java version: 21
• Model Provider: OpenAI
• Relevant classes: AudioTranscriptionResponse, AudioTranscription, AudioTranscriptionMetadata

Steps to reproduce

1. Configure audio transcription with JSON response format:

   spring.ai.openai.audio.options.response-format=verbose_json
   

2. Transcribe audio containing metadata (e.g., segments/timestamps)
3. Inspect AudioTranscriptionResponse:

   AudioTranscriptionResponse response = client.call(prompt);
   response.getResults().get(0); // Only contains text
   response.getMetadata();   // Empty, lacks JSON body fields
   

Expected behavior

• Full JSON response should be accessible in metadata
• AudioTranscription should include:

  class AudioTranscription(
    // not sure if this should be a JSON String 
    // maybe nullable(or empty) fields such as segments and words in `AudioTranscription` or `AudioTranscriptionResponse` might work
    private final String text; 
    private AudioTranscriptionMetadata transcriptionMetadata;
  ) {}
  

• Backward compatibility maintained for non-JSON responses

Minimal Complete Reproducible example

@Test
void shouldReturnFullJsonMetadata() {
	OpenAiAudioTranscriptionOptions transcriptionOptions = OpenAiAudioTranscriptionOptions.builder()
			.responseFormat(TranscriptResponseFormat.VERBOSE_JSON)
			.temperature(0f)
			.build();
	AudioTranscriptionPrompt transcriptionRequest = new AudioTranscriptionPrompt(this.audioFile,
			transcriptionOptions);
	AudioTranscriptionResponse response = this.transcriptionModel.call(transcriptionRequest);
	assertThat(response.getResults()).hasSize(1);

	// this fails
	assertThat(response.getResults().get(0))
			.isEqualTo(new AudioTranscription("[{\"id\":0,\"seek\":0,\"start\":0.0,\"end\":11.0,\"text\":\"And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.\",\"tokens\":[50364,400,370,452,7177,6280,11,1029,406,437,428,1941,393,360,337,291,11,1029,437,291,393,360,337,428,1941,13,50914],\"temperature\":0.0,\"avgLogprob\":-0.3053564,\"compressionRatio\":1.3544304,\"noSpeechProb\":0.03939722}]"));

	// or maybe metadata should not be empty?
	assertThat(response.getMetadata().isEmpty()).isTrue();
}

Start by reading AudioTranscriptionResponse, AudioTranscription, and AudioTranscriptionMetadata, then run the provided shouldReturnFullJsonMetadata reproducer with verbose_json. The change is done when full JSON metadata is accessible for JSON responses while non-JSON transcription behavior remains backward compatible.