import os import sys import importlib.util import site import json import torch import gradio as gr import torchaudio import numpy as np from huggingface_hub import snapshot_download, hf_hub_download import subprocess import re import spaces # 创建一个全局变量来跟踪已下载的资源 # Create a global variable to track downloaded resources downloaded_resources = { "configs": False, "tokenizer_vq32": False, "tokenizer_vq8192": False, "ar_Vq32ToVq8192": False, "ar_PhoneToVq8192": False, "fmt_Vq8192ToMels": False, "vocoder": False } def install_espeak(): """Detect and install espeak-ng dependency""" try: # Check if espeak-ng is already installed result = subprocess.run(["which", "espeak-ng"], capture_output=True, text=True) if result.returncode != 0: print("Detected espeak-ng not installed in the system, attempting to install...") # Try to install espeak-ng and its data using apt-get subprocess.run(["apt-get", "update"], check=True) # Install espeak-ng and the corresponding language data package subprocess.run(["apt-get", "install", "-y", "espeak-ng", "espeak-ng-data"], check=True) print("espeak-ng and its data packages installed successfully!") else: print("espeak-ng is already installed in the system.") # Even if already installed, try to update data to ensure integrity (optional but sometimes helpful) # print("Attempting to update espeak-ng data...") # subprocess.run(["apt-get", "update"], check=True) # subprocess.run(["apt-get", "install", "--only-upgrade", "-y", "espeak-ng-data"], check=True) # Verify Chinese support (optional) try: voices_result = subprocess.run(["espeak-ng", "--voices=cmn"], capture_output=True, text=True, check=True) if "cmn" in voices_result.stdout: print("espeak-ng supports 'cmn' language.") else: print("Warning: espeak-ng is installed, but 'cmn' language still seems unavailable.") except Exception as e: print(f"Error verifying espeak-ng Chinese support (may not affect functionality): {e}") except Exception as e: print(f"Error installing espeak-ng: {e}") print("Please try to run manually: apt-get update && apt-get install -y espeak-ng espeak-ng-data") # Install espeak before all other operations install_espeak() def patch_langsegment_init(): try: # Try to find the location of the LangSegment package spec = importlib.util.find_spec("LangSegment") if spec is None or spec.origin is None: print("Unable to locate LangSegment package.") return # Build the path to __init__.py init_path = os.path.join(os.path.dirname(spec.origin), '__init__.py') if not os.path.exists(init_path): print(f"LangSegment __init__.py file not found at: {init_path}") # Try to find in site-packages, applicable in some environments for site_pkg_path in site.getsitepackages(): potential_path = os.path.join(site_pkg_path, 'LangSegment', '__init__.py') if os.path.exists(potential_path): init_path = potential_path print(f"Found __init__.py in site-packages: {init_path}") break else: # If the loop ends normally (no break) print(f"Also unable to find __init__.py in site-packages") return print(f"Attempting to read LangSegment __init__.py: {init_path}") with open(init_path, 'r') as f: lines = f.readlines() modified = False new_lines = [] target_line_prefix = "from .LangSegment import" for line in lines: stripped_line = line.strip() if stripped_line.startswith(target_line_prefix): if 'setLangfilters' in stripped_line or 'getLangfilters' in stripped_line: print(f"Found line that needs modification: {stripped_line}") # Remove setLangfilters and getLangfilters modified_line = stripped_line.replace(',setLangfilters', '') modified_line = modified_line.replace(',getLangfilters', '') # Ensure comma handling is correct (e.g., if they are the last items) modified_line = modified_line.replace('setLangfilters,', '') modified_line = modified_line.replace('getLangfilters,', '') # If they are the only extra imports, remove any redundant commas modified_line = modified_line.rstrip(',') new_lines.append(modified_line + '\n') modified = True print(f"Modified line: {modified_line.strip()}") else: new_lines.append(line) # Line