A collection of large publically available speech datasets for biometric speaker identification, veriffication and diarization, voice cloning, speech enhancement and denoising tasks. Because most audio datasets focus on speech-to-text domain, the motivation was to separate speech datasets for solving text-independent tasks.

• VoxCeleb - audio-visual dataset collected from interviews of famous people; consists of over 7000 speakers of different ethnicities, accents, and ages; 2000 hours of data collected in real-world environmets with backgroud noises.

• VoxMovies - audio dataset with 856 speakers from the VoxCeleb dataset with varying emotion,accents and noise.

• VCTK - 48kHz audio dataset consisting of 110 speakers with different accents reading the same 400 sentences; each sentence is recorded by two different microphones.

• LibriSpeech - 1000 hours of 16kHz audio data, contains both clean and noisy examples.

• LibriAdapt - dataset constructed from LibriSpeech with demain shifts due to different microphones, speaker accents and acoustic environments.

• LibriTTS - dataset constructed from LibriSpeech more suited for speech synthesis, consists of 585 hours of English speech, 24kHz sampling rate, utterances separated when sentences end.

• TIMIT - 16kHz audio data of 630 speakers, each reading 10 identical sentences.

• M-AILABS - 16kHz audio data in 8 languages, contains speaker idenities and genders constructed from audiobooks.

• DCASE - a collection of background sounds that usually occur on the street, in a transport, in an office or at home, such as baby crying, car breaking, people walking, alarms, screams etc. Can be used to train models robust to real-environment background noises.