Request: Alternative Pretrained models for UK accent

[munceyboyjoe]

Whilst this tool is fascinating and fun to play with, the pretrained models made available for download impart a US accent on UK English speakers. I don't have the compute power to train models, does anyone know of available models for download that would fair better with UK accents?

[deeepgandhi]

Yeah, I face a similar problem. I think if you could find datasets of speakers in the UK accent. You can finetune the speaker encoding model. I think that can make the results better.

[ghost]

As @deepgandhideep mentioned, you can finetune a single-speaker model and I have provided directions here. The training can be completed in less than a day without a GPU. #437 (comment)

[ghost]

I tried training on VCTK, but there is not enough data to make a good voice cloning model. Many of the speakers in that dataset have unsuitable accents (American, Canadian, Australian, Indian). Furthermore, the vocabulary in that corpus lacks the variety that we have in LibriSpeech/TTS which makes for a better synthesizer.

If anyone wants this, please provide the dataset and transcripts in a usable format (preferably the one in #437 (comment) ). Commonvoice is a potential dataset as you can use the metadata to filter by accent. You could also explore the speakers in the Spoken Wikipedia Corpora (https://nats.gitlab.io/swc/).

[ghost]

Closing this issue due to inactivity. Anyone is welcome to reopen it if taking this project on, or if a dataset is located that isn't VCTK. It is possible to train a voice cloning model on VCTK, but the result doesn't meet my standards.