is fine, keep as is else: new_lines.append(line) # Non-target line, keep as is if modified: print(f"Attempting to write back modified LangSegment __init__.py to: {init_path}") try: with open(init_path, 'w') as f: f.writelines(new_lines) print("LangSegment __init__.py modified successfully.") # Try to reload the module to make changes effective (may not work, depending on import chain) try: import LangSegment importlib.reload(LangSegment) print("LangSegment module has been attempted to reload.") except Exception as reload_e: print(f"Error reloading LangSegment (may have no impact): {reload_e}") except PermissionError: print(f"Error: Insufficient permissions to modify {init_path}. Consider modifying requirements.txt.") except Exception as write_e: print(f"Other error occurred when writing LangSegment __init__.py: {write_e}") else: print("LangSegment __init__.py doesn't need modification.") except ImportError: print("LangSegment package not found, unable to fix.") except Exception as e: print(f"Unexpected error occurred when fixing LangSegment package: {e}") # Execute the fix before all other imports (especially Amphion) that might trigger LangSegment patch_langsegment_init() # Clone Amphion repository if not os.path.exists("Amphion"): subprocess.run(["git", "clone", "https://github.com/open-mmlab/Amphion.git"]) os.chdir("Amphion") else: if not os.getcwd().endswith("Amphion"): os.chdir("Amphion") # Add Amphion to the path if os.path.dirname(os.path.abspath("Amphion")) not in sys.path: sys.path.append(os.path.dirname(os.path.abspath("Amphion"))) # Ensure needed directories exist os.makedirs("wav", exist_ok=True) os.makedirs("ckpts/Vevo", exist_ok=True) from models.vc.vevo.vevo_utils import VevoInferencePipeline, save_audio, load_wav # Download and setup config files def setup_configs(): if downloaded_resources["configs"]: print("Config files already downloaded, skipping...") return config_path = "models/vc/vevo/config" os.makedirs(config_path, exist_ok=True) config_files = [ "PhoneToVq8192.json", "Vocoder.json", "Vq32ToVq8192.json", "Vq8192ToMels.json", "hubert_large_l18_c32.yaml", ] for file in config_files: file_path = f"{config_path}/{file}" if not os.path.exists(file_path): try: file_data = hf_hub_download( repo_id="amphion/Vevo", filename=f"config/{file}", repo_type="model", ) os.makedirs(os.path.dirname(file_path), exist_ok=True) # Copy file to target location subprocess.run(["cp", file_data, file_path]) except Exception as e: print(f"Error downloading config file {file}: {e}") downloaded_resources["configs"] = True setup_configs() # Device configuration device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") print(f"Using device: {device}") # Initialize pipeline dictionary inference_pipelines = {} # Download all necessary model resources at startup def preload_all_resources(): print("Preloading all model resources...") # Download configuration files setup_configs() # Store the downloaded model paths global downloaded_content_tokenizer_path global downloaded_content_style_tokenizer_path global downloaded_ar_vq32_path global downloaded_ar_phone_path global downloaded_fmt_path global downloaded_vocoder_path # Download Content Tokenizer (vq32) if not downloaded_resources["tokenizer_vq32"]: print("Preloading Content Tokenizer (vq32)...") local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["tokenizer/vq32/*"], ) downloaded_content_tokenizer_path = local_dir downloaded_resources["tokenizer_vq32"] = True print("Content Tokenizer (vq32) download completed") # Download Content-Style Tokenizer (vq8192) if not downloaded_resources["tokenizer_vq8192"]: print("Preloading Content-Style Tokenizer (vq8192)...") local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["tokenizer/vq8192/*"], ) downloaded_content_style_tokenizer_path = local_dir downloaded_resources["tokenizer_vq8192"] = True print("Content-Style Tokenizer (vq8192) download completed") # Download Autoregressive Transformer (Vq32ToVq8192) if not downloaded_resources["ar_Vq32ToVq8192"]: print("Preloading