|
--- |
|
license: mit |
|
datasets: |
|
- kaist-audio-book |
|
language: |
|
- ko |
|
pipeline_tag: text-to-speech |
|
--- |
|
|
|
# SpeechT5 Text To Speech |
|
|
|
DON'T USE THIS. This is a model that failed to train. However, someone saw my code and asked me to share this model, so I put it up. |
|
|
|
|
|
```py |
|
import torch |
|
from transformers import AutoTokenizer, SpeechT5HifiGan, SpeechT5ForTextToSpeech |
|
|
|
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan") |
|
tokenizer = AutoTokenizer.from_pretrained("Bingsu/speecht5_test") |
|
model = SpeechT5ForTextToSpeech.from_pretrained("Bingsu/speecht5_test") |
|
|
|
emb_url = "https://huggingface.co/Bingsu/speecht5_test/resolve/main/speaker_embedding.pt" |
|
emb_sd = torch.hub.load_state_dict_from_url(emb_url, map_location="cpu") |
|
emb = torch.nn.Embedding(model.config.num_speakers, model.config.speaker_embedding_dim) |
|
emb.load_state_dict(emb_sd) |
|
``` |
|
|
|
```py |
|
@torch.inference_mode() |
|
def gen(text: str, speaker_id: int = 0): |
|
inputs = tokenizer(text, return_tensors="pt") |
|
s_id = torch.tensor(speaker_id) |
|
|
|
speaker_embeddings = emb(s_id).unsqueeze(0) |
|
speech = model.generate_speech(inputs.input_ids, speaker_embeddings=speaker_embeddings, vocoder=vocoder) |
|
return speech.numpy() |
|
``` |
|
|
