How to optimize the mispronunciation of Chinese polyphonic characters

[LSliu666]

How to optimize the mispronunciation of Chinese polyphonic characters

[SWivid]

@LSliu666 Hi, see #193

[LSliu666]

Okay, thank you。Is there any optimization in determining the pronunciation of polyphonic characters based on specific phrases @SWivid

[SWivid]

@LSliu666 Mainly for dataset. A good dataset covering most cases and transcribed just correctly will let model do better on this.

[LSliu666]

I am using Gradio. How to indicate the pinyin of the text so that I won't read it incorrectly

[LSliu666]

@SWivid

[SWivid]

Currently no such feature in Gradio yet. Need to see through the code, part as in #193. Welcome PR~

[LSliu666]

Is code inference possible？

[SWivid]

What do you mean with code inference, the infer-cli?

[LSliu666]

yes

[SWivid]

Yes, it is just as we suggest above

Need to see through the code, part as in https://github.com/SWivid/F5-TTS/issues/193.

Or simply you could do like: 好奇 -> 浩奇

[LSliu666]

f5-tts_infer-cli \ --model "F5-TTS" \ --ref_audio "ref_audio.wav" \ --ref_text "The content, subtitle or transcription of reference audio." \ --gen_text "Some text you want TTS model generate for you." Fill in Pinyin in gen_text

[LSliu666]

Yeah, I can change the homophone to his homophone.thanks

[LSliu666]

将多音字改成它的同音字

[LSliu666]

Pursuing accurate pronunciation rather than word accuracy, haha, interesting

[SWivid]

@LSliu666 Mainly for dataset. A good dataset covering most cases and transcribed just correctly will let model do better on this.

@LSliu666 Or you could just provide us with some data, we will take that to training.

Pursuing accurate pronunciation rather than word accuracy, haha, interesting

It's easy to be an armchair critic.

[LSliu666]

Sorry, I don't have any relevant data

[LSliu666]

追求发音准确而不是单词准确，哈哈，音符

The meaning of this is that I have found an alternative method for mispronouncing polyphonic characters, which is to replace the Chinese character of a polyphonic character with its homophone, so that its pronunciation is correct, pursuing consistency in pronunciation without worrying about whether the character is correct. Just like 穿着 can read as chuanzhe，If I write 穿卓 can read as chuanzhuo this is the right place to study

[SWivid]

@LSliu666 Sure, hope this temporary strategy would help. We don't mean to limit model's ability on that, rather the model is limit with dataset (if you could just provide a bunch of perfect data, we could just train with unicodes and let model learn semantic things itself)

Feel free to open issue if further questions.

[LSliu666]

Okay, thank you for your answer