app.py

"""
🎵 Lyric Sync — Gradio Space
Automatic perfect song lyric acquisition and synchronization.

Pipeline:
1. Vocal separation (Demucs htdemucs)
2. Word-level transcription (Whisper large-v3)
3. Song identification (AcoustID / transcript search)
4. Lyrics fetching (LRCLIB)
5. Sequence alignment (transfer timings to correct lyrics)
6. Timing refinement (onset/offset detection)
"""

import json
import logging
import os
import re
import subprocess
import tempfile
import time
import unicodedata
from dataclasses import dataclass, field
from difflib import SequenceMatcher
from pathlib import Path
from typing import Optional

import gradio as gr
import librosa
import numpy as np
import requests
import spaces
import torch
import torchaudio

logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
logger = logging.getLogger(__name__)


# ═══════════════════════════════════════════════════════════════════════════════
# AUDIO LOADING (ffmpeg-based, handles all formats)
# ═══════════════════════════════════════════════════════════════════════════════

def load_audio_as_tensor(audio_path: str, target_sr: int = 44100) -> tuple[torch.Tensor, int]:
    """
    Load any audio file into a torch tensor using librosa (ffmpeg backend).
    Returns: (tensor [channels, samples], sample_rate)
    Always returns stereo at target_sr.
    """
    # librosa.load handles MP3, FLAC, OGG, WAV, etc. via soundfile/ffmpeg
    y, sr = librosa.load(audio_path, sr=target_sr, mono=False)
    
    # y shape: (samples,) if mono or (channels, samples) if stereo
    if y.ndim == 1:
        # Mono → stereo by duplicating
        wav = torch.from_numpy(y).float().unsqueeze(0).repeat(2, 1)
    else:
        wav = torch.from_numpy(y).float()
        if wav.shape[0] > 2:
            wav = wav[:2]
    
    return wav, target_sr


# ═══════════════════════════════════════════════════════════════════════════════
# DATA CLASSES
# ═══════════════════════════════════════════════════════════════════════════════

@dataclass
class TimedWord:
    word: str
    start: float
    end: float
    confidence: float = 1.0

    @property
    def duration(self) -> float:
        return self.end - self.start


@dataclass
class SongInfo:
    title: str
    artist: str
    album: Optional[str] = None
    method: str = "unknown"


# ═══════════════════════════════════════════════════════════════════════════════
# STEP 1: VOCAL SEPARATION (Demucs)
# ═══════════════════════════════════════════════════════════════════════════════

_demucs_model = None


def get_demucs_model(device="cpu"):
    global _demucs_model
    if _demucs_model is None:
        from demucs.pretrained import get_model
        _demucs_model = get_model("htdemucs")
        _demucs_model.eval()
        logger.info("Demucs htdemucs loaded")
    _demucs_model.to(device)
    return _demucs_model


def separate_vocals(audio_path: str, device: str = "cpu") -> tuple[np.ndarray, int, np.ndarray, int]:
    """
    Separate vocals from audio.
    Returns: (vocals_16k_mono, 16000, vocals_44k_mono, 44100)
    """
    from demucs.apply import apply_model

    model = get_demucs_model(device)

    # Load audio using librosa (handles MP3, FLAC, OGG, etc.)
    wav, sr = load_audio_as_tensor(audio_path, target_sr=44100)

    # Run separation
    wav_batch = wav.unsqueeze(0).to(device)
    with torch.no_grad():
        sources = apply_model(
            model, wav_batch, device=device,
            shifts=1, split=True, overlap=0.25, progress=False,
        )
    # sources: [1, 4, 2, N] — drums, bass, other, vocals
    vocal_idx = model.sources.index("vocals")
    vocals_stereo = sources[0, vocal_idx].cpu()  # [2, N] at 44100
    vocals_44k = vocals_stereo.mean(dim=0)  # mono

    # Resample to 16kHz for Whisper
    vocals_16k = torchaudio.functional.resample(
        vocals_44k.unsqueeze(0), 44100, 16000
    ).squeeze(0)

    return vocals_16k.numpy(), 16000, vocals_44k.numpy(), 44100


# ═══════════════════════════════════════════════════════════════════════════════
# STEP 2: TRANSCRIPTION (Whisper)
# ═══════════════════════════════════════════════════════════════════════════════

