from inference.infer_tool import Svc from vextract.vocal_extract import VEX import gradio as gr import os # os.environ['CUDA_VISIBLE_DEVICES'] = '1,2' class VitsGradio: def __init__(self): self.so = Svc() self.v = VEX() self.lspk = [] self.modelPaths = [] for root, dirs, files in os.walk("checkpoints"): for dir in dirs: self.modelPaths.append(dir) with gr.Blocks(title="Sovits Singing Synthesis Tool") as self.Vits: gr.Markdown( """ # Singing Synthesis Tool - Please select the voice model, device, and operating mode in sequence, then click "Load Model" - The input audio needs to be clean vocals """ ) with gr.Tab("Vocal Extraction"): with gr.Row(): with gr.Column(): sample_audio = gr.Audio(label="Input Audio") extractAudioBtn = gr.Button("Extract Vocals") with gr.Row(): with gr.Column(): self.sample_vocal_output = gr.Audio(label="Output Audio") self.sample_accompaniment_output = gr.Audio() extractAudioBtn.click(self.v.separate, inputs=[sample_audio], outputs=[self.sample_vocal_output, self.sample_accompaniment_output], show_progress=True, api_name="extract") with gr.Tab("Singing Synthesis"): with gr.Row(visible=False) as self.VoiceConversion: with gr.Column(): with gr.Row(): with gr.Column(): self.srcaudio = gr.Audio(label="Input Audio") self.btnVC = gr.Button("Speaker Conversion") with gr.Column(): with gr.Row(): with gr.Column(): self.dsid0 = gr.Dropdown(label="Target Character", choices=self.lspk) self.tran = gr.Slider(label="Pitch Shift", maximum=60, minimum=-60, step=1, value=0) self.th = gr.Slider(label="Slice Threshold", maximum=32767, minimum=-32768, step=0.1, value=-40) self.ns = gr.Slider(label="Noise Level", maximum=1.0, minimum=0.0, step=0.1, value=0.4) with gr.Row(): self.VCOutputs = gr.Audio() self.btnVC.click(self.so.inference, inputs=[self.srcaudio, self.dsid0, self.tran, self.th, self.ns], outputs=[self.VCOutputs], show_progress=True, api_name="run") with gr.Row(visible=False) as self.VoiceBatchConversion: with gr.Column(): with gr.Row(): with gr.Column(): self.srcaudio = gr.Files(label="Upload Multiple Audio Files", file_types=['.wav'], interactive=True) self.btnVC = gr.Button("Speaker Conversion") with gr.Column(): with gr.Row(): with gr.Column(): self.dsid1 = gr.Dropdown(label="Target Character", choices=self.lspk) self.tran = gr.Slider(label="Pitch Shift", maximum=60, minimum=-60, step=1, value=0) self.th = gr.Slider(label="Slice Threshold", maximum=32767, minimum=-32768, step=0.1, value=-40) self.ns = gr.Slider(label="Noise Level", maximum=1.0, minimum=0.0, step=0.1, value=0.4) with gr.Row(): self.VCOutputs = gr.File(label="Output Zip File", interactive=False) self.btnVC.click(self.batch_inference, inputs=[self.srcaudio, self.dsid1, self.tran, self.th, self.ns], outputs=[self.VCOutputs], show_progress=True, api_name="batch") with gr.Row(): with gr.Column(): modelstrs = gr.Dropdown(label="Model", choices=self.modelPaths, value=self.modelPaths[0], type="value") devicestrs = gr.Dropdown(label="Device", choices=["cpu", "cuda"], value="cuda", type="value") isbatchmod = gr.Radio(label="Operating Mode", choices=["single", "batch"], value="single", info="single: Single file processing. batch: Batch processing supports uploading multiple files") btnMod = gr.Button("Load Model") btnMod.click(self.loadModel, inputs=[modelstrs, devicestrs, isbatchmod], outputs=[self.dsid0, self.dsid1, self.VoiceConversion, self.VoiceBatchConversion], show_progress=True, api_name="switch") def batch_inference(self, files, chara, tran, slice_db, ns, progress=gr.Progress()): from zipfile import ZipFile from scipy.io import wavfile import uuid temp_directory = "temp" if not os.path.exists(temp_directory): os.mkdir(temp_directory) progress(0.00, desc="Initializing Directory") tmp_workdir_name = f"{temp_directory}/batch_{uuid.uuid4()}" if not os.path.exists(tmp_workdir_name): os.mkdir(tmp_workdir_name) progress(0.10, desc="Initializing Directory") output_files = [] for idx, file in enumerate(files): filename = os.path.basename(file.name) progress(0.10 + (0.70 / float(len(files))) * (idx + 1.00), desc=f"Processing Audio {(idx + 1)}/{len(files)}: {filename}") print(f"{idx}, {file}, {filename}") sampling_rate, audio = wavfile.read(file.name) output_sampling_rate, output_audio = self.so.inference((sampling_rate, audio), chara=chara, tran=tran, slice_db=slice_db, ns=ns) new_filepath = f"{tmp_workdir_name}/{filename}" wavfile.write(filename=new_filepath, rate=output_sampling_rate, data=output_audio) output_files.append(new_filepath) progress(0.70, desc="Audio Processing Complete") zipfilename = f"{tmp_workdir_name}/output.zip" with ZipFile(zipfilename, "w") as zip_obj: for idx, filepath in enumerate(output_files): zip_obj.write(filepath, os.path.basename(filepath)) progress(0.80, desc="Compression Complete") # todo: remove data progress(1.00, desc="Cleaning Up") return zipfilename def loadModel(self, path, device, process_mode): self.lspk = [] print(f"path: {path}, device: {device}") self.so.set_device(device) print(f"device set.") self.so.load_checkpoint(path) print(f"checkpoint loaded") for spk, sid in self.so.hps_ms.spk.items(): self.lspk.append(spk) print(f"LSPK: {self.lspk}") if process_mode == "single": VChange = gr.update(visible=True) VBChange = gr.update(visible=False) else: VChange = gr.update(visible=False) VBChange = gr.update(visible=True) SD0Change = gr.update(choices=self.lspk, value=self.lspk[0]) SD1Change = gr.update(choices=self.lspk, value=self.lspk[0]) print("All set. Updating display") return [SD0Change, SD1Change, VChange, VBChange] if __name__ == "__main__": grVits = VitsGradio() grVits.Vits\ .queue(concurrency_count=20, status_update_rate=5.0)\ .launch(server_port=7870, share=True, show_api=False)