import gradio as gr import numpy as np import soundfile as sf import spaces import torch import torchaudio import librosa import yaml import tempfile import os from huggingface_hub import hf_hub_download from transformers import AutoFeatureExtractor, WhisperModel from torch.nn.utils import parametrizations from modules.commons import build_model, load_checkpoint, recursive_munch from modules.campplus.DTDNN import CAMPPlus from modules.bigvgan import bigvgan from modules.rmvpe import RMVPE from modules.audio import mel_spectrogram # ---------------------------- # Optimization Settings # ---------------------------- # Set the number of threads to the number of CPU cores torch.set_num_threads(os.cpu_count()) # Enable optimized backends torch.backends.openmp.enabled = True torch.backends.mkldnn.enabled = True torch.backends.cudnn.enabled = False torch.backends.cuda.enabled = False torch.set_grad_enabled(False) # Force CPU usage device = torch.device("cpu") print(f"[DEVICE] | Using device: {device}") # ---------------------------- # Load Models and Configuration # ---------------------------- def load_custom_model_from_hf(repo_id, model_filename="pytorch_model.bin", config_filename="config.yml"): os.makedirs("./checkpoints", exist_ok=True) model_path = hf_hub_download(repo_id=repo_id, filename=model_filename, cache_dir="./checkpoints") if config_filename is None: return model_path config_path = hf_hub_download(repo_id=repo_id, filename=config_filename, cache_dir="./checkpoints") return model_path, config_path # Load DiT model dit_checkpoint_path, dit_config_path = load_custom_model_from_hf( "Plachta/Seed-VC", "DiT_seed_v2_uvit_whisper_small_wavenet_bigvgan_pruned.pth", "config_dit_mel_seed_uvit_whisper_small_wavenet.yml" ) config = yaml.safe_load(open(dit_config_path, 'r')) model_params = recursive_munch(config['model_params']) model = build_model(model_params, stage='DiT') # Debug: Print model keys to identify correct key print(f"[INFO] | Model keys: {model.keys()}") hop_length = config['preprocess_params']['spect_params']['hop_length'] sr = config['preprocess_params']['sr'] # Load DiT checkpoints model, _, _, _ = load_checkpoint(model, None, dit_checkpoint_path, load_only_params=True, ignore_modules=[], is_distributed=False) for key in model: model[key].eval() model[key].to(device) print("[INFO] | DiT model loaded and set to eval mode.") model.cfm.estimator.setup_caches(max_batch_size=1, max_seq_length=8192) # Ensure 'CAMPPlus' is correctly imported and defined try: campplus_model = CAMPPlus(feat_dim=80, embedding_size=192) print("[INFO] | CAMPPlus model instantiated.") except NameError: print("[ERROR] | CAMPPlus is not defined. Please check the import path and ensure CAMPPlus is correctly defined.") raise # Set weights_only=True for security campplus_ckpt_path = load_custom_model_from_hf("funasr/campplus", "campplus_cn_common.bin", config_filename=None) campplus_state = torch.load(campplus_ckpt_path, map_location="cpu", weights_only=True) campplus_model.load_state_dict(campplus_state) campplus_model.eval() campplus_model.to(device) print("[INFO] | CAMPPlus model loaded, set to eval mode, and moved to CPU.") # Load BigVGAN model bigvgan_model = bigvgan.BigVGAN.from_pretrained('nvidia/bigvgan_v2_22khz_80band_256x', use_cuda_kernel=False) bigvgan_model.remove_weight_norm() bigvgan_model = bigvgan_model.eval().to(device) print("[INFO] | BigVGAN model loaded, weight norm removed, set to eval mode, and moved to CPU.") # Load FAcodec model ckpt_path, config_path = load_custom_model_from_hf("Plachta/FAcodec", 