processing_saga.py

import numpy as np
from transformers import ProcessorMixin

class SagaProcessor(ProcessorMixin):
    attributes = ["feature_extractor", "tokenizer"]

    def __init__(self, feature_extractor, tokenizer, **kwargs):
        super().__init__(feature_extractor, tokenizer, **kwargs)
        self.target_sr = 16000

    def process_audio(self, audio, sampling_rate):
        if int(sampling_rate) == self.target_sr:
            return audio

        src_len = audio.shape[0]
        dst_len = int(round(src_len * (float(self.target_sr) / float(sampling_rate))))
        
        if dst_len <= 1:
            return np.zeros((0,), dtype=np.float32)

        src_x = np.linspace(0.0, 1.0, num=src_len, endpoint=False)
        dst_x = np.linspace(0.0, 1.0, num=dst_len, endpoint=False)
        return np.interp(dst_x, src_x, audio).astype(np.float32)

    def get_prompt(self):
        messages = [
            {"role": "system", "content": ""},
            {"role": "user", "content": [{"type": "audio", "audio": ""}]},
        ]
        prompt = self.tokenizer.apply_chat_template(
            messages,
            add_generation_prompt=True,
            tokenize=False,
        )
        return prompt + "language Danish<asr_text>"