| import torch |
| import random |
| import torchaudio |
| import numpy as np |
| from typing import Union |
| from config import SharedConfig |
| import torchaudio.transforms as T |
|
|
|
|
| def load_and_resample(wav_file: str, target_sr: Union[int, None]): |
| wav_seg, orig_sr = torchaudio.load(wav_file) |
| wav_seg = wav_seg.mean(0) |
| if target_sr is None: |
| return wav_seg |
| return resample(wav_seg, orig_sr, target_sr) |
|
|
|
|
| def resample(wav_seg: torch.Tensor, orig_sr: int, target_sr: int): |
| resampler = T.Resample(orig_freq=orig_sr, new_freq=target_sr) |
| return resampler(wav_seg) |
|
|
|
|
| def normalize(wav_seg: torch.Tensor): |
| return wav_seg / wav_seg.abs().max() |
|
|
|
|
| def instrument_to_release_time(instrument): |
| instrument = instrument.strip() |
| if instrument == "keyboard": |
| return random.uniform(0.1, 1.0) |
| if instrument == "mallet": |
| return random.uniform(0.1, 1.0) |
| if instrument == "organ": |
| return random.uniform(0.1, 1.0) |
| if instrument == "guitar": |
| return random.uniform(0.1, 0.5) |
| if instrument == "bass": |
| return random.uniform(0.1, 0.2) |
| if instrument == "string": |
| return random.uniform(0.8, 1.0) |
| if instrument == "brass": |
| return random.uniform(0.1, 1.0) |
| if instrument == "reed": |
| return random.uniform(0.8, 1.0) |
| if instrument == "flute": |
| return random.uniform(0.8, 1.0) |
| if instrument == "synth": |
| return random.uniform(0.1, 1.0) |
| if instrument == "vocal": |
| return random.uniform(0.1, 1.0) |
| if instrument == "other": |
| return random.uniform(0.1, 1.0) |
| if instrument == "drums": |
| return 0 |
| raise ValueError(f"Unknown instrument: {instrument}") |
|
|
|
|
| class AudioSegmenter: |
| def __init__(self, config: SharedConfig): |
| self.input_sec = config.input_sec |
|
|
| def chunk_audio_and_notes(self, waveform: np.ndarray): |
| waveform = torch.from_numpy(waveform) |
| assert waveform.ndim == 1 |
| wave_length = waveform.shape[0] |
| chunks = list(waveform.split(int(self.input_sec * self.sample_rate), dim=0)) |
| return chunks |
|
|