'pytorch_model.bin', 'config.yml') codec_config = yaml.safe_load(open(config_path)) codec_model_params = recursive_munch(codec_config['model_params']) codec_encoder = build_model(codec_model_params, stage="codec") ckpt_params = torch.load(ckpt_path, map_location="cpu", weights_only=True) for key in codec_encoder: codec_encoder[key].load_state_dict(ckpt_params[key], strict=False) codec_encoder = {k: v.eval().to(device) for k, v in codec_encoder.items()} print("[INFO] | FAcodec model loaded, set to eval mode, and moved to CPU.") # Load Whisper model with float32 and compatible size whisper_name = model_params.speech_tokenizer.whisper_name if hasattr(model_params.speech_tokenizer, 'whisper_name') else "openai/whisper-small" whisper_model = WhisperModel.from_pretrained(whisper_name, torch_dtype=torch.float32).to(device) del whisper_model.decoder # Remove decoder as it's not used whisper_feature_extractor = AutoFeatureExtractor.from_pretrained(whisper_name) print(f"[INFO] | Whisper model '{whisper_name}' loaded with dtype {whisper_model.dtype} and moved to CPU.") # Generate mel spectrograms with optimized parameters mel_fn_args = { "n_fft": 1024, "win_size": 1024, "hop_size": 256, "num_mels": 80, "sampling_rate": sr, "fmin": 0, "fmax": None, "center": False } to_mel = lambda x: mel_spectrogram(x, **mel_fn_args) # Load F0 conditioned model dit_checkpoint_path_f0, dit_config_path_f0 = load_custom_model_from_hf( "Plachta/Seed-VC", "DiT_seed_v2_uvit_whisper_base_f0_44k_bigvgan_pruned_ft_ema.pth", "config_dit_mel_seed_uvit_whisper_base_f0_44k.yml" ) config_f0 = yaml.safe_load(open(dit_config_path_f0, 'r')) model_params_f0 = recursive_munch(config_f0['model_params']) model_f0 = build_model(model_params_f0, stage='DiT') hop_length_f0 = config_f0['preprocess_params']['spect_params']['hop_length'] sr_f0 = config_f0['preprocess_params']['sr'] # Load F0 model checkpoints model_f0, _, _, _ = load_checkpoint(model_f0, None, dit_checkpoint_path_f0, load_only_params=True, ignore_modules=[], is_distributed=False) for key in model_f0: model_f0[key].eval() model_f0[key].to(device) print("[INFO] | F0 conditioned DiT model loaded and set to eval mode.") model_f0.cfm.estimator.setup_caches(max_batch_size=1, max_seq_length=8192) # Load F0 extractor model_path = load_custom_model_from_hf("lj1995/VoiceConversionWebUI", "rmvpe.pt", None) rmvpe = RMVPE(model_path, is_half=False, device=device) print("[INFO] | RMVPE model loaded and moved to CPU.") mel_fn_args_f0 = { "n_fft": config_f0['preprocess_params']['spect_params']['n_fft'], "win_size": config_f0['preprocess_params']['spect_params']['win_length'], "hop_size": config_f0['preprocess_params']['spect_params']['hop_length'], "num_mels": 80, # Ensure this matches the primary model "sampling_rate": sr_f0, "fmin": 0, "fmax": None, "center": False } to_mel_f0 = lambda x: mel_spectrogram(x, **mel_fn_args_f0) # Load BigVGAN 44kHz model bigvgan_44k_model = bigvgan.BigVGAN.from_pretrained('nvidia/bigvgan_v2_44khz_128band_512x', use_cuda_kernel=False) bigvgan_44k_model.remove_weight_norm() bigvgan_44k_model = bigvgan_44k_model.eval().to(device) print("[INFO] | BigVGAN 44kHz model loaded, weight norm removed, set to eval mode, and moved to CPU.") # ---------------------------- # Helper Functions # ---------------------------- def adjust_f0_semitones(f0_sequence, n_semitones): factor = 2 ** (n_semitones / 12) return f0_sequence * factor def crossfade(chunk1, chunk2, overlap): fade_out = np.cos(np.linspace(0, np.pi / 2, overlap)) ** 2 fade_in = np.cos(np.linspace(np.pi / 2, 0, overlap)) ** 