Will alignment work with transliterated/romanized text?

[bonsai-byte]

Suppose I provide transliterated/romanized Japanese/Chinese and provide the language as ja/zh, will the alignment still work reliably? Or is that not recommended? This also potentially applies to languages from the Indian subcontinent where the written form of the language is often in romanized form.

[jianfch]

Although I haven't extensively tested it to confirm whether is it as reliable as normal text, it works well for the few clips I tested. I suspect the model (especially the larger ones) have learned to generalize pretty well from the large amount of training data. Given how common romanized transcriptions are, there might even be some in that data.

The text might not be properly split into words for the output if language is set to one the following: zh, ja, th, lo, my
Setting language='en' will fix it, assuming language='en' does not negatively affect the alignment accuracy.