_whisper_pipe = None


def get_whisper_pipeline(device="cpu"):
    global _whisper_pipe
    if _whisper_pipe is None:
        from transformers import pipeline as hf_pipeline
        _whisper_pipe = hf_pipeline(
            task="automatic-speech-recognition",
            model="openai/whisper-large-v3",
            dtype=torch.float16 if device == "cuda" else torch.float32,
            device=device,
            model_kwargs={"attn_implementation": "sdpa"},
        )
        logger.info(f"Whisper large-v3 pipeline loaded on {device}")
    return _whisper_pipe


def transcribe_vocals(audio: np.ndarray, sr: int = 16000) -> list[TimedWord]:
    pipe = get_whisper_pipeline("cuda" if torch.cuda.is_available() else "cpu")

    result = pipe(
        {"array": audio.astype(np.float32), "sampling_rate": sr},
        return_timestamps="word",
        generate_kwargs={
            "language": "english",
            "task": "transcribe",
        },
        chunk_length_s=30,
        stride_length_s=5,
    )

    words = []
    for chunk in result.get("chunks", []):
        text = chunk["text"].strip()
        ts = chunk.get("timestamp", (None, None))
        if text and ts[0] is not None and ts[1] is not None:
            words.append(TimedWord(word=text, start=ts[0], end=ts[1]))

    return words


# ═══════════════════════════════════════════════════════════════════════════════
# STEP 3: SONG IDENTIFICATION
# ═══════════════════════════════════════════════════════════════════════════════

def identify_song(audio_path, user_artist="", user_title="", transcript_text=""):
    if user_artist.strip() and user_title.strip():
        return SongInfo(title=user_title.strip(), artist=user_artist.strip(), method="user_provided")

    acoustid_key = os.environ.get("ACOUSTID_API_KEY")
    if acoustid_key:
        result = _acoustid_identify(audio_path, acoustid_key)
        if result:
            return result

    if transcript_text:
        words = transcript_text.split()
        if len(words) >= 5:
            mid = len(words) // 2
            fragment = " ".join(words[max(0, mid-5):mid+5])
            result = _search_lrclib(fragment)
            if result:
                return result
    return None


def _acoustid_identify(audio_path, api_key):
    try:
        result = subprocess.run(
            ["fpcalc", "-json", "-length", "120", audio_path],
            capture_output=True, text=True, timeout=30,
        )
        if result.returncode != 0:
            return None
        fp_data = json.loads(result.stdout)
        resp = requests.post("https://api.acoustid.org/v2/lookup", data={
            "client": api_key, "duration": fp_data["duration"],
            "fingerprint": fp_data["fingerprint"],
            "meta": "recordings releasegroups", "format": "json",
        }, timeout=15)
        resp.raise_for_status()
        data = resp.json()
        if data.get("status") != "ok" or not data.get("results"):
            return None
        best = max(data["results"], key=lambda r: r.get("score", 0))
        if best.get("score", 0) < 0.5 or not best.get("recordings"):
            return None
        rec = best["recordings"][0]
        artist = rec.get("artists", [{}])[0].get("name", "Unknown")
        album = rec.get("releasegroups", [{}])[0].get("title") if rec.get("releasegroups") else None
        return SongInfo(title=rec.get("title", "Unknown"), artist=artist, album=album, method="acoustid")
    except Exception as e:
        logger.warning(f"AcoustID failed: {e}")
        return None


def _search_lrclib(query):
    try:
        resp = requests.get("https://lrclib.net/api/search", params={"q": query}, timeout=10)
        if resp.status_code == 200:
            results = resp.json()
            if results:
                best = results[0]
                return SongInfo(title=best.get("trackName", "Unknown"), artist=best.get("artistName", "Unknown"), album=best.get("albumName"), method="lrclib_search")
    except Exception as e:
        logger.debug(f"LRCLIB search failed: {e}")
    return None


# ═══════════════════════════════════════════════════════════════════════════════
# STEP 4: LYRICS FETCHING
# ═══════════════════════════════════════════════════════════════════════════════

def fetch_lyrics(song):
    params = {"artist_name": song.artist, "track_name": song.title}
    if song.album:
        params["album_name"] = song.album
    try:
        resp = requests.get("https://lrclib.net/api/get", params=params, timeout=10)
        if resp.status_code == 200:
            data = resp.json()
            plain = data.get("plainLyrics") or data.get("syncedLyrics", "")
            if plain:
                plain = re.sub(r"\[\d{2}:\d{2}\.\d{2,3}\]\s*", "", plain)
                words = plain.split()
                if words:
                    return plain, words
    except Exception as e:
        logger.warning(f"LRCLIB fetch failed: {e}")
    try:
        resp = requests.get("https://lrclib.net/api/search", params={"q": f"{song.artist} {song.title}"}, timeout=10)
        if resp.status_code == 200:
            results = resp.json()
            if results:
                plain = results[0].get("plainLyrics") or results[0].get("syncedLyrics", "")
                plain = re.sub(r"\[\d{2}:\d{2}\.\d{2,3}\]\s*", "", plain)
                words = plain.split()
                if words:
                    return plain, words
    except Exception as e:
        logger.debug(f"LRCLIB search fallback failed: {e}")
    return None