2 chunk2[:overlap] = chunk2[:overlap] * fade_in + chunk1[-overlap:] * fade_out return chunk2 # ---------------------------- # Voice Conversion Function # ---------------------------- @torch.no_grad() @torch.inference_mode() def voice_conversion(source, target, diffusion_steps, length_adjust, inference_cfg_rate, f0_condition, auto_f0_adjust, pitch_shift): print("[INFO] | Voice conversion started.") inference_module = model if not f0_condition else model_f0 mel_fn = to_mel if not f0_condition else to_mel_f0 bigvgan_fn = bigvgan_model if not f0_condition else bigvgan_44k_model sr_current = 22050 if not f0_condition else 44100 hop_length_current = 256 if not f0_condition else 512 max_context_window = sr_current // hop_length_current * 30 overlap_wave_len = 16 * hop_length_current bitrate = "320k" # Load audio using librosa print("[INFO] | Loading source and reference audio.") source_audio, _ = librosa.load(source, sr=sr_current) ref_audio, _ = librosa.load(target, sr=sr_current) # Clip reference audio to 25 seconds ref_audio = ref_audio[:sr_current * 25] print(f"[INFO] | Source audio length: {len(source_audio)/sr_current:.2f}s, Reference audio length: {len(ref_audio)/sr_current:.2f}s") # Convert audio to tensors source_audio_tensor = torch.tensor(source_audio).unsqueeze(0).float().to(device) ref_audio_tensor = torch.tensor(ref_audio).unsqueeze(0).float().to(device) # Resample to 16kHz ref_waves_16k = torchaudio.functional.resample(ref_audio_tensor, sr_current, 16000) converted_waves_16k = torchaudio.functional.resample(source_audio_tensor, sr_current, 16000) # Generate Whisper features print("[INFO] | Generating Whisper features for source audio.") if converted_waves_16k.size(-1) <= 16000 * 30: alt_inputs = whisper_feature_extractor( [converted_waves_16k.squeeze(0).cpu().numpy()], return_tensors="pt", return_attention_mask=True, sampling_rate=16000 ) alt_input_features = whisper_model._mask_input_features( alt_inputs.input_features, attention_mask=alt_inputs.attention_mask ).to(device) alt_outputs = whisper_model.encoder( alt_input_features.to(torch.float32), head_mask=None, output_attentions=False, output_hidden_states=False, return_dict=True ) S_alt = alt_outputs.last_hidden_state.to(torch.float32) S_alt = S_alt[:, :converted_waves_16k.size(-1) // 320 + 1] print(f"[INFO] | S_alt shape: {S_alt.shape}") else: # Process in chunks print("[INFO] | Processing source audio in chunks.") overlapping_time = 5 # seconds chunk_size = 16000 * 30 # 30 seconds overlap_size = 16000 * overlapping_time S_alt_list = [] buffer = None traversed_time = 0 total_length = converted_waves_16k.size(-1) while traversed_time < total_length: if buffer is None: chunk = converted_waves_16k[:, traversed_time:traversed_time + chunk_size] else: chunk = torch.cat([ buffer, converted_waves_16k[:, traversed_time:traversed_time + chunk_size - overlap_size] ], dim=-1) alt_inputs = whisper_feature_extractor( [chunk.squeeze(0).cpu().numpy()], return_tensors="pt", return_attention_mask=True, sampling_rate=16000 ) alt_input_features = whisper_model._mask_input_features( alt_inputs.input_features, attention_mask=alt_inputs.attention_mask ).to(device) alt_outputs = whisper_model.encoder( alt_input_features.to(torch.float32), head_mask=None, output_attentions=False, output_hidden_states=False, return_dict=True ) S_chunk = alt_outputs.last_hidden_state.to(torch.float32) S_chunk = S_chunk[:, :chunk.size(-1) // 320 + 1] print(f"[INFO] | Processed chunk with S_chunk shape: {S_chunk.shape}") if traversed_time == 0: S_alt_list.append(S_chunk) else: