- These are some tips to generate the best sounding text-to-speech. + Here are some quick tips to generate the best sounding text-to-speech.
- A TTS stroke is when the generated audio becomes garbled and - incoherent. + (This is not a typical problem.)
+- This can be fixed by adding an ending punctuation as without - one, this will happen frequently. If the message still messes - up, add a comma at the start and end so the program can take a - brief pause and usually corrects it. + Some of our very earliest models were built using a limited amount of data + and are sometimes prone to producing speech that sounds like “babbling” + or incoherent vocal convulsions. +
++ This problem can typically be fixed by + {" "} + adding ending punctuation to your sentence. + If the audio still has problems, + {" "} + add a comma at the start and end of your text + {" "} + so that the program can add brief pauses. This usually corrects the problem. +
++ If the incoherent audio is at the end, you can always edit it with software + such as Audacity.
- Sometimes the TTS program cannot accurately read the word or you - want a word pronounced a certain way. You can experiment with - it. For example, to say the word "ass" with a certain hit, spell - it as "asz". Or another example, Games = Gaymes. + Sometimes a TTS model will not pronounce words correctly or in the + way you like. + You can experiment with misspellings of words until you get something + that sounds better. For example “peekahchuu” might sound different + than “pikachu”. +
+
+ Many of our models also support Arpabet phonetic annotation. It's an advanced feature you can
+ learn to use by joining our
To get a certain emotion and personality in the message you - want, add fluff. To do this, add the first sentence with words - that you have no intention of using. Like bunch of cursing and - such. " - - Fuck you bitch! I hate you! - {" "} - I cant believe you said that to me!". You can remove the first - sentence you had no intention of using in an audio editor of - your choice like Audacity. + want, add some extra emotional “conditioning”. +
++ To do this, think of a strong emotional sentence or two you could start with. + Then add this to the very beginning of your text. + For example, to add some anger to “I can't believe you said that to me!”, + you could input the following instead: +
++ “Darn it! I hate you! I can't believe you said that to me!” +
++ You can later remove these extra leading sentences with an audio editor such as Audacity, leaving + just the speech you want.
- This only works if the dataset has yelling and transcript - properly. When you do the transcript of a character that yells, - add an exclamation mark on the transcript clip that has yelling - and periods for normal speech. This will easier trigger angry - emotions. But the yelling will likely not happen with smaller - datasets or if the yelling is not at all in the dataset. + While not all models support it, some of our models incorporate rich emotional samples + that were transcripted with an emphasis mark – typically an exclamation mark (!).