# ═══════════════════════════════════════════════════════════════════════════════
# STEP 5: SEQUENCE ALIGNMENT
# ═══════════════════════════════════════════════════════════════════════════════

def normalize_word(word):
    word = unicodedata.normalize("NFKD", word).lower()
    word = re.sub(r"[^\w']", "", word).strip("'")
    return word


def align_words(asr_words, ref_words):
    if not asr_words or not ref_words:
        return asr_words or []
    asr_norm = [normalize_word(w.word) for w in asr_words]
    ref_norm = [normalize_word(w) for w in ref_words]
    asr_valid = [(i, n) for i, n in enumerate(asr_norm) if n]
    ref_valid = [(i, n) for i, n in enumerate(ref_norm) if n]
    if not asr_valid or not ref_valid:
        return asr_words
    asr_indices, asr_normalized = zip(*asr_valid)
    ref_indices, ref_normalized = zip(*ref_valid)
    ref_set = set(ref_normalized)
    asr_for_matching = []
    for aw in asr_normalized:
        if aw in ref_set:
            asr_for_matching.append(aw)
        else:
            best_match, best_ratio = aw, 0.0
            for rw in ref_set:
                if abs(len(aw) - len(rw)) > max(len(aw), len(rw)) * 0.4:
                    continue
                ratio = SequenceMatcher(None, aw, rw).ratio()
                if ratio > best_ratio:
                    best_ratio, best_match = ratio, rw
            asr_for_matching.append(best_match if best_ratio >= 0.75 else aw)
    sm = SequenceMatcher(None, asr_for_matching, list(ref_normalized), autojunk=False)
    opcodes = sm.get_opcodes()
    result = [TimedWord(word=w, start=0.0, end=0.0, confidence=0.0) for w in ref_words]
    for tag, i1, i2, j1, j2 in opcodes:
        if tag == "equal":
            for asr_pos, ref_pos in zip(range(i1, i2), range(j1, j2)):
                orig_asr, orig_ref = asr_indices[asr_pos], ref_indices[ref_pos]
                result[orig_ref] = TimedWord(word=ref_words[orig_ref], start=asr_words[orig_asr].start, end=asr_words[orig_asr].end, confidence=asr_words[orig_asr].confidence)
        elif tag == "replace":
            t_start = asr_words[asr_indices[i1]].start
            t_end = asr_words[asr_indices[i2 - 1]].end
            n_ref = j2 - j1
            duration = t_end - t_start
            for k, ref_pos in enumerate(range(j1, j2)):
                orig_ref = ref_indices[ref_pos]
                result[orig_ref] = TimedWord(word=ref_words[orig_ref], start=t_start + k * duration / n_ref, end=t_start + (k + 1) * duration / n_ref, confidence=0.5)
    _fill_gaps(result)
    return result


def _fill_gaps(words):
    for i in range(len(words)):
        if words[i].start > 0 or words[i].end > 0:
            continue
        prev_end, next_start = 0.0, None
        for j in range(i - 1, -1, -1):
            if words[j].end > 0:
                prev_end = words[j].end
                break
        for j in range(i + 1, len(words)):
            if words[j].start > 0:
                next_start = words[j].start
                break
        if next_start is None:
            next_start = prev_end + 0.3
        gap_count = 0
        for j in range(i, len(words)):
            if words[j].start > 0 or words[j].end > 0:
                break
            gap_count += 1
        position = 0
        for j in range(i - 1, -1, -1):
            if words[j].start > 0 or words[j].end > 0:
                break
            position += 1
        t_per_word = (next_start - prev_end) / max(gap_count, 1)
        words[i].start = prev_end + position * t_per_word
        words[i].end = prev_end + (position + 1) * t_per_word
        words[i].confidence = 0.3


