app.py

import torch  
torch.manual_seed(160923)  

import gradio as gr  
from huggingface_hub import hf_hub_download  
from InferenceInterfaces.ControllableInterface import ControllableInterface  
from Utility.utils import float2pcm, load_json_from_path  

import matplotlib.pyplot as plt  
import librosa  
import librosa.display  
import numpy as np  
import io  
from PIL import Image  
import threading  

def generate_spectrogram_image(wav, sr):  
    fig, ax = plt.subplots(figsize=(4, 1.5))  
    D = librosa.amplitude_to_db(librosa.stft(wav, n_fft=512), ref=np.max)  
    img = librosa.display.specshow(D, sr=sr, x_axis='time', y_axis='log', ax=ax)  
    fig.colorbar(img, ax=ax)  
    buf = io.BytesIO()  
    plt.savefig(buf, format='png', bbox_inches='tight', pad_inches=0.05)  
    plt.close(fig)  
    buf.seek(0)  
    return Image.open(buf)  

class TTSWebUI:  

    def __init__(self,  
                 gpu_id="cpu",  
                 title="Controllable Text-to-Speech for over 7000 Languages",  
                 article="Thank you Hugging Face 🤗 for the GPU!<br>More: https://github.com/DigitalPhonetics/IMS-Toucan",  
                 tts_model_path=None,  
                 vocoder_model_path=None,  
                 embedding_gan_path=None,  
                 available_artificial_voices=10):  

        path_to_iso_list = hf_hub_download(repo_id="Flux9665/ToucanTTS", filename="iso_to_fullname.json")  
        iso_to_name = load_json_from_path(path_to_iso_list)  
        self.text_selection = [f"{iso_to_name[iso]} ({iso})" for iso in iso_to_name]  

        self.controllable_ui = ControllableInterface(  
            gpu_id=gpu_id,  
            available_artificial_voices=available_artificial_voices,  
            tts_model_path=tts_model_path,  
            vocoder_model_path=vocoder_model_path,  
            embedding_gan_path=embedding_gan_path  
        )  

        self.title = title  
        self.article = article  
        self.available_artificial_voices = available_artificial_voices  

    def read(self,  
             prompt,  
             language,  
             prosody_creativity,  
             duration_scaling_factor,  
             voice_seed,  
             emb1,  
             reference_audio):  

        lang_code = language[-4:-1]  
        result = [None]  

        def run_tts():  
            try:  
                print("[INFO] Running TTS with prompt:", prompt)  
                result[0] = self.controllable_ui.read(  
                    prompt,  
                    reference_audio,  
                    lang_code,  
                    lang_code,  
                    voice_seed,  
                    prosody_creativity,  
                    duration_scaling_factor,  
                    1.0, 1.0, 1.0,  
                    emb1,  
                    0.0, 0.0, 0.0, 0.0, 0.0,  
                    -24.0  
                )  
            except Exception as e:  
                print("[ERROR] Exception during TTS:", e)  
                result[0] = e  

        thread = threading.Thread(target=run_tts)  
        thread.start()  
        thread.join()  # ❗ Đã bỏ timeout để không bị cắt sớm trên CPU chậm

        if thread.is_alive():  
            print("[WARNING] TTS thread still alive after join → Timeout logic (shouldn't happen now)")  
            return None, generate_spectrogram_image(np.zeros(16000), 16000)  

        if isinstance(result[0], Exception):  
            print("[ERROR] TTS returned exception object:", result[0])  
            raise result[0]  

        if result[0] is None:  
            print("[ERROR] TTS returned None — possible silent failure")  
            return None, generate_spectrogram_image(np.zeros(16000), 16000)  

        sr, wav, _ = result[0]  

        print("[INFO] TTS success — sample rate:", sr, " | waveform shape:", wav.shape)  

        fig = generate_spectrogram_image(wav, sr)  
        return (sr, float2pcm(wav)), fig  

    def launch(self):  
        gr.Interface(  
            fn=self.read,  
            inputs=[  
                gr.Textbox(lines=2, placeholder="Type something...", value="What I cannot create, I do not understand.", label="Text input"),  
                gr.Dropdown(self.text_selection, type="value", value='English (eng)', label="Select the Language"),  
                gr.Slider(0.0, 0.8, step=0.1, value=0.5, label="Prosody Creativity"),  
                gr.Slider(0.7, 1.3, step=0.1, value=1.0, label="Faster - Slower"),  
                gr.Slider(0, self.available_artificial_voices, step=1, value=5, label="Random Voice Seed"),  
                gr.Slider(-10.0, 10.0, step=0.1, value=0.0, label="Gender"),  
                gr.Audio(type="filepath", show_label=True, container=True, label="[OPTIONAL] Voice to Clone"),  
            ],  
            outputs=[  
                gr.Audio(type="numpy", label="Speech"),  
                gr.Image(label="Visualization")  
            ],  
            title=self.title,  
            allow_flagging="never",  
            description=self.article,  
            theme=gr.themes.Ocean(primary_hue="amber", secondary_hue="orange")  
        ).launch()  

if __name__ == '__main__':  
    app = TTSWebUI(gpu_id="cpu")  
    app.launch()