2cent-tts-60m

Converted the models from https://github.com/taylorchu/2cent-tts to .safetensors. Below is inference code:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained(
    "timBoML/2cent-tts-60m"
)
tokenizer = AutoTokenizer.from_pretrained("timBoML/2cent-tts-60m")

phones = "həlˈoʊ aɪɐm tˈuː sˈɛnt tˌiːtˌiːˈɛs" # using espeak-ng

input_ids = (
    tokenizer.encode(phones, add_special_tokens=False)
    + tokenizer.encode("<s>", add_special_tokens=False)
    + [4136]
)

input_ids = torch.tensor(input_ids).unsqueeze(0)

generated_ids = model.generate(
      input_ids=input_ids,
      max_new_tokens=2048,

  )

generated_ids = generated_ids.squeeze()

tokens = generated_ids[input_ids.shape[1]:]

first_audio_token = tokenizer.encode("<audio_0>")[-1]

tokens = tokens - first_audio_token


import locale
import torchaudio.transforms as T
import os
import torch
from snac import SNAC
locale.getpreferredencoding = lambda: "UTF-8"

snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")

def redistribute_codes(code_list):
  layer_1 = []
  layer_2 = []
  layer_3 = []
  for i in range((len(code_list)+1)//7):
    layer_1.append(code_list[7*i])
    layer_2.append(code_list[7*i+1])
    layer_3.append(code_list[7*i+2])
    layer_3.append(code_list[7*i+3])
    layer_2.append(code_list[7*i+4])
    layer_3.append(code_list[7*i+5])
    layer_3.append(code_list[7*i+6])
  codes = [torch.tensor(layer_1).unsqueeze(0),
         torch.tensor(layer_2).unsqueeze(0),
         torch.tensor(layer_3).unsqueeze(0)]
  audio_hat = snac_model.decode(codes)
  return audio_hat

sample = redistribute_codes(tokens)

from IPython.display import Audio, display
display(Audio(sample.detach().squeeze().to("cpu").numpy(), rate=24000))