Autoregressive Transformer (Vq32ToVq8192)...") local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["contentstyle_modeling/Vq32ToVq8192/*"], ) downloaded_ar_vq32_path = local_dir downloaded_resources["ar_Vq32ToVq8192"] = True print("Autoregressive Transformer (Vq32ToVq8192) download completed") # Download Autoregressive Transformer (PhoneToVq8192) if not downloaded_resources["ar_PhoneToVq8192"]: print("Preloading Autoregressive Transformer (PhoneToVq8192)...") local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["contentstyle_modeling/PhoneToVq8192/*"], ) downloaded_ar_phone_path = local_dir downloaded_resources["ar_PhoneToVq8192"] = True print("Autoregressive Transformer (PhoneToVq8192) download completed") # Download Flow Matching Transformer if not downloaded_resources["fmt_Vq8192ToMels"]: print("Preloading Flow Matching Transformer (Vq8192ToMels)...") local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vq8192ToMels/*"], ) downloaded_fmt_path = local_dir downloaded_resources["fmt_Vq8192ToMels"] = True print("Flow Matching Transformer (Vq8192ToMels) download completed") # Download Vocoder if not downloaded_resources["vocoder"]: print("Preloading Vocoder...") local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vocoder/*"], ) downloaded_vocoder_path = local_dir downloaded_resources["vocoder"] = True print("Vocoder download completed") print("All model resources preloading completed!") # Initialize path variables to store downloaded model paths downloaded_content_tokenizer_path = None downloaded_content_style_tokenizer_path = None downloaded_ar_vq32_path = None downloaded_ar_phone_path = None downloaded_fmt_path = None downloaded_vocoder_path = None # Preload all resources before creating the Gradio interface preload_all_resources() def get_pipeline(pipeline_type): if pipeline_type in inference_pipelines: return inference_pipelines[pipeline_type] # Initialize pipeline based on the required pipeline type if pipeline_type == "style" or pipeline_type == "voice": # Use already downloaded Content Tokenizer if downloaded_resources["tokenizer_vq32"]: content_tokenizer_ckpt_path = os.path.join( downloaded_content_tokenizer_path, "tokenizer/vq32/hubert_large_l18_c32.pkl" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["tokenizer/vq32/*"], ) content_tokenizer_ckpt_path = os.path.join( local_dir, "tokenizer/vq32/hubert_large_l18_c32.pkl" ) # Use already downloaded Content-Style Tokenizer if downloaded_resources["tokenizer_vq8192"]: content_style_tokenizer_ckpt_path = os.path.join( downloaded_content_style_tokenizer_path, "tokenizer/vq8192" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["tokenizer/vq8192/*"], ) content_style_tokenizer_ckpt_path = os.path.join(local_dir, "tokenizer/vq8192") # Use already downloaded Autoregressive Transformer ar_cfg_path = "./models/vc/vevo/config/Vq32ToVq8192.json" if downloaded_resources["ar_Vq32ToVq8192"]: ar_ckpt_path = os.path.join( downloaded_ar_vq32_path, "contentstyle_modeling/Vq32ToVq8192" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["contentstyle_modeling/Vq32ToVq8192/*"], ) ar_ckpt_path = os.path.join(local_dir, "contentstyle_modeling/Vq32ToVq8192") # Use already downloaded Flow Matching Transformer fmt_cfg_path = "./models/vc/vevo/config/Vq8192ToMels.json" if downloaded_resources["fmt_Vq8192ToMels"]: fmt_ckpt_path = os.path.join( downloaded_fmt_path, "acoustic_modeling/Vq8192ToMels" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vq8192ToMels/*"], ) fmt_ckpt_path = os.path.join(local_dir, "acoustic_modeling/Vq8192ToMels") # Use already downloaded Vocoder vocoder_cfg_path = "./models/vc/vevo/config/Vocoder.json" if downloaded_resources["vocoder"]: vocoder_ckpt_path = os.path.join( downloaded_vocoder_path, "acoustic_modeling/Vocoder" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vocoder/*"], ) vocoder_ckpt_path = os.path.join(local_dir, "acoustic_modeling/Vocoder") # Initialize pipeline inference_pipeline = VevoInferencePipeline( content_tokenizer_ckpt_path=content_tokenizer_ckpt_path, content_style_tokenizer_ckpt_path=content_style_tokenizer_ckpt_path, ar_cfg_path=ar_cfg_path, ar_ckpt_path=ar_ckpt_path, fmt_cfg_path=fmt_cfg_path, fmt_ckpt_path=fmt_ckpt_path, vocoder_cfg_path=vocoder_cfg_path, vocoder_ckpt_path=vocoder_ckpt_path, device=device, ) elif pipeline_type == "timbre": # Use already downloaded Content-Style Tokenizer if downloaded_resources["tokenizer_vq8192"]: content_style_tokenizer_ckpt_path = os.path.join( downloaded_content_style_tokenizer_path, "tokenizer/vq8192" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["tokenizer/vq8192/*"], ) content_style_tokenizer_ckpt_path = os.path.join(local_dir, "tokenizer/vq8192") # Use already downloaded Flow Matching Transformer fmt_cfg_path = "./models/vc/vevo/config/Vq8192ToMels.json" if downloaded_resources["fmt_Vq8192ToMels"]: fmt_ckpt_path = os.path.join( downloaded_fmt_path, "acoustic_modeling/Vq8192ToMels" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vq8192ToMels/*"], ) fmt_ckpt_path = os.path.join(local_dir, "acoustic_modeling/Vq8192ToMels") # Use already downloaded Vocoder vocoder_cfg_path = "./models/vc/vevo/config/Vocoder.json" if downloaded_resources["vocoder"]: vocoder_ckpt_path = os.path.join( downloaded_vocoder_path, "acoustic_modeling/Vocoder" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vocoder/*"], ) vocoder_ckpt_path = os.path.join(local_dir, "acoustic_modeling/Vocoder") # Initialize pipeline inference_pipeline = VevoInferencePipeline( content_style_tokenizer_ckpt_path=content_style_tokenizer_ckpt_path, fmt_cfg_path=fmt_cfg_path, fmt_ckpt_path=fmt_ckpt_path, vocoder_cfg_path=vocoder_cfg_path, vocoder_ckpt_path=vocoder_ckpt_path, device=device, ) elif pipeline_type == "tts": # Use already downloaded Content-Style Tokenizer if downloaded_resources["tokenizer_vq8192"]: content_style_tokenizer_ckpt_path = os.path.join( downloaded_content_style_tokenizer_path, "tokenizer/vq8192" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["tokenizer/vq8192/*"], ) content_style_tokenizer_ckpt_path = os.path.join(local_dir, "tokenizer/vq8192") # Use already downloaded Autoregressive Transformer (TTS specific) ar_cfg_path = "./models/vc/vevo/config/PhoneToVq8192.json" if downloaded_resources["ar_PhoneToVq8192"]: ar_ckpt_path = os.path.join( downloaded_ar_phone_path, "contentstyle_modeling/PhoneToVq8192" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["contentstyle_modeling/PhoneToVq8192/*"], ) ar_ckpt_path = os.path.join(local_dir, "contentstyle_modeling/PhoneToVq8192") # Use already downloaded Flow Matching Transformer fmt_cfg_path = "./models/vc/vevo/config/Vq8192ToMels.json" if downloaded_resources["fmt_Vq8192ToMels"]: fmt_ckpt_path = os.path.join( downloaded_fmt_path, "acoustic_modeling/Vq8192ToMels" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vq8192ToMels/*"], ) fmt_ckpt_path = os.path.join(local_dir, "acoustic_modeling/Vq8192ToMels") # Use already downloaded Vocoder vocoder_cfg_path = "./models/vc/vevo/config/Vocoder.json" if downloaded_resources["vocoder"]: vocoder_ckpt_path = os.path.join( downloaded_vocoder_path, "acoustic_modeling/Vocoder" ) else: # Fallback to direct download local_dir = snapshot_download( repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vocoder/*"], ) vocoder_ckpt_path = os.path.join(local_dir, "acoustic_modeling/Vocoder") # Initialize pipeline inference_pipeline = VevoInferencePipeline( content_style_tokenizer_ckpt_path=content_style_tokenizer_ckpt_path, ar_cfg_path=ar_cfg_path, ar_ckpt_path=ar_ckpt_path, fmt_cfg_path=fmt_cfg_path, fmt_ckpt_path=fmt_ckpt_path, vocoder_cfg_path=vocoder_cfg_path, vocoder_ckpt_path=vocoder_ckpt_path, device=device, ) # Cache pipeline instance inference_pipelines[pipeline_type] = inference_pipeline return inference_pipeline # Implement VEVO functionality functions @spaces.GPU() def vevo_style(content_wav, style_wav): temp_content_path = "wav/temp_content.wav" temp_style_path = "wav/temp_style.wav" output_path = "wav/output_vevostyle.wav" # Check and process audio data if content_wav is None or