When will the german version come?
When will the german version come?
This is 3 in 1 version it should work with German but may be not very good spelling numbers
I mean it kind of works but it sounds like i'm talking to my colleague Marek from work who just barely speaks german with the thickest accent you have ever heard xD
Yeah, same here ... was putting high hopes into this.
Anyways @Gregniuki could you give some details on your training?
Is it a finetune of the base model or retraining from scratch? How many hours of German/Polish/English audio have been used? What hyperparams were used for training.
Also, in case it was a single German speaker, could you upload a reference sample. Maybe this could trigger better output ... like here, where Japanese worked best with reference audio from the training dataset: https://youtu.be/GmketyZW2c4
It's fine-tune it retains some Chinese speach.
Was trained for 15 epoch with learning rate 7.5e-5 than another 15 epoch 2e-5. IIt's 100 hours English , 100 hours Polish and 50 hours German from Emily dataset and some YouTube podcasts transcribed with Whisper . Trained on 4090 with 9600 audio samples batch size . Audios were trimmed to 14 seconds . I may train it for another day with learning relate 1e-5. It's using standard char tokenizer, but maybe use of #coquitts phonemizer function like would work better and accents wouldn't mix then.
Thanks a lot. Will give it another shot, in case there's an update.
I mean it kind of works but it sounds like i'm talking to my colleague Marek from work who just barely speaks german with the thickest accent you have ever heard xD
hahaha