---
library_name: transformers
language:
- zh
license: mit
base_model: microsoft/Phi-4-multimodal-instruct
tags:
- automatic-speech-recognition
- audio
- speech
- generated_from_trainer
datasets:
- JacobLinCool/common_voice_19_0_zh-TW
metrics:
- wer
- cer
model-index:
- name: Phi-4-multimodal-instruct-commonvoice-zh-tw
  results:
  - task:
      type: automatic-speech-recognition
      name: Automatic Speech Recognition
    dataset:
      name: JacobLinCool/common_voice_19_0_zh-TW
      type: JacobLinCool/common_voice_19_0_zh-TW
    metrics:
    - type: wer
      value: 31.18
      name: Wer
    - type: cer
      value: 6.67
      name: Cer
---

# Phi-4-multimodal-instruct-commonvoice-zh-tw

This model is a fine-tuned version of [microsoft/Phi-4-multimodal-instruct](https://huggingface.co/microsoft/Phi-4-multimodal-instruct) on the [Common Voice 19.0 Taiwanese Mandarin dataset](https://huggingface.co/datasets/JacobLinCool/common_voice_19_0_zh-TW).

- WER: 31.18%
- CER: 6.67%

## Model description

Phi-4-multimodal-instruct-commonvoice-zh-tw is a multimodal language model fine-tuned for Automated Speech Recognition (ASR) of Taiwanese Mandarin (zh-TW). The base model is Microsoft's Phi-4-multimodal-instruct, which was further trained on speech transcription tasks.

The model accepts audio input and produces Traditional Chinese text transcriptions. It has been specifically optimized to recognize Taiwanese Mandarin speech patterns and vocabulary.

## Intended uses & limitations

This model is intended for:
- Transcribing spoken Taiwanese Mandarin to text
- Automated subtitling/captioning for zh-TW content
- Speech-to-text applications requiring Taiwanese Mandarin support

Limitations:
- Performance may vary with background noise, speaking speed, or accents
- The model performs best with clear audio input
- Specialized terminology or domain-specific vocabulary may have lower accuracy

## Training and evaluation data

The model was fine-tuned on Common Voice 19.0 Taiwanese Mandarin dataset. Common Voice is a crowdsourced speech dataset containing contributions from volunteers who record themselves reading sentences in various languages.

The evaluation was performed on the test split of the same dataset, consisting of 5,013 samples.

## Training procedure

The model was trained using LoRA adapters focused on the speech recognition components of the base model, allowing for efficient fine-tuning while preserving the general capabilities of the underlying Phi-4 model.

### Prompt format

This model follows the prompt template from the original paper. For speech recognition tasks, the audio input is provided inline with a simple instruction:

```
<|user|>
<|audio_1|> Transcribe the audio clip into text.
<|assistant|>
[Transcription output in Traditional Chinese]
<|end|>
```

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 4e-05
- train_batch_size: 4
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 32
- total_train_batch_size: 128
- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.95) and epsilon=1e-07 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 50
- num_epochs: 2

### Training results

The model achieved the following performance metrics on the test set:
- Word Error Rate (WER): 31.18%
- Character Error Rate (CER): 6.67%
- Number of evaluation samples: 5,013

### Framework versions

- Transformers 4.49.0
- Pytorch 2.4.1+cu124
- Datasets 3.3.2
- Tokenizers 0.21.1

## How to use

```python
import torch
from transformers import AutoProcessor, AutoModelForCausalLM
import librosa

AUDIO_PATH = "test.wav"

MODEL = "JacobLinCool/Phi-4-multimodal-instruct-commonvoice-zh-tw"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
USE_FA = True

processor = AutoProcessor.from_pretrained(MODEL, trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    torch_dtype=torch.bfloat16 if USE_FA else torch.float32,
    _attn_implementation="flash_attention_2" if USE_FA else "sdpa",
    trust_remote_code=True,
).to(DEVICE)

audio, sr = librosa.load(AUDIO_PATH, sr=16000)

# Prepare the user message and generate the prompt
user_message = {
    "role": "user",
    "content": "<|audio_1|> Transcribe the audio clip into text.",
}
prompt = processor.tokenizer.apply_chat_template(
    [user_message], tokenize=False, add_generation_prompt=True
)

# Build the inputs for the model
inputs = processor(text=prompt, audios=[(audio, sr)], return_tensors="pt")
inputs = {k: v.to(model.device) if hasattr(v, "to") else v for k, v in inputs.items()}

# Generate transcription without gradients
with torch.no_grad():
    generated_ids = model.generate(
        **inputs,
        eos_token_id=processor.tokenizer.eos_token_id,
        max_new_tokens=64,
        do_sample=False,
    )

# Decode the generated token IDs into a human-readable transcription
transcription = processor.decode(
    generated_ids[0, inputs["input_ids"].shape[1] :],
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False,
)

# Print the transcription
print(transcription)
```