style_wav is None: raise ValueError("Please upload audio files") # Process audio format if isinstance(content_wav, tuple) and len(content_wav) == 2: if isinstance(content_wav[0], np.ndarray): content_data, content_sr = content_wav else: content_sr, content_data = content_wav # Ensure single channel if len(content_data.shape) > 1 and content_data.shape[1] > 1: content_data = np.mean(content_data, axis=1) # Resample to 24kHz if content_sr != 24000: content_tensor = torch.FloatTensor(content_data).unsqueeze(0) content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000) content_sr = 24000 else: content_tensor = torch.FloatTensor(content_data).unsqueeze(0) # Normalize volume content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95 else: raise ValueError("Invalid content audio format") if isinstance(style_wav[0], np.ndarray): style_data, style_sr = style_wav else: style_sr, style_data = style_wav # Ensure single channel if len(style_data.shape) > 1 and style_data.shape[1] > 1: style_data = np.mean(style_data, axis=1) # Resample to 24kHz if style_sr != 24000: style_tensor = torch.FloatTensor(style_data).unsqueeze(0) style_tensor = torchaudio.functional.resample(style_tensor, style_sr, 24000) style_sr = 24000 else: style_tensor = torch.FloatTensor(style_data).unsqueeze(0) # Normalize volume style_tensor = style_tensor / (torch.max(torch.abs(style_tensor)) + 1e-6) * 0.95 # Print debug information print(f"Content audio shape: {content_tensor.shape}, sample rate: {content_sr}") print(f"Style audio shape: {style_tensor.shape}, sample rate: {style_sr}") # Save audio torchaudio.save(temp_content_path, content_tensor, content_sr) torchaudio.save(temp_style_path, style_tensor, style_sr) try: # Get pipeline pipeline = get_pipeline("style") # Inference gen_audio = pipeline.inference_ar_and_fm( src_wav_path=temp_content_path, src_text=None, style_ref_wav_path=temp_style_path, timbre_ref_wav_path=temp_content_path, ) # Check if generated audio is numerical anomaly if torch.isnan(gen_audio).any() or torch.isinf(gen_audio).any(): print("Warning: Generated audio contains NaN or Inf values") gen_audio = torch.nan_to_num(gen_audio, nan=0.0, posinf=0.95, neginf=-0.95) print(f"Generated audio shape: {gen_audio.shape}, max: {torch.max(gen_audio)}, min: {torch.min(gen_audio)}") # Save generated audio save_audio(gen_audio, output_path=output_path) return output_path except Exception as e: print(f"Error during processing: {e}") import traceback traceback.print_exc() raise e @spaces.GPU() def vevo_timbre(content_wav, reference_wav): temp_content_path = "wav/temp_content.wav" temp_reference_path = "wav/temp_reference.wav" output_path = "wav/output_vevotimbre.wav" # Check and process audio data if content_wav is None or reference_wav is None: raise ValueError("Please upload audio files") # Process content audio format if isinstance(content_wav, tuple) and len(content_wav) == 2: if isinstance(content_wav[0], np.ndarray): content_data, content_sr = content_wav else: content_sr, content_data = content_wav # Ensure single channel if len(content_data.shape) > 1 and content_data.shape[1] > 1: content_data = np.mean(content_data, axis=1) # Resample to 24kHz if content_sr != 24000: content_tensor = torch.FloatTensor(content_data).unsqueeze(0) content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000) content_sr = 24000 else: content_tensor = torch.FloatTensor(content_data).unsqueeze(0) # Normalize volume content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95 else: raise ValueError("Invalid content audio format") # Process reference audio format if isinstance(reference_wav, tuple) and len(reference_wav) == 2: if isinstance(reference_wav[0], np.ndarray): reference_data, reference_sr = reference_wav else: reference_sr, reference_data = reference_wav # Ensure single channel if len(reference_data.shape) > 1 and reference_data.shape[1] > 1: reference_data = np.mean(reference_data, axis=1) # Resample to 24kHz if reference_sr != 24000: reference_tensor = torch.FloatTensor(reference_data).unsqueeze(0) reference_tensor = torchaudio.functional.resample(reference_tensor, reference_sr, 24000) reference_sr = 24000 else: reference_tensor = torch.FloatTensor(reference_data).unsqueeze(0) # Normalize volume reference_tensor = reference_tensor / (torch.max(torch.abs(reference_tensor)) + 1e-6) * 0.95 else: raise ValueError("Invalid reference audio format") # Print debug information print(f"Content audio shape: {content_tensor.shape}, sample rate: {content_sr}") print(f"Reference audio shape: {reference_tensor.shape}, sample rate: {reference_sr}") # Save uploaded audio torchaudio.save(temp_content_path, content_tensor, content_sr) torchaudio.save(temp_reference_path, reference_tensor, reference_sr) try: # Get pipeline pipeline = get_pipeline("timbre") # Inference gen_audio = pipeline.inference_fm( src_wav_path=temp_content_path, timbre_ref_wav_path=temp_reference_path, flow_matching_steps=32, ) # Check if generated audio is numerical anomaly if torch.isnan(gen_audio).any() or torch.isinf(gen_audio).any(): print("Warning: Generated audio contains NaN or Inf values") gen_audio = torch.nan_to_num(gen_audio, nan=0.0, posinf=0.95, neginf=-0.95) print(f"Generated audio shape: {gen_audio.shape}, max: {torch.max(gen_audio)}, min: {torch.min(gen_audio)}") # Save generated audio save_audio(gen_audio, output_path=output_path) return output_path except Exception as e: print(f"Error during processing: {e}") import traceback traceback.print_exc() raise e @spaces.GPU() def vevo_voice(content_wav, style_reference_wav, timbre_reference_wav): temp_content_path = "wav/temp_content.wav" temp_style_path = "wav/temp_style.wav" temp_timbre_path = "wav/temp_timbre.wav" output_path = "wav/output_vevovoice.wav" # Check and process audio data if content_wav is None or style_reference_wav is None or timbre_reference_wav is None: raise ValueError("Please upload all required audio files") # Process content audio format if isinstance(content_wav, tuple) and len(content_wav) == 2: if isinstance(content_wav[0], np.ndarray): content_data, content_sr = content_wav else: content_sr, content_data = content_wav # Ensure single channel if len(content_data.shape) > 1 and content_data.shape[1] > 1: content_data = np.mean(content_data, axis=1) # Resample to 24kHz if content_sr != 24000: content_tensor = torch.FloatTensor(content_data).unsqueeze(0) content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000) content_sr = 24000 else: content_tensor = torch.FloatTensor(content_data).unsqueeze(0) # Normalize volume content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95 else: raise ValueError("Invalid content audio format") # Process style reference audio format if isinstance(style_reference_wav, tuple) and len(style_reference_wav) == 2: if isinstance(style_reference_wav[0], np.ndarray): style_data, style_sr = style_reference_wav else: style_sr, style_data = style_reference_wav # Ensure single channel if len(style_data.shape) > 1 and style_data.shape[1] > 1: style_data = np.mean(style_data, axis=1) # Resample to 24kHz if style_sr != 24000: style_tensor = torch.FloatTensor(style_data).unsqueeze(0) style_tensor = torchaudio.functional.resample(style_tensor, style_sr, 24000) style_sr = 24000 else: style_tensor = torch.FloatTensor(style_data).unsqueeze(0) # Normalize volume style_tensor = style_tensor / (torch.max(torch.abs(style_tensor)) + 1e-6) * 0.95 else: raise ValueError("Invalid style reference audio format") # Process timbre reference audio format if isinstance(timbre_reference_wav, tuple) and len(timbre_reference_wav) == 2: if isinstance(timbre_reference_wav[0], np.ndarray): timbre_data, timbre_sr = timbre_reference_wav else: timbre_sr, timbre_data = timbre_reference_wav # Ensure single channel if len(timbre_data.shape) > 1 and timbre_data.shape[1] > 1: timbre_data = np.mean(timbre_data, axis=1) # Resample to 24kHz if timbre_sr != 24000: timbre_tensor = torch.FloatTensor(timbre_data).unsqueeze(0) timbre_tensor = torchaudio.functional.resample(timbre_tensor, timbre_sr, 24000) timbre_sr = 24000 else: