little thing that converts english text into animal crossing voice sounds.

here's what it does, with an example roughly approximating the procedure:

1. start with user input
   bunny world
2. use DeepPhonemizer to turn user input into IPA
   bʌni wɜːld
3. parse the IPA into animal crossing voicebank samples (which consist of arabic numerals, latin letters, and japanese moras)
   [[b a n i] [w a r d]]
[[ba ni] [wa ru do]]
4. stitch together the audio samples and play the resulting sound!
   🎵

sorry about the code

get this thing working

look at requirements.txt and install those things. pytorch cannot be 2.6 because DeepPhonemizer doesn't pass weights_only=False to torch.load. i could probably fix this by forking DeepPhonemizer or copying the implementation of load_checkpoint into my code, but i don't really want to.

if you don't already have pytorch installed, may i suggest specifically pulling the cpu-only package from their site like this:
pip install torch~=2.5.0 --index-url https://download.pytorch.org/whl/cpu
so that you don't have to download one quadrillion nvidia packages you don't need. otherwise pip install -r requirements.txt is fine for the rest of the packages

you need to prepare a voicebank at a location specified in config.py (the default is voices/GAL). it should be a collection of wav files named [0-9].wav, [A-Z].wav, [kgsztdnhbpmyrw]?[aiueo].wav.

here's a link to an audio rip from New Leaf by someone called JJ314. if you download that and extract it to a subdirectory called voices_untitled in this directory, you can run the provided voices-rename.py to receive three voicebanks BOY, GAL, and MAN.

now, you need a model for DeepPhonemizer. this has been tested exclusively with their provided pretrained model latin_ipa_forward.pt in en_uk (find it in their README). the code expects that file with that name in the current directory, but you can change it in config.py.

if you want to use another language, or a model trained on american english, you'll surely have to make some serious alterations to my translation layer. the translation rules are tuned to what the aforementioned model emits and nothing more.

using it

there's a little sample CLI included if you run acvoice.py directly. use --help for more information. you need ffplay installed to get the repl to produce audio.

if you want to use it programmatically, you can import it and

1. phonemizer to turn english text into IPA
2. parse_phonemes to turn IPA into a voicebank wordlist
3. stitch_audio to turn a voicebank wordlist into a pydub AudioSegment
4. do whatever you want after that

i have deliberately restricted my use of the config.py to the sample CLI. if you import as a library, you don't need it and the API will not use its values.