| import soundfile as sf |
| import numpy as np |
| import soxr |
|
|
| def resample(samples: np.ndarray, orig_sr: int, target_sr: int) -> np.ndarray: |
| """ |
| Resamples a mono float array to target_sr with soxr (anti-aliased, HQ). |
| Returns the input unchanged when the rates already match. |
| """ |
| if orig_sr == target_sr: |
| return samples |
| return soxr.resample(samples, orig_sr, target_sr, quality="HQ").astype(np.float32) |
|
|
| def condition_audio(samples: np.ndarray, sample_rate: int = 16000, pad_s: float = 0.2, |
| target_peak: float = 0.9, max_gain: float = 10.0) -> np.ndarray: |
| """ |
| Prepares a mono clip for the model: removes DC offset, scales the peak to |
| target_peak (gain capped at max_gain so near-silence is not blown up), and |
| pads pad_s seconds of zeros at both ends. An all-zero clip skips scaling. |
| Returns float32, 1-D. |
| """ |
| x = np.asarray(samples, dtype=np.float32).reshape(-1) |
| if x.size: |
| x = x - np.mean(x, dtype=np.float32) |
| peak = float(np.max(np.abs(x))) |
| if peak > 0.0: |
| x = x * min(target_peak / peak, max_gain) |
| pad = np.zeros(int(round(pad_s * sample_rate)), dtype=np.float32) |
| return np.concatenate([pad, x.astype(np.float32), pad]) |
|
|
| def load_and_resample(file_path: str, target_sr: int = 16000) -> tuple[np.ndarray, float]: |
| """ |
| Loads an audio file, converts stereo to mono if needed, |
| and resamples to target_sr using soxr if required. |
| Returns: |
| samples: np.ndarray (float32, 1D array) |
| duration_s: float |
| """ |
| data, sr = sf.read(file_path, dtype="float32") |
| |
| |
| if data.ndim > 1: |
| data = np.mean(data, axis=1) |
| |
| |
| if sr != target_sr: |
| data = resample(data, sr, target_sr) |
| sr = target_sr |
| |
| duration_s = len(data) / float(sr) |
| return data, duration_s |
|
|
