Phi4-mm responds that the no audio before and after text.

hi,
I am trying to run Phi4-mm using Librispeech data. Im following the steps mentioned in
https://github.com/microsoft/onnxruntime-genai/blob/ec8649357f29f159cea86d289ad413b71231f0e6/examples/python/phi4-mm.py.

based on the logs, the audio_embeds value is as follows.
Name: audio_embeds
Shape: [1, 349, 80]
Type: Float16
Location: CPU
Values: Starts with 9.42188 9.32812 8.96875 6.90234 6.05078 ...
Stats:
Min: 0 at index [169]
Max: 25.7656 at index [25985]
Mean: 14.144
StdDev: 4.09015

The response I get from the model is "Since there's nothing before or after the provided text, I can't transcribe anything. If you want to transcribe a different text or need help with something else, feel free to provide the information!" with temperature=1.0.
For temperature<1.0, response is empty.

I am able to get correct response with only text based prompt like "tell me a joke" with my setup.

Start with examples/python/phi4-mm.py and reproduce the Librispeech run using the reported audio_embeds shape and prompt. Compare the multimodal invocation with the working text-only prompt and inspect the generated response at different temperatures. Done means the example produces a meaningful transcription for the supplied audio rather than the no-audio response.

Written by the indexing model from the issue text.