# ═══════════════════════════════════════════════════════════════════════════════
# STEP 6: TIMING REFINEMENT
# ═══════════════════════════════════════════════════════════════════════════════

def refine_timings(vocals_44k, words, sr=44100, hop_length=256):
    if len(vocals_44k) == 0 or not words:
        return words
    odf = librosa.onset.onset_strength(y=vocals_44k, sr=sr, hop_length=hop_length, n_fft=1024, fmin=80.0, fmax=4000.0, aggregate=np.median, detrend=True)
    onset_frames = librosa.onset.onset_detect(onset_envelope=odf, sr=sr, hop_length=hop_length, backtrack=True, units='frames', pre_max=2, post_max=2, pre_avg=2, post_avg=4, delta=0.05, wait=8)
    rms = librosa.feature.rms(y=vocals_44k, frame_length=1024, hop_length=hop_length)[0]
    rms_smooth = np.convolve(rms, np.ones(7) / 7, mode='same')
    search_frames = int(0.08 * sr / hop_length)
    refined = []
    for word in words:
        w = TimedWord(word=word.word, start=word.start, end=word.end, confidence=word.confidence)
        approx_frame = librosa.time_to_frames(w.start, sr=sr, hop_length=hop_length)
        lo, hi = max(0, approx_frame - search_frames), min(len(odf) - 1, approx_frame + search_frames)
        candidates = onset_frames[(onset_frames >= lo) & (onset_frames <= hi)]
        if len(candidates) > 0:
            w.start = librosa.frames_to_time(candidates[np.argmin(np.abs(candidates - approx_frame))], sr=sr, hop_length=hop_length)
        end_frame = librosa.time_to_frames(w.end, sr=sr, hop_length=hop_length)
        end_search = int(0.05 * sr / hop_length)
        elo, ehi = max(0, end_frame - end_search), min(len(rms_smooth) - 1, end_frame + end_search)
        if elo < ehi:
            rms_db = librosa.amplitude_to_db(rms_smooth[elo:ehi + 1] + 1e-10, ref=rms_smooth.max() + 1e-10)
            silent = np.where(rms_db < -40.0)[0]
            if len(silent) > 0:
                w.end = librosa.frames_to_time(elo + silent[0], sr=sr, hop_length=hop_length)
        if w.end <= w.start + 0.03:
            w.end = w.start + 0.03
        refined.append(w)
    for i in range(len(refined) - 1):
        if refined[i].end > refined[i + 1].start:
            mid = (refined[i].end + refined[i + 1].start) / 2
            refined[i] = TimedWord(word=refined[i].word, start=refined[i].start, end=mid, confidence=refined[i].confidence)
            refined[i + 1] = TimedWord(word=refined[i + 1].word, start=mid, end=refined[i + 1].end, confidence=refined[i + 1].confidence)
    return refined


# ═══════════════════════════════════════════════════════════════════════════════
# OUTPUT FORMATTERS
# ═══════════════════════════════════════════════════════════════════════════════

def _lrc_ts(seconds):
    return f"{int(seconds // 60):02d}:{seconds % 60:05.2f}"

def _srt_ts(seconds):
    h, m, s = int(seconds // 3600), int((seconds % 3600) // 60), seconds % 60
    return f"{h:02d}:{m:02d}:{int(s):02d},{int((s % 1) * 1000):03d}"

def format_enhanced_lrc(words, line_gap=1.0):
    if not words: return ""
    lines, current_line = [], []
    for word in words:
        if current_line and word.start - current_line[-1].end > line_gap:
            lines.append(f"[{_lrc_ts(current_line[0].start)}] " + " ".join(f"<{_lrc_ts(w.start)}> {w.word}" for w in current_line) + f" <{_lrc_ts(current_line[-1].end)}>")
            current_line = []
        current_line.append(word)
    if current_line:
        lines.append(f"[{_lrc_ts(current_line[0].start)}] " + " ".join(f"<{_lrc_ts(w.start)}> {w.word}" for w in current_line) + f" <{_lrc_ts(current_line[-1].end)}>")
    return "\n".join(lines)

def format_standard_lrc(words, line_gap=1.0):
    if not words: return ""
    lines, current_line = [], []
    for word in words:
        if current_line and word.start - current_line[-1].end > line_gap:
            lines.append(f"[{_lrc_ts(current_line[0].start)}] " + " ".join(w.word for w in current_line))
            current_line = []
        current_line.append(word)
    if current_line:
        lines.append(f"[{_lrc_ts(current_line[0].start)}] " + " ".join(w.word for w in current_line))
    return "\n".join(lines)

def format_srt(words, line_gap=1.0, max_words=10):
    if not words: return ""
    entries, current_line = [], []
    for word in words:
        if current_line and (word.start - current_line[-1].end > line_gap or len(current_line) >= max_words):
            entries.append(current_line)
            current_line = []
        current_line.append(word)
    if current_line: entries.append(current_line)
    return "\n".join(f"{idx}\n{_srt_ts(lw[0].start)} --> {_srt_ts(lw[-1].end)}\n{' '.join(w.word for w in lw)}\n" for idx, lw in enumerate(entries, 1))

def format_json(words):
    return [{"word": w.word, "start": round(w.start, 3), "end": round(w.end, 3), "confidence": round(w.confidence, 3)} for w in words]

