Speaking issues Coqui TTS Tacotron2 DDC model: Difference between revisions
Jump to navigation
Jump to search
m (Just an edit for demo) Tags: Reverted Visual edit |
m (Rollback changes) Tags: Manual revert Visual edit |
||
| Line 17: | Line 17: | ||
* Nelson Mandela | * Nelson Mandela | ||
* pergola | * pergola | ||
==== Mitigation ==== | ==== Mitigation ==== | ||
Latest revision as of 23:44, 12 November 2021
General[edit | edit source]
Most models are trained with a dot, exclamation or question mark at the end. So always end a sentence to avoid model synthesizing weird output.
- TTS model: tts_models/en/ljspeech/tacotron2-DDC
- Vocoder model: vocoder_models/en/ljspeech/hifigan_v2
Input string formatting[edit | edit source]
Phrases ending in "ah"[edit | edit source]
"ah" at end of sentence generally produces strange results. Short names produce a 12 second clip.
Examples:
- Nelson Mandela
- pergola
Mitigation[edit | edit source]
If at the end of the input, adding punctuation to the end synthesizes correctly:
Example "Nelson Mandela" > "Nelson Mandela."
Acronyms[edit | edit source]
To speak acronyms as letters it needs to be formatted as:
"A. B. C. news"
Not:
"ABC news" "A.B.C. news"
Mispronounced Words[edit | edit source]
- video