timbre_tensor = torch.FloatTensor(timbre_data).unsqueeze(0) # Normalize volume timbre_tensor = timbre_tensor / (torch.max(torch.abs(timbre_tensor)) + 1e-6) * 0.95 else: raise ValueError("Invalid timbre reference audio format") # Print debug information print(f"Content audio shape: {content_tensor.shape}, sample rate: {content_sr}") print(f"Style reference audio shape: {style_tensor.shape}, sample rate: {style_sr}") print(f"Timbre reference audio shape: {timbre_tensor.shape}, sample rate: {timbre_sr}") # Save uploaded audio torchaudio.save(temp_content_path, content_tensor, content_sr) torchaudio.save(temp_style_path, style_tensor, style_sr) torchaudio.save(temp_timbre_path, timbre_tensor, timbre_sr) try: # Get pipeline pipeline = get_pipeline("voice") # Inference gen_audio = pipeline.inference_ar_and_fm( src_wav_path=temp_content_path, src_text=None, style_ref_wav_path=temp_style_path, timbre_ref_wav_path=temp_timbre_path, ) # Check if generated audio is numerical anomaly if torch.isnan(gen_audio).any() or torch.isinf(gen_audio).any(): print("Warning: Generated audio contains NaN or Inf values") gen_audio = torch.nan_to_num(gen_audio, nan=0.0, posinf=0.95, neginf=-0.95) print(f"Generated audio shape: {gen_audio.shape}, max: {torch.max(gen_audio)}, min: {torch.min(gen_audio)}") # Save generated audio save_audio(gen_audio, output_path=output_path) return output_path except Exception as e: print(f"Error during processing: {e}") import traceback traceback.print_exc() raise e @spaces.GPU() def vevo_tts(text, ref_wav, timbre_ref_wav=None, style_ref_text=None, src_language="en", ref_language="en", style_ref_text_language="en"): temp_ref_path = "wav/temp_ref.wav" temp_timbre_path = "wav/temp_timbre.wav" output_path = "wav/output_vevotts.wav" # Check and process audio data if ref_wav is None: raise ValueError("Please upload a reference audio file") # Process reference audio format if isinstance(ref_wav, tuple) and len(ref_wav) == 2: if isinstance(ref_wav[0], np.ndarray): ref_data, ref_sr = ref_wav else: ref_sr, ref_data = ref_wav # Ensure single channel if len(ref_data.shape) > 1 and ref_data.shape[1] > 1: ref_data = np.mean(ref_data, axis=1) # Resample to 24kHz if ref_sr != 24000: ref_tensor = torch.FloatTensor(ref_data).unsqueeze(0) ref_tensor = torchaudio.functional.resample(ref_tensor, ref_sr, 24000) ref_sr = 24000 else: ref_tensor = torch.FloatTensor(ref_data).unsqueeze(0) # Normalize volume ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95 else: raise ValueError("Invalid reference audio format") # Print debug information print(f"Reference audio shape: {ref_tensor.shape}, sample rate: {ref_sr}") if style_ref_text: print(f"Style reference text: {style_ref_text}, language: {style_ref_text_language}") # Save uploaded audio torchaudio.save(temp_ref_path, ref_tensor, ref_sr) if timbre_ref_wav is not None: if isinstance(timbre_ref_wav, tuple) and len(timbre_ref_wav) == 2: if isinstance(timbre_ref_wav[0], np.ndarray): timbre_data, timbre_sr = timbre_ref_wav else: timbre_sr, timbre_data = timbre_ref_wav # Ensure single channel if len(timbre_data.shape) > 1 and timbre_data.shape[1] > 1: timbre_data = np.mean(timbre_data, axis=1) # Resample to 24kHz if timbre_sr != 24000: timbre_tensor = torch.FloatTensor(timbre_data).unsqueeze(0) timbre_tensor = torchaudio.functional.resample(timbre_tensor, timbre_sr, 24000) timbre_sr = 24000 else: timbre_tensor = torch.FloatTensor(timbre_data).unsqueeze(0) # Normalize volume timbre_tensor = timbre_tensor / (torch.max(torch.abs(timbre_tensor)) + 1e-6) * 0.95 print(f"Timbre reference audio shape: {timbre_tensor.shape}, sample rate: {timbre_sr}") torchaudio.save(temp_timbre_path, timbre_tensor, timbre_sr) else: raise ValueError("Invalid timbre reference audio format") else: temp_timbre_path = temp_ref_path try: # Get pipeline pipeline = get_pipeline("tts") # Inference gen_audio = pipeline.inference_ar_and_fm( src_wav_path=None, src_text=text, style_ref_wav_path=temp_ref_path, timbre_ref_wav_path=temp_timbre_path, style_ref_wav_text=style_ref_text, src_text_language=src_language, style_ref_wav_text_language=style_ref_text_language, ) # Check