|
from bark.generation import load_codec_model, generate_text_semantic, grab_best_device |
|
from encodec.utils import convert_audio |
|
import torchaudio |
|
import torch |
|
import os |
|
import gradio |
|
|
|
|
|
def clone_voice(audio_filepath, text, dest_filename, progress=gradio.Progress(track_tqdm=True)): |
|
if len(text) < 1: |
|
raise gradio.Error('No transcription text entered!') |
|
|
|
use_gpu = not os.environ.get("BARK_FORCE_CPU", False) |
|
progress(0, desc="Loading Codec") |
|
model = load_codec_model(use_gpu=use_gpu) |
|
progress(0.25, desc="Converting WAV") |
|
|
|
|
|
device = grab_best_device(use_gpu) |
|
wav, sr = torchaudio.load(audio_filepath) |
|
wav = convert_audio(wav, sr, model.sample_rate, model.channels) |
|
wav = wav.unsqueeze(0).to(device) |
|
progress(0.5, desc="Extracting codes") |
|
|
|
|
|
with torch.no_grad(): |
|
encoded_frames = model.encode(wav) |
|
codes = torch.cat([encoded[0] for encoded in encoded_frames], dim=-1).squeeze() |
|
|
|
|
|
seconds = wav.shape[-1] / model.sample_rate |
|
|
|
semantic_tokens = generate_text_semantic(text, max_gen_duration_s=seconds, top_k=50, top_p=.95, temp=0.7) |
|
|
|
|
|
codes = codes.cpu().numpy() |
|
|
|
import numpy as np |
|
output_path = dest_filename + '.npz' |
|
np.savez(output_path, fine_prompt=codes, coarse_prompt=codes[:2, :], semantic_prompt=semantic_tokens) |
|
return "Finished" |
|
|