Method _add_noise

espnet2/train/preprocessor.py:309–381

(
        self,
        speech,
        power,
        noises,
        noise_db_low,
        noise_db_high,
        tgt_fs=None,
        single_channel=False,
    )

        return speech, rir

    def _add_noise(
        self,
        speech,
        power,
        noises,
        noise_db_low,
        noise_db_high,
        tgt_fs=None,
        single_channel=False,
    ):
        nsamples = speech.shape[1]
        noise_path = np.random.choice(noises)
        noise = None
        if noise_path is not None:
            noise_db = np.random.uniform(noise_db_low, noise_db_high)
            with soundfile.SoundFile(noise_path) as f:
                fs = f.samplerate
                if tgt_fs and fs != tgt_fs:
                    nsamples_ = int(nsamples / tgt_fs * fs) + 1
                else:
                    nsamples_ = nsamples
                if f.frames == nsamples_:
                    noise = f.read(dtype=np.float64, always_2d=True)
                elif f.frames < nsamples_:
                    if f.frames / nsamples_ < self.short_noise_thres:
                        logging.warning(
                            f"Noise ({f.frames}) is much shorter than "
                            f"speech ({nsamples_}) in dynamic mixing"
                        )
                    offset = np.random.randint(0, nsamples_ - f.frames)
                    # noise: (Time, Nmic)
                    noise = f.read(dtype=np.float64, always_2d=True)
                    # Repeat noise
                    noise = np.pad(
                        noise,
                        [(offset, nsamples_ - f.frames - offset), (0, 0)],
                        mode="wrap",
                    )
                else:
                    offset = np.random.randint(0, f.frames - nsamples_)
                    f.seek(offset)
                    # noise: (Time, Nmic)
                    noise = f.read(nsamples_, dtype=np.float64, always_2d=True)
                    if len(noise) != nsamples_:
                        raise RuntimeError(f"Something wrong: {noise_path}")
            if single_channel:
                num_ch = noise.shape[1]
                chs = [np.random.randint(num_ch)]
                noise = noise[:, chs]
            # noise: (Nmic, Time)
            noise = noise.T
            if tgt_fs and fs != tgt_fs:
                logging.warning(
                    f"Resampling noise to match the sampling rate ({fs} -> {tgt_fs} Hz)"
                )
                noise = librosa.resample(
                    noise, orig_sr=fs, target_sr=tgt_fs, res_type="kaiser_fast"
                )