if generated audio is numerical anomaly if torch.isnan(gen_audio).any() or torch.isinf(gen_audio).any(): print("Warning: Generated audio contains NaN or Inf values") gen_audio = torch.nan_to_num(gen_audio, nan=0.0, posinf=0.95, neginf=-0.95) print(f"Generated audio shape: {gen_audio.shape}, max: {torch.max(gen_audio)}, min: {torch.min(gen_audio)}") # Save generated audio save_audio(gen_audio, output_path=output_path) return output_path except Exception as e: print(f"Error during processing: {e}") import traceback traceback.print_exc() raise e # Create Gradio interface with gr.Blocks(title="Vevo: Controllable Zero-Shot Voice Imitation with Self-Supervised Disentanglement") as demo: gr.Markdown("# Vevo: Controllable Zero-Shot Voice Imitation with Self-Supervised Disentanglement") # Add link tag line with gr.Row(elem_id="links_row"): gr.HTML("""
arXiv Paper ICLR Paper HuggingFace Model GitHub Repo
""") with gr.Tab("Vevo-Timbre"): gr.Markdown("### Vevo-Timbre: Maintain style but transfer timbre") with gr.Row(): with gr.Column(): timbre_content = gr.Audio(label="Source Audio", type="numpy") timbre_reference = gr.Audio(label="Timbre Reference", type="numpy") timbre_button = gr.Button("Generate") with gr.Column(): timbre_output = gr.Audio(label="Result") timbre_button.click(vevo_timbre, inputs=[timbre_content, timbre_reference], outputs=timbre_output) with gr.Tab("Vevo-Style"): gr.Markdown("### Vevo-Style: Maintain timbre but transfer style (accent, emotion, etc.)") with gr.Row(): with gr.Column(): style_content = gr.Audio(label="Source Audio", type="numpy") style_reference = gr.Audio(label="Style Reference", type="numpy") style_button = gr.Button("Generate") with gr.Column(): style_output = gr.Audio(label="Result") style_button.click(vevo_style, inputs=[style_content, style_reference], outputs=style_output) with gr.Tab("Vevo-Voice"): gr.Markdown("### Vevo-Voice: Transfers both style and timbre with separate references") with gr.Row(): with gr.Column(): voice_content = gr.Audio(label="Source Audio", type="numpy") voice_style_reference = gr.Audio(label="Style Reference", type="numpy") voice_timbre_reference = gr.Audio(label="Timbre Reference", type="numpy") voice_button = gr.Button("Generate") with gr.Column(): voice_output = gr.Audio(label="Result") voice_button.click(vevo_voice, inputs=[voice_content, voice_style_reference, voice_timbre_reference], outputs=voice_output) with gr.Tab("Vevo-TTS"): gr.Markdown("### Vevo-TTS: Text-to-speech with separate style and timbre references") with gr.Row(): with gr.Column(): tts_text = gr.Textbox(label="Target Text", placeholder="Enter text to synthesize...", lines=3) tts_src_language = gr.Dropdown(["en", "zh", "de", "fr", "ja", "ko"], label="Text Language", value="en") tts_reference = gr.Audio(label="Style Reference", type="numpy") tts_style_ref_text = gr.Textbox(label="Style Reference Text", placeholder="Enter style reference text...", lines=3) tts_style_ref_text_language = gr.Dropdown(["en", "zh", "de", "fr", "ja", "ko"], label="Style Reference Text Language", value="en") tts_timbre_reference = gr.Audio(label="Timbre Reference", type="numpy") tts_button = gr.Button("Generate") with gr.Column(): tts_output = gr.Audio(label="Result") tts_button.click( vevo_tts, inputs=[tts_text, tts_reference, tts_timbre_reference, tts_style_ref_text, tts_src_language, tts_style_ref_text_language], outputs=tts_output ) gr.Markdown(""" ## About VEVO VEVO is a versatile voice synthesis and conversion model that offers four main functionalities: 1. **Vevo-Style**: Maintains timbre but transfers style (accent, emotion, etc.) 2. **Vevo-Timbre**: Maintains style but transfers timbre 3. **Vevo-Voice**: Transfers both style and timbre with separate references 4. **Vevo-TTS**: Text-to-speech with separate style and timbre references For more information, visit the [Amphion project](https://github.com/open-mmlab/Amphion) """) # Launch application demo.launch()