Vikentii Pankov, Valeria Pronina, Alexander Kuzmin, Maksim Borisov, Nikita Usoltsev, Xingshan Zeng, Alexander Golubkov, Nikolai Ermolenko, Aleksandra Shirshova, Yulia Matveeva
Abstract
We address zero-shot TTS systems’ noise-robustness problem by proposing a dual-objective training for the speaker encoder using self-supervised DINO loss. This approach enhances the speaker encoder with the speech synthesis objective, capturing a wider range of speech characteristics beneficial for voice cloning. At the same time, the DINO objective improves speaker representation learning, ensuring robustness to noise and speaker discriminability. Experiments demonstrate significant improvements in subjective metrics under both clean and noisy conditions, outperforming traditional speaker-encoder-based TTS systems. Additionally, we explore training zero-shot TTS on noisy, unlabeled data. Our two-stage training strategy, leveraging self-supervised speech models to distinguish between noisy and clean speech, shows notable advances in similarity and naturalness, especially with noisy training datasets, compared to the ASR-transcription-based approach.
[*] this page is intended for research demonstration purposes only.
System architecture
Robustness to noises at inference and general similarity improvement
text | speaker | gender | noise | reference | DINO-VITS | AAM-Softmax-VITS | YourTTS | BYOL-A |
---|---|---|---|---|---|---|---|---|
HE PASSED THROUGH HENLEY SAINT ALBANS AND CAME SO NEAR TO LONDON AS HARROW ON THE HILL | adi | male | street | |||||
THE INFLUENCE WITH THE TIMAEUS HAS EXERCISED UPON POSTERITY IS DUE PARTLY TO A MISUNDERSTANDING | casey | male | marathon | |||||
BY QUICK MARCHES THROUGH THESE INACCESSIBLE MOUNTAINS THAT GENERAL FREED HIMSELF FROM THE SUPERIOR FORCES OF THE COVENANTERS | casey | male | outdoor | |||||
HER SEA GOING QUALITIES WERE EXCELLENT AND WOULD HAVE AMPLY SUFFICED FOR A CIRCUMNAVIGATION OF THE GLOBE | jess | female | booth | |||||
I CAN'T SEE IT IN THAT LIGHT SAID THE OLD LAWYER | jess | female | booth | |||||
I DID NOT WRONG MYSELF SO BUT I PLACED A WRONG ON THEE | jess | female | hairdryer | |||||
THIS REALITY BEGINS TO EXPLAIN THE DARK POWER AND OTHERWORLDLY FASCINATION OF TWENTY THOUSAND LEAGUES UNDER THE SEAS | jess | female | hairdryer | |||||
HER SEA GOING QUALITIES WERE EXCELLENT AND WOULD HAVE AMPLY SUFFICED FOR A CIRCUMNAVIGATION OF THE GLOBE | jess | female | hairdryer | |||||
IF SHE COULD ONLY SEE PHRONSIE FOR JUST ONE MOMENT | jess | female | street | |||||
IT IS MANIFEST THAT MAN IS NOW SUBJECT TO MUCH VARIABILITY | kallmekris | female | booth | |||||
IN SHORT HE BECOMES A PROMINENT FIGURE IN LONDON SOCIETY AND IF HE IS NOT CAREFUL SOMEBODY WILL SAY SO | kallmekris | female | cafe | |||||
EVEN SO I HAD JUST RETURNED FROM AN ARDUOUS JOURNEY EXHAUSTED AND BADLY NEEDING A REST | kallmekris | female | cafe | |||||
FOR A FULL HOUR HE HAD PACED UP AND DOWN WAITING BUT HE COULD WAIT NO LONGER | kallmekris | female | cafe | |||||
IT WAS THE SCARLET LETTER IN ANOTHER FORM THE SCARLET LETTER ENDOWED WITH LIFE | kallmekris | female | cafe | |||||
UPON THIS MADAME DEIGNED TO TURN HER EYES LANGUISHINGLY TOWARDS THE COMTE OBSERVING | kallmekris | female | street | |||||
THEY WERE CERTAINLY NO NEARER THE SOLUTION OF THEIR PROBLEM | kallmekris | female | waterfall | |||||
IT IS SIXTEEN YEARS SINCE JOHN BERGSON DIED | kallmekris | female | waterfall | |||||
SOMEONE ELSE TOLD A STORY NOT PARTICULARLY EFFECTIVE WHICH I SAW HE WAS NOT FOLLOWING | kallmekris | female | waterfall | |||||
CONCORD RETURNED TO ITS PLACE AMIDST THE TENTS | kallmekris | female | waterfall | |||||
