app.py

import gradio as gr
import torch
from speechbrain.inference.TTS import Tacotron2
from speechbrain.inference.vocoders import HIFIGAN
from scipy.io.wavfile import write

# 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):
    # Tacotron2でmel spectrogramを生成
    # テキストを直接入力として、LongTensorでラップする
    mel_output, mel_length, alignment = tacotron2.encode_batch([text])

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

    # torch tensorをwavfileとして保存
    # sampling rate を 22050に設定
    write("speech.wav", rate=22050, data=waveforms.squeeze().cpu().numpy())
    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()