Method _load_noise

espnet2/train/preprocessor.py:2146–2179

(self, speech, speech_db, noises, noise_db_low, noise_db_high)

        return speech, rir

    def _load_noise(self, speech, speech_db, noises, noise_db_low, noise_db_high):
        nsamples = speech.shape[1]
        noise_path = np.random.choice(noises)
        noise = None
        if noise_path is not None:
            noise_snr = np.random.uniform(noise_db_low, noise_db_high)
            with soundfile.SoundFile(noise_path) as f:
                if f.frames == nsamples:
                    noise = f.read(dtype=np.float64)
                elif f.frames < nsamples:
                    # noise: (Time,)
                    noise = f.read(dtype=np.float64)
                    # Repeat noise
                    noise = np.pad(
                        noise,
                        (0, nsamples - f.frames),
                        mode="wrap",
                    )
                else:
                    offset = np.random.randint(0, f.frames - nsamples)
                    f.seek(offset)
                    # noise: (Time,)
                    noise = f.read(nsamples, dtype=np.float64)
                    if len(noise) != nsamples:
                        raise RuntimeError(f"Something wrong: {noise_path}")
            # noise: (Nmic, Time)
            noise = noise[None, :]

            noise_power = np.mean(noise**2)
            noise_db = 10 * np.log10(noise_power + 1e-4)
            scale = np.sqrt(10 ** ((speech_db - noise_db - noise_snr) / 10))

            noise = noise * scale
        return noise

    def _apply_data_augmentation(self, speech):
        # speech: (Nmic, Time)