The model was initially fine-tuned using the original ASR prompt: "Transcribe the audio clip into text."
This prompt is language agnostic—as described in the model paper:

The ASR prompt for Phi-4-Multimodal is “Transcribe the audio clip into text.”, which is language agnostic. We notice that the model can learn to recognize in the target language perfectly without providing language information, while Qwen2-audio and Gemini-2.0-Flash require the language information in the prompt to obtain the optimal ASR performance.

However, we found that using a language-defining prompt, such as: "Transcribe the Turkish audio." leads to better performance.
See: ysdede/Phi-4-mm-inst-asr-turkish

Training Results

When benchmarked with the original ASR prompt "Transcribe the audio clip into text.", the evaluation results were as follows:

Before Fine-Tuning:
- WER: 153.84
- CER: 82.57
After Fine-Tuning:
- WER: 64.76
- CER: 29.85

Inference

Load generation_config and processor from the base model as a quick fix to use the default generation settings.

Note: The new models currently lack high-quality fine-tuning scripts. When saving a fine-tuned model using model.save_pretrained(), the processor configuration—including essential audio parameters—is not automatically saved. This omission can lead to errors during inference due to the model’s complex architecture. Loading these components from the base model ensures that all critical settings are properly included.

python

generation_config = GenerationConfig.from_pretrained(
    'microsoft/Phi-4-multimodal-instruct', 'generation_config.json'
)
processor = AutoProcessor.from_pretrained(
    'microsoft/Phi-4-multimodal-instruct', trust_remote_code=True
)

Framework versions

Transformers 4.46.1
Pytorch 2.5.1+cu124
Datasets 3.3.2
Tokenizers 0.20.3

ysdede/Phi-4-mm-inst-asr-turkish-3

Phi-4-mm-inst-asr-turkish-3

Training Prompt

Training Results

Inference

Framework versions

ysdede/
Phi-4-mm-inst-asr-turkish-3