batch.waveforms_from_files_into_numpys(start_time_milliseconds=0, end_time_milliseconds=0, frame_rate_hz=0, num_channels=0, convert_to_mono=False, zero_pad_ending=False, repeat_pad_ending=False, resample_mode=0, decoding_backend=0, num_workers=0)

Uses multithreading in Rust to decode many audio files in parallel, directly returning a list of NumPyNamedResult objects instead of Waveform objects.

This method is not as ergonomic as waveforms_from_files() for decoding audio files in parallel. However, this method avoids unnecessary memory allocations and Python GIL contention.

Remember to raise exceptions when needed

waveforms_from_files_into_numpy() will return

exceptions but will not raise them for you. It is your responsibility to check every exception field for a not-None exception that you can raise or intentionally ignore.

  • filenames (list[str]) – A list of filenames–each as str–to decode in parallel.

  • start_time_milliseconds (int, optional) – We discard any audio before this millisecond offset. By default, this does nothing and the audio is decoded from the beginning. Negative offsets are invalid.

  • end_time_milliseconds (int, optional) – We discard any audio after this millisecond offset. By default, this does nothing and the audio is decoded all the way to the end. If start_time_milliseconds is specified, then end_time_milliseconds must be greater. The resulting

  • frame_rate_hz (int, optional) – A destination frame rate to resample the audio to. Do not specify this parameter if you wish Babycat to preserve the audio’s original frame rate. This does nothing if frame_rate_hz is equal to the audio’s original frame rate.

  • num_channels (int, optional) – Set this to a positive integer n to select the first n channels stored in the audio file. By default, Babycat will return all of the channels in the original audio. This will raise an exception if you specify a num_channels greater than the actual number of channels in the audio.

  • convert_to_mono (bool, optional) – Set to True to average all channels into a single monophonic (mono) channel. If num_channels = n is also specified, then only the first n channels will be averaged. Note that convert_to_mono cannot be set to True while also setting num_channels = 1.

  • zero_pad_ending (bool, optional) – If you set this to True, Babycat will zero-pad the ending of the decoded waveform to ensure that the output waveform’s duration is exactly end_time_milliseconds - start_time_milliseconds. By default, zero_pad_ending = False, in which case the output waveform will be shorter than end_time_milliseconds - start_time_milliseconds if the input audio is shorter than end_time_milliseconds.

  • resample_mode (int, optional) – If you set frame_rate_hz to resample the audio when decoding, you can also set resample_mode to pick which resampling backend to use. The babycat.resample_mode submodule contains the various available resampling algorithms compiled into Babycat. By default, Babycat resamples audio using libsamplerate at its highest-quality setting.

  • decoding_backend (int, optional) – Sets the audio decoding backend to use. Defaults to the Symphonia backend.

  • num_workers (int, optional) – The number of threads–Rust threads, not Python threads–to use for parallel decoding of the audio files in filenames. By default, Babycat creates the same number of threads as the number of logical CPU cores on your machine.


A list of objects that contain a ndarray for every successful encoding and a Python exception for every failed encoding. Look at the "Raises" section of Waveform.decode_from_filename() for a list of possible exceptions that can be returned by this method.

Return type