CMU ARCTIC Concat15 This dataset contains 140 speech samples formed by concatening utterances from the CMU ARCTIC speech corpus [1]. The dataset is male/female balanced and contains 7 of each. There are in addition 10 samples for each speaker. A single sample was formed by concatenating samples from the CMU ARCTIC corpus until the length exceeds 15 seconds. The following speakers were selected female: axb, clb, eey, ljm, lnh, slp, slt, male: aew, ahw, aup, awb, bdl, fem, gka. Use the dataset A JSON file containing the metadata is provided. The file is structured as follows. { "fs": <sampling rate>, "files": [ "first_file.wav", <all the other files> ], "sorted": { "male": { "aew": [ "cmu_arctic_male_aew_1.wav", <rest of this speaker's files> ], <rest of the male speakers> }, "female": { <all the female speakers similar to male speakers> } } } There are two functions provided to help selecting the files, sampling and wav_read_center. The first will select a number of distinct subset of speakers. sampling(num_subsets, num_speakers, metadata_file, gender_balanced=False, seed=None): This function will pick automatically and at random subsets speech samples from a list generated using this file. Parameters ---------- num_subsets: int Number of subsets to create num_speakers: int Number of distinct speakers desired in a subset metadata_file: str Location of the metadata file gender_balanced: bool, optional If True, the subsets will have a the same number of male/female speakers when `num_speakers` is even, and one extra male, when `num_speakers` is odd. Default is `False`. seed: int, optional When a seed is provided, the random number generator is fixed to a deterministic state. This is useful for getting consistently the same set of speakers. The initial state of the random number generator is restored at the end of the function. When not provided, the random number generator is used without setting the seed. Returns ------- A list of `num_subsets` lists of wav filenames, each containing `num_speakers` entries. The second reads a bunch of wav files, adjust their length so that they’re all the same size and puts them in a numpy array. wav_read_center(wav_list, center=True, seed=None): Read a bunch of wav files, equalize their length and puts them in a numpy array Parameters ---------- wav_list: list of str A list of file names, the file names should be of format wav and monaural center: bool, optional When True (default), the signals will be centered, otherwise, only their end will be zero padded seed: int Provides a seed for the random number generator. When this is provided, center option is ignored and the beginning of segments is placed at random within the maximum length available Returns ------- ndarray (n_files, n_samples) A 2D array that contains one signal per row So a typical use could be like this. from generate_samples import sampling, wav_read_center # Create 10 groups of 3 speakers, deterministic groups = sampling(10, 3, "metadata.json", seed=0) # Read all the sound files in the first group # shape: (n_signals, n_samples) signals = wav_read_center(groups[0]) Generate the dataset The dataset was generated with the generate_samples.py script which relies on numpy, scipy, and pyroomacoustics. It can be re-generated with the following command. python ./generate_samples.py -s 7 -n 10 -d 15 [--cmudir /path/to/CMU/corpus] If the --cmudir option is not provided, the whole CMU ARCTIC corpus will be downloaded automatically. Also, the dataset will be cached in a file called cmu_arctic.dat. This file takes about a gigabyte of disk space, so you might want to remove it when you’re done. References [1] J. Kominek and A. W. Black, “CMU ARCTIC databases for speech synthesis,” CMU-LTI-03-177, 2003.