app.py

import streamlit as st
import tempfile
from pydub import AudioSegment
from pyannote.audio import Pipeline
from faster_whisper import WhisperModel
from docx import Document
from io import BytesIO
import colorsys

# ページ設定
st.set_page_config(page_title="話者分離付き文字起こし", layout="centered")
st.title("🧠 話者分離付き文字起こしアプリ")

# Whisperモデル選択
model_size = st.selectbox("Whisperモデルを選択", ["tiny", "base", "small", "medium", "large-v2"], index=2)

# Hugging Face トークン入力
token_input = st.text_input("🔐 Hugging Face アクセストークンを入力", type="password")

# 音声アップロード
uploaded_file = st.file_uploader("🎵 音声ファイルをアップロード（mp3, wav, m4a）", type=["mp3", "wav", "m4a"])

# カラーパレット生成
def generate_color_palette(n):
    colors = []
    for i in range(n):
        hue = i / n
        lightness = 0.85
        saturation = 0.6
        rgb = colorsys.hls_to_rgb(hue, lightness, saturation)
        hex_color = '#%02x%02x%02x' % tuple(int(c * 255) for c in rgb)
        colors.append(hex_color)
    return colors

# 処理スタート
if uploaded_file and token_input:
    st.audio(uploaded_file)

    if st.button("▶️ 文字起こしスタート"):
        status = st.info("準備中…")
        progress = st.progress(0)

        try:
            # 音声を.wavに変換
            with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
                audio = AudioSegment.from_file(uploaded_file)
                audio.export(tmp.name, format="wav")
                audio_path = tmp.name

            progress.progress(20)
            status.info("🔍 話者分離中...")
            pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization", use_auth_token=token_input)
            diarization = pipeline(audio_path)

            progress.progress(50)
            status.info("📝 Whisperで文字起こし中...")
            model = WhisperModel(model_size, compute_type="int8")
            segments, _ = model.transcribe(audio_path, language="ja", vad_filter=True)

            # 話者割当
            word_blocks = []
            for segment in segments:
                start = segment.start
                speaker = "unknown"
                for turn in diarization.itertracks(yield_label=True):
                    if turn[0].start <= start <= turn[0].end:
                        speaker = turn[2]
                        break
                word_blocks.append((speaker, segment.text.strip()))

            progress.progress(80)
            status.success("✅ 完了！")

            # 表示と色分け
            st.subheader("🗣️ 話者ごとの文字起こし")
            unique_speakers = sorted(set(s for s, _ in word_blocks))
            colors = generate_color_palette(len(unique_speakers))
            color_map = {spk: col for spk, col in zip(unique_speakers, colors)}

            for speaker, text in word_blocks:
                st.markdown(
                    f"<div style='background-color:{color_map[speaker]}; padding:8px; border-radius:5px; margin-bottom:6px;'>"
                    f"<b>{speaker}</b>: {text}</div>",
                    unsafe_allow_html=True
                )

            # Wordファイル出力
            doc = Document()
            for speaker, text in word_blocks:
                doc.add_paragraph(f"{speaker}: {text}")
            docx_io = BytesIO()
            doc.save(docx_io)
            docx_io.seek(0)

            st.download_button("💾 Wordファイルでダウンロード", docx_io, file_name="transcription.docx", mime="application/vnd.openxmlformats-officedocument.wordprocessingml.document")

            progress.progress(100)

        except Exception as e:
            st.error(f"❌ エラーが発生しました:\n\n{e}")
elif uploaded_file and not token_input:
    st.warning("🔐 Hugging Face のトークンを入力してください。")