THE SALIENT FEATURES OF THIS DEVELOPMENT OF DOMESTIC SERVICE HAVE ALREADY BEEN INDICATED | myra | female | booth | |||||
THE PARIS PLANT LIKE THAT AT THE CRYSTAL PALACE WAS A TEMPORARY EXHIBIT | nadine | female | inside | |||||
GENTLEMEN TO YOUR POSTS WHEREUPON SAINT AIGNAN AND VILLEROY TOOK THEIR LEAVE | nadine | female | loud | |||||
BUT THE MORE FORGETFULNESS HAD THEN PREVAILED THE MORE POWERFUL WAS THE FORCE OF REMEMBRANCE WHEN SHE AWOKE | nadine | female | outdoor | |||||
YOU WILL FIND ME CONTINUALLY SPEAKING OF FOUR MEN TITIAN HOLBEIN TURNER AND TINTORET IN ALMOST THE SAME TERMS | nadine | female | outdoor | |||||
ALSO A POPULAR CONTRIVANCE WHEREBY LOVE MAKING MAY BE SUSPENDED BUT NOT STOPPED DURING THE PICNIC SEASON | nadine | female | outdoor | |||||
WELL MOTHER SAID THE YOUNG STUDENT LOOKING UP WITH A SHADE OF IMPATIENCE | nadine | female | outdoor | |||||
LENGTH OF SERVICE FOURTEEN YEARS THREE MONTHS AND FIVE DAYS | nadine | female | restaurant | |||||
MISS LAKE DECLINED THE CARRIAGE TO NIGHT | nadine | female | restaurant | |||||
WHAT IS YOUR COUNTRY OLAF HAVE YOU ALWAYS BEEN A THRALL THE THRALL'S EYES FLASHED | sawyer | male | crowdstreet | |||||
HE BEGAN A CONFUSED COMPLAINT AGAINST THE WIZARD WHO HAD VANISHED BEHIND THE CURTAIN ON THE LEFT | sawyer | male | crowdstreet | |||||
YES DEAD THESE FOUR YEARS AN A GOOD JOB FOR HER TOO | sawyer | male | inside | |||||
IN EVERY WAY THEY SOUGHT TO UNDERMINE THE AUTHORITY OF SAINT PAUL | sawyer | male | restaurant | |||||
YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION | sawyer | male | restaurant |
Training from noisy untranscribed data
text | speaker | gender | noise | reference | Whisper transcripts from original training data | Whisper transcripts with 35% of training data augmented with noises (SNR=0db) | HuBERT units from original training data | HuBERT units with 35% of training data augmented with noises (SNR=0db) |
---|---|---|---|---|---|---|---|---|
AND HOW ODD THE DIRECTIONS WILL LOOK | kallmekris | female | cafe | |||||
THE PARIS PLANT LIKE THAT AT THE CRYSTAL PALACE WAS A TEMPORARY EXHIBIT | nadine | female | inside | |||||
THE DEMOCRATIC COMMITTEE FIGURED OUT A WAY TO DO THIS | myra | female | street | |||||
I DID NOT WRONG MYSELF SO BUT I PLACED A WRONG ON THEE | jess | female | hairdryer | |||||
IT WAS THE SCARLET LETTER IN ANOTHER FORM THE SCARLET LETTER ENDOWED WITH LIFE | kallmekris | female | cafe | |||||
MY FRIEND'S TEMPER HAD NOT IMPROVED SINCE HE HAD BEEN DEPRIVED OF THE CONGENIAL SURROUNDINGS OF BAKER STREET | kallmekris | female | booth | |||||
IN EVERY WAY THEY SOUGHT TO UNDERMINE THE AUTHORITY OF SAINT PAUL | sawyer | male | restaurant | |||||
LENGTH OF SERVICE FOURTEEN YEARS THREE MONTHS AND FIVE DAYS | nadine | female | restaurant | |||||
THE STOP AT QUEENSTOWN THE TEDIOUS PASSAGE UP THE MERSEY WERE THINGS THAT HE NOTED DIMLY THROUGH HIS GROWING IMPATIENCE | casey | male | booth | |||||
NATURE OF THE EFFECT PRODUCED BY EARLY IMPRESSIONS | nadine | female | booth | |||||
IT IS SIXTEEN YEARS SINCE JOHN BERGSON DIED | kallmekris | female | waterfall | |||||
WHY ARE WE TO BE DENIED EACH OTHER'S SOCIETY | nadine | female | outdoor | |||||
MISS LAKE DECLINED THE CARRIAGE TO NIGHT | jess | female | booth | |||||
THEY WERE CERTAINLY NO NEARER THE SOLUTION OF THEIR PROBLEM | kallmekris | female | waterfall | |||||
STUFF IT INTO YOU HIS BELLY COUNSELLED HIM | kallmekris | female | wind | |||||
CONCORD RETURNED TO ITS PLACE AMIDST THE TENTS | kallmekris | female | waterfall | |||||
IF SHE COULD ONLY SEE PHRONSIE FOR JUST ONE MOMENT | nadine | female | restaurant | |||||
