app.py

import torch
import torchaudio
import gradio as gr
import torch.nn.functional as F
from transformers import WavLMForXVector, Wav2Vec2FeatureExtractor

# 準備模型
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = WavLMForXVector.from_pretrained("microsoft/wavlm-base-sv").to(device)
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("microsoft/wavlm-base-sv")

# 音訊處理函式
def preprocess(audio):
    if audio is None:
        return None
    waveform, sr = torchaudio.load(audio)
    if sr != 16000:
        waveform = torchaudio.functional.resample(waveform, sr, 16000)
    return waveform.squeeze(0)

# 取得 normalized embedding
def get_embedding(waveform):
    inputs = feature_extractor(waveform.numpy(), sampling_rate=16000, return_tensors="pt", padding=True)
    inputs = {k: v.to(device) for k, v in inputs.items()}
    with torch.no_grad():
        embedding = model(**inputs).embeddings
    return F.normalize(embedding, p=2, dim=1)

# 主處理函式
def compare_audio(native_audio, user_audio):

    native_wav = preprocess(native_audio)
    user_wav = preprocess(user_audio)

    if native_wav is None or user_wav is None:
        return "請上傳兩段語音"

    emb1 = get_embedding(native_wav)
    emb2 = get_embedding(user_wav)
    similarity = F.cosine_similarity(emb1, emb2).item()
    score = round(similarity * 100, 2)  # 轉換為 0~100 分數

    # 評語
    if score > 90:
        feedback = "非常接近！你模仿得很好 👏"
    elif score > 75:
        feedback = "不錯，再接再厲 👍"
    elif score > 60:
        feedback = "有些相似，但還有改進空間 🙂"
    else:
        feedback = "相似度不高，請再試一次 😅"

    return f"相似度分數：{score}/100\n{feedback}"

# Gradio UI
title = "🎤 語音模仿評分器"
description = "上傳 native speaker 的語音，以及你模仿的語音，系統會幫你評分你的發音相似度。"

demo = gr.Interface(
    fn=compare_audio,
    inputs=[
        gr.Audio(type="filepath", label="📢 Native Speaker 語音"),
        gr.Audio(type="filepath", label="🗣️ 你的模仿錄音"),
    ],
    outputs="text",
    title=title,
    description=description,
)

if __name__ == "__main__":
    demo.launch()