Model Card for whisper-large-v3-formosan-iso-prompt

This model is a early fine-tuned version of the Taiwanese indigenous openai/whisper-large-v3, which uses the ids of each dialect as prompts during training.
Note: we use indonesian as whisper language id

Dialect and Id

• 阿美語: ami
• 賽德克語: sdq
• 太魯閣語: trv

Training process

The training of the model was performed with the following hyperparameters

• Batch size: 32
• Epochs: 4
• Warmup Steps: 1170
• Total Steps: 11700
• Learning rate: 7e-5
• Data augmentation: No

How to use

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
model_id = "formospeech/whisper-large-v3-formosan-iso-prompt"
dialect_id = "ami"
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)
processor = AutoProcessor.from_pretrained(model_id)
pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=128,
    chunk_length_s=30,
    batch_size=16,
    torch_dtype=torch_dtype,
    device=device,
)
generate_kwargs = {"language": "id", "prompt_ids": torch.from_numpy(processor.get_prompt_ids(dialect_id)).to(device)}
transcription = pipe("path/to/my_audio.wav", generate_kwargs=generate_kwargs)
print(transcription.replace(f" {dialect_id}", ""))