SoFilNet: Self-Supervised Source-Filter Model of Speech Production

Audio samples

Samples

Utterances from the TIMIT test set (VTR-FORMANTS subset), unseen during training. All audio is 16 kHz.

Utterance	Original	Parameters only	Estimated magnitude, original phase	Original magnitude, estimated phase
dr1/felc0/si2016 Female speaker “Heave on those ropes the boat's come unstuck.”
dr2/mwew0/sx11 Male speaker “He will allow a rare lie.”
dr2/fpas0/si944 Female speaker “If the farm is rented the rent must be paid.”
dr8/mjln0/si1449 Male speaker “Chief interest in radiosterilization resides in the military services.”
dr4/fjlm0/si1043 Female speaker “Not surprisingly this approach did not work.”
dr4/mtls0/si2000 Male speaker “We can get it if we dig he said patiently.”
dr7/fdhc0/si929 Female speaker “But such cases were in the past unusual.”
dr7/mgrt0/si2080 Male speaker “He would not carry a brief case.”