skip_frames = 50 * overlapping_time S_alt_list.append(S_chunk[:, skip_frames:]) buffer = chunk[:, -overlap_size:] traversed_time += chunk_size - overlap_size S_alt = torch.cat(S_alt_list, dim=1) print(f"[INFO] | Final S_alt shape after chunk processing: {S_alt.shape}") # Original Whisper features print("[INFO] | Generating Whisper features for reference audio.") ori_waves_16k = torchaudio.functional.resample(ref_audio_tensor, sr_current, 16000) ori_inputs = whisper_feature_extractor( [ori_waves_16k.squeeze(0).cpu().numpy()], return_tensors="pt", return_attention_mask=True, sampling_rate=16000 ) ori_input_features = whisper_model._mask_input_features( ori_inputs.input_features, attention_mask=ori_inputs.attention_mask ).to(device) ori_outputs = whisper_model.encoder( ori_input_features.to(torch.float32), head_mask=None, output_attentions=False, output_hidden_states=False, return_dict=True ) S_ori = ori_outputs.last_hidden_state.to(torch.float32) S_ori = S_ori[:, :ori_waves_16k.size(-1) // 320 + 1] print(f"[INFO] | S_ori shape: {S_ori.shape}") # Generate mel spectrograms print("[INFO] | Generating mel spectrograms.") mel = mel_fn(source_audio_tensor.float()) mel2 = mel_fn(ref_audio_tensor.float()) print(f"[INFO] | Mel spectrogram shapes: mel={mel.shape}, mel2={mel2.shape}") # Length adjustment target_lengths = torch.LongTensor([int(mel.size(2) * length_adjust)]).to(mel.device) target2_lengths = torch.LongTensor([mel2.size(2)]).to(mel2.device) print(f"[INFO] | Target lengths: {target_lengths.item()}, {target2_lengths.item()}") # Extract style features print("[INFO] | Extracting style features from reference audio.") feat2 = torchaudio.compliance.kaldi.fbank( ref_waves_16k, num_mel_bins=80, dither=0, sample_frequency=16000 ) feat2 = feat2 - feat2.mean(dim=0, keepdim=True) style2 = campplus_model(feat2.unsqueeze(0)) print(f"[INFO] | Style2 shape: {style2.shape}") # F0 Conditioning if f0_condition: print("[INFO] | Performing F0 conditioning.") F0_ori = rmvpe.infer_from_audio(ref_waves_16k[0], thred=0.5) F0_alt = rmvpe.infer_from_audio(converted_waves_16k[0], thred=0.5) F0_ori = torch.from_numpy(F0_ori).to(device)[None].float() F0_alt = torch.from_numpy(F0_alt).to(device)[None].float() voiced_F0_ori = F0_ori[F0_ori > 1] voiced_F0_alt = F0_alt[F0_alt > 1] log_f0_alt = torch.log(F0_alt + 1e-5) voiced_log_f0_ori = torch.log(voiced_F0_ori + 1e-5) voiced_log_f0_alt = torch.log(voiced_F0_alt + 1e-5) median_log_f0_ori = torch.median(voiced_log_f0_ori) median_log_f0_alt = torch.median(voiced_log_f0_alt) # Shift F0 levels shifted_log_f0_alt = log_f0_alt.clone() if auto_f0_adjust: shifted_log_f0_alt[F0_alt > 1] = ( log_f0_alt[F0_alt > 1] - median_log_f0_alt + median_log_f0_ori ) shifted_f0_alt = torch.exp(shifted_log_f0_alt) if pitch_shift != 0: shifted_f0_alt[F0_alt > 1] = adjust_f0_semitones(shifted_f0_alt[F0_alt > 1], pitch_shift) print("[INFO] | F0 conditioning completed.") else: F0_ori = None F0_alt = None shifted_f0_alt = None print("[INFO] | F0 conditioning not applied.") # Length Regulation print("[INFO] | Applying length regulation.") cond, _, _, _, _ = inference_module.length_regulator(S_alt, ylens=target_lengths, n_quantizers=3, f0=shifted_f0_alt) prompt_condition, _, _, _, _ = inference_module.length_regulator(S_ori, ylens=target2_lengths, n_quantizers=3, f0=F0_ori) print(f"[INFO] | Cond shape: {cond.shape}, Prompt condition shape: {prompt_condition.shape}") # Initialize variables for audio generation max_source_window = max_context_window - mel2.size(2) processed_frames = 0 generated_wave_chunks = [] print("[INFO] | Starting inference and audio generation.") while processed_frames < cond.size(1): chunk_cond = cond[:, processed_frames:processed_frames + max_source_window] is_last_chunk = processed_frames + max_source_window >= cond.size(1) cat_condition = torch.cat([prompt_condition, chunk_cond], dim=1) # Perform inference vc_target = inference_module.cfm.inference( cat_condition, torch.LongTensor([cat_condition.size(1)]).to(mel2.device), mel2, style2, None, diffusion_steps, inference_cfg_rate=inference_cfg_rate ) vc_target = vc_target[:, :, mel2.size(2):] print(f"[INFO] | vc_target shape: {vc_target.shape}") # Generate waveform using BigVGAN vc_wave = bigvgan_fn(vc_target.float())[0] print(f"[INFO] | vc_wave shape: {vc_wave.shape}") # Handle the generated waveform output_wave = vc_wave[0].cpu().numpy() generated_wave_chunks.append(output_wave) # Fix: Ensure processed_frames increments correctly to avoid infinite loop processed_frames += vc_target.size(2) # Changed from 'vc_target.size(2) - 16' to 'vc_target.size(2)' print(f"[INFO] | Processed frames updated to: {processed_frames}") # Concatenate all generated wave chunks final_audio = np.concatenate(generated_wave_chunks).astype(np.float32) # Normalize the audio to ensure it's within [-1.0, 1.0] max_val = np.max(np.abs(final_audio)) if max_val > 1.0: final_audio = final_audio / max_val print("[INFO] | Final audio normalized.") # Save the audio to a temporary WAV file print("[INFO] | Saving final audio to a temporary WAV file.") with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_file: sf.write(tmp_file.name, final_audio, sr_current, format='WAV') temp_file_path = tmp_file.name print(f"[INFO] | Final audio saved to {temp_file_path}") return temp_file_path # ---------------------------- # Bypass GPU Initialization Error (KEEP THIS FUNCTION AND NEVER DELETE, OTHERWISE IT WILL ERROR DUE TO NOT AT LEAST ONE FUNCTION HAVING USE OF GPU) # ---------------------------- @spaces.GPU(duration=15) def gpu(): return # ---------------------------- # Gradio Interface # ---------------------------- description = ( "🪄 **Voice Conversion Tool**\n\n" "Upload your **Source Audio** and **Reference Audio** files to perform voice conversion. " "Adjust the sliders and checkboxes to customize the conversion process." ) inputs = [ gr.Audio(type="filepath", label="Source Audio"), gr.Audio(type="filepath", label="Reference Audio"), gr.Slider(minimum=1, maximum=100, value=25, step=1, label="Diffusion Steps", info="Default is 25. Use 50-100 for best quality."), gr.Slider(minimum=0.5, maximum=2.0, step=0.1, value=1.0, label="Length Adjustment", info="<1.0 to speed up speech, >1.0 to slow down speech."), gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=0.7, label="Inference CFG Rate", info="Has a subtle influence."), gr.Checkbox(label="Use F0 Conditioned Model", value=False, info="Must be enabled for singing voice conversion."), gr.Checkbox(label="Auto F0 Adjustment", value=True, info="Roughly adjusts F0 to match target voice. Only works when 'Use F0 Conditioned Model' is enabled."), gr.Slider(label='Pitch Shift (semitones)', minimum=-12, maximum=12, step=1, value=0, info="Pitch shift in semitones. Only works when 'Use F0 Conditioned Model' is enabled."), ] # Set outputs to a single gr.Audio component with type="filepath" outputs = gr.Audio(label="Full Output Audio", type="filepath") gr.Interface( fn=voice_conversion, description=description, inputs=inputs, outputs=outputs, title="Seed Voice Conversion", cache_examples=False, allow_flagging="never" ).launch(share=True)