Skip to content

altering_pronunciation.md

Latest commit

 

History

History
87 lines (70 loc) · 4.2 KB

altering_pronunciation.md

File metadata and controls

87 lines (70 loc) · 4.2 KB
Raw

How to modify pronunciation

  • Thanks to Thorsten Mueller for this helpful tutorial which gives tips on how to add custom pronunciation rules to espeak-ng, which piper relies on.
  • If you notice that words aren't being pronounced correctly even when they are recorded correctly in your audio dataset, it is likely because the phonemes supplied by espeak-ng were incorrect when your model was trained.
  • Custom pronunciation rules must be applied before you train your model. You will need to retrain your model if you want to change its pronunciation later. Spending some time getting your pronunciation figured out before training a lot of voices is highly recommended.

Creating custom pronunciation rules

  • Clone the espeak-ng git repo to get the files you will need to modify: git clone https://github.com/espeak-ng/espeak-ng
  • For whichever language(s) you are using, you will need 3 dictionary source files from espeak-ng/dictsource (substitute the language code of the language you are modifying for xx):
  1. xx_list
  2. xx_rules
  3. xx_emoji
  4. xx_extra - this file is not supplied, it is a plain text file that you must create yourself (see below)
  • Each line in xx_extra begins with the text to be pronounced, followed by the pronunciation as represented in Kirschenbaum format (Also known as ASCII-IPA or erkIPA format).
#en_extra file format example
Noella  no'El:V O

Alternative option - use AI to create your pronunciation file - (thanks D34DC3N73R!)

D34DC3N73R reported having great success converting IPA to Kirschenbaum / ASCII IPA using an LLM. They also had very good results using an LLM to generate custom pronunciations (ie en_extra file entries) for specific accents. I haven't tried it yet but this looks like a very promising option.

Example prompt:

I am developing a New York accent voice model for Piper, 
which utilizes eSpeak-NG as its phoneme synthesis engine. 
Provide a list of the 100 words most likely to be mispronounced 
by eSpeak-NG due to the distinct features of the New York accent. 
Display each word in a list with its Kirschenbaum (ASCII-IPA) 
conversion format, as shown in these examples:

garage g@'rA:Z 
smart smA:rt 
sabotage sab@,tA:3 

Do not include any explanations or additional text within the list itself.

Example output:

Here is a list of the top 100 words most likely to be mispronounced by eSpeak-NG in a New York accent, 
presented in Kirschenbaum (ASCII-IPA) format without explanations:

coffee k'A:fi
water w'A:t@
dog dA:g
talk tA:k
call kA:l
car kA:
park pA:k
sure S@
door dO:
floor flO:
long lA:Ng
island 'A:l@nd
thought TA:t

etc...

Activating custom pronunciation rules

For people using the textymcspeechy-piper docker image (applies to most users)

  • Espeak-ng is preinstalled in the textymcspeechy-piper docker image. The commands to compile pronunciation must be run as root within the docker container, and must be run every time the container starts.
  • There are scripts for both manually and automatically activating your custom pronunciations provided. Instructions for these are found here.

For non-containerized installs of espeak-ng (more convenient for testing):

  • Compile the new pronunciation rule to make it active.
  # IMPORTANT-  run this from the directory where xx_list, xx_rules, xx_emoji, and xx_extra are located.

  sudo espeak-ng --compile=xx  # <--- where xx is the 2 letter language code.
  • Test the new pronunciation, eg
espeak-ng "Testing the corrected pronunciation of Noella" --ipa

espeak-ng -v ja "ピクルス" --ipa

espeak-ng -v ru "Пожалуйста" --ipa
  • Changes to espeak-ng's pronunciation rules are applied system-wide on non-containerized installs and only have to be applied once. If you need to undo your custom pronunciations, revise or delete xx_extra and compile your rules again as above.