def format_ass(words, line_gap=1.0):
    header = "[Script Info]\nTitle: Lyric Sync\nScriptType: v4.00+\nPlayResX: 1920\nPlayResY: 1080\n\n[V4+ Styles]\nFormat: Name, Fontname, Fontsize, PrimaryColour, SecondaryColour, OutlineColour, BackColour, Bold, Italic, Underline, StrikeOut, ScaleX, ScaleY, Spacing, Angle, BorderStyle, Outline, Shadow, Alignment, MarginL, MarginR, MarginV, Encoding\nStyle: Default,Arial,48,&H00FFFFFF,&H000000FF,&H00000000,&H64000000,-1,0,0,0,100,100,0,0,1,2,1,2,10,10,40,1\n\n[Events]\nFormat: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text\n"
    line_groups, current = [], []
    for word in words:
        if current and word.start - current[-1].end > line_gap:
            line_groups.append(current)
            current = []
        current.append(word)
    if current: line_groups.append(current)
    def ass_ts(s): return f"{int(s//3600)}:{int((s%3600)//60):02d}:{s%60:05.2f}"
    events = []
    for lw in line_groups:
        karaoke = " ".join(f"{{\\kf{int(w.duration * 100)}}}{w.word}" for w in lw)
        events.append(f"Dialogue: 0,{ass_ts(lw[0].start)},{ass_ts(lw[-1].end)},Default,,0,0,0,,{karaoke}")
    return header + "\n".join(events)


# ═══════════════════════════════════════════════════════════════════════════════
# MAIN PIPELINE
# ═══════════════════════════════════════════════════════════════════════════════

@spaces.GPU(duration=300)
def run_pipeline(audio_path: str, artist: str, title: str, progress=gr.Progress()):
    if audio_path is None:
        raise gr.Error("Please upload an audio file.")
    start_time = time.time()
    status_log = []
    def log(msg):
        status_log.append(msg)
        logger.info(msg)

    device = "cuda" if torch.cuda.is_available() else "cpu"
    log(f"🖥️ Using device: {device}")

    progress(0.05, desc="🎵 Separating vocals (Demucs htdemucs)...")
    log("🎵 Separating vocals with Demucs htdemucs...")
    vocals_16k, sr_16k, vocals_44k, sr_44k = separate_vocals(audio_path, device=device)
    log(f"   ✓ Vocals: {len(vocals_16k)/sr_16k:.1f}s extracted")

    progress(0.30, desc="📝 Transcribing vocals (Whisper large-v3)...")
    log("📝 Transcribing with Whisper large-v3 (word-level timestamps)...")
    asr_words = transcribe_vocals(vocals_16k, sr_16k)
    transcript_text = " ".join(w.word for w in asr_words)
    log(f"   ✓ Transcribed: {len(asr_words)} words")

    progress(0.55, desc="🔍 Identifying song...")
    log("🔍 Identifying song...")
    song = identify_song(audio_path, artist, title, transcript_text)
    if song:
        log(f"   ✓ Identified: {song.artist} — {song.title} (via {song.method})")
    else:
        log("   ⚠ Could not identify song — using raw transcription")

    ref_words = None
    if song:
        progress(0.65, desc="📖 Fetching reference lyrics (LRCLIB)...")
        log("📖 Fetching lyrics from LRCLIB...")
        lyrics_result = fetch_lyrics(song)
        if lyrics_result:
            _, ref_words = lyrics_result
            log(f"   ✓ Lyrics: {len(ref_words)} words")
        else:
            log("   ⚠ No lyrics found — using raw transcription")

    if ref_words:
        progress(0.75, desc="🔗 Aligning transcript to reference lyrics...")
        log("🔗 Aligning ASR transcript to reference lyrics...")
        synced_words = align_words(asr_words, ref_words)
        matched = sum(1 for w in synced_words if w.confidence >= 0.8)
        total = len(synced_words)
        log(f"   ✓ Aligned: {matched}/{total} direct matches ({matched/total*100:.0f}%)" if total else "   ✓ Aligned")
    else:
        synced_words = asr_words
        log("   Using raw ASR words (no reference lyrics)")

    progress(0.88, desc="⚙️ Refining timings (onset/offset detection)...")
    log("⚙️ Refining timings via audio analysis (onset detection, librosa)...")
    synced_words = refine_timings(vocals_44k, synced_words, sr=sr_44k)
    log("   ✓ Timing refinement complete")

    progress(0.96, desc="📄 Formatting outputs...")
    lrc_enhanced = format_enhanced_lrc(synced_words)
    lrc_standard = format_standard_