ON ARRIVING AT HOME AT MY OWN RESIDENCE I FOUND THAT OUR SALON WAS FILLED WITH A BRILLIANT COMPANY | kallmekris | female | cafe | |||||
JOHN TAYLOR WHO HAD SUPPORTED HER THROUGH COLLEGE WAS INTERESTED IN COTTON | casey | male | booth | |||||
FOR A FULL HOUR HE HAD PACED UP AND DOWN WAITING BUT HE COULD WAIT NO LONGER | kallmekris | female | cafe | |||||
WHAT IS YOUR COUNTRY OLAF HAVE YOU ALWAYS BEEN A THRALL THE THRALL'S EYES FLASHED | sawyer | male | crowdstreet | |||||
IT ENGENDERS A WHOLE WORLD LA PEGRE FOR WHICH READ THEFT AND A HELL LA PEGRENNE FOR WHICH READ HUNGER | nadine | female | outdoor | |||||
SATURDAY AUGUST FIFTEENTH THE SEA UNBROKEN ALL ROUND NO LAND IN SIGHT | kallmekris | female | cafe | |||||
HE HAD HIS HAND UPON LAKE'S SHOULDER | kallmekris | female | cafe | |||||
IT WAS THE INDIAN WHOSE DARK SILHOUETTE APPEARED SUDDENLY UPON HIS BLIND | nadine | female | background | |||||
IN A GENERAL WAY THOUGH NOT WHOLLY NOR CONSISTENTLY THESE TWO GROUPS COINCIDE | sawyer | male | booth | |||||
HER SEA GOING QUALITIES WERE EXCELLENT AND WOULD HAVE AMPLY SUFFICED FOR A CIRCUMNAVIGATION OF THE GLOBE | jess | female | hairdryer | |||||
HE KNEW THE SILVER FLEECE HIS AND ZORA'S MUST BE RUINED | kallmekris | female | cafe | |||||
HE BEGAN A CONFUSED COMPLAINT AGAINST THE WIZARD WHO HAD VANISHED BEHIND THE CURTAIN ON THE LEFT | sawyer | male | crowdstreet | |||||
YES DEAD THESE FOUR YEARS AN A GOOD JOB FOR HER TOO | sawyer | male | inside | |||||
I CAN'T SEE IT IN THAT LIGHT SAID THE OLD LAWYER | jess | female | booth | |||||
BUT CAP'N BILL MADE NO SUCH ATTEMPT KNOWING IT WOULD BE USELESS | myra | female | street | |||||
THE PITY THAT WE MUST COME AND GO | nadine | female | restaurant | |||||
EVEN SO I HAD JUST RETURNED FROM AN ARDUOUS JOURNEY EXHAUSTED AND BADLY NEEDING A REST | kallmekris | female | cafe | |||||
ALSO A POPULAR CONTRIVANCE WHEREBY LOVE MAKING MAY BE SUSPENDED BUT NOT STOPPED DURING THE PICNIC SEASON | nadine | female | outdoor | |||||
BY QUICK MARCHES THROUGH THESE INACCESSIBLE MOUNTAINS THAT GENERAL FREED HIMSELF FROM THE SUPERIOR FORCES OF THE COVENANTERS | nadine | female | restaurant | |||||
MISS LAKE DECLINED THE CARRIAGE TO NIGHT | nadine | female | restaurant | |||||
THIS WAS WHAT DID THE MISCHIEF SO FAR AS THE RUNNING AWAY WAS CONCERNED | myra | female | street | |||||
CONCORD RETURNED TO ITS PLACE AMIDST THE TENTS | adi | male | booth | |||||
HE HAD HIS HAND UPON LAKE'S SHOULDER | nadine | female | outdoor |
Broader Impacts
As voice cloning and conversion technologies progress and grow in quality so does the risk of their misuse for malicious purposes. That is why special attention should be payed to developing mitigation strategies. One such strategy is to control the use of voice cloning systems and ensure inability to use them in open-domain speech generation without user consent. Another strategy is the development of fake audio detection systems. In our case the discriminator jointly trained with our DINO-VITS model can serve as the first step to such a solution. For its evaluation we took real speech recordings from the full CHiME3 dataset and then measured its accuracy separately on the outputs of each of the models mentioned in this paper. The F-score was above 95% for all such evalua tions. This shows that GAN discriminators seem to show a potential for efficiently detecting synthesized speech by models that share the core generator architecture with the one they were trained with (HiFi-GAN decoder in our case). However a more extensive study of this issue should be in order so as to fully ensure minimization of risk of harm to the end user related to the said technologies.