import gradio as gr
import torch
import soundfile as sf
from speechbrain.inference.TTS import Tacotron2
from speechbrain.inference.vocoders import HIFIGAN
from speechbrain.dataio.dataio import read_audio

# モデルのロード
hifi_gan = HIFIGAN.from_hparams(source="speechbrain/tts-hifigan-ljspeech", savedir="tmpdir_vocoder")
tacotron2 = Tacotron2.from_hparams(source="speechbrain/tts-tacotron2-ljspeech", savedir="tmpdir_tts")

# 推論関数の定義
def synthesize_speech(text):
    # テキストをトークンIDに変換
    text = text.lower()
    tokenized = tacotron2.hparams.tokenize(text, phonemize=False)
    # トークンIDをLong型のテンソルに変換
    tokens = torch.LongTensor(tokenized)

    # Tacotron2でmel spectrogramを生成
    mel_output, mel_length, alignment = tacotron2.encode_batch(tokens)

    # HiFi-GANでmel spectrogramから音声を生成
    waveforms = hifi_gan.decode_batch(mel_output)

    # 音声を .wav 形式で保存
    sf.write("speech.wav", waveforms.squeeze().cpu().numpy(), samplerate=hifi_gan.hparams.sample_rate)
    return "speech.wav"

# Gradioインターフェースの作成
iface = gr.Interface(
    fn=synthesize_speech,
    inputs=gr.Textbox(lines=5, label="Input Text"),
    outputs=gr.Audio(label="Output Audio", type="filepath"),
    title="TTS Demo",
    description="Enter text to synthesize speech."
)

iface.launch()