In the past our model would fail to generate some sentences. This was caused by very short or long sentences, as these sentences are generally not seen during training.
In this update, we have significantly improved the stability and you will no longer see failure-to-generate errors (hopefully!).
Without getting into the deep technical details, the model has several fall backs methods to generate audio no matter what you type.
If the model needs to use one of the fall back methods, there may be a slight change in accent. The audio quality might be less than normal - but something will generate and no words will be skipped.
This was one of the biggest requests from our early access users.
So thanks for all your feedback and help as we build out the Replica tools. Up next we've been working on the voice creation process.
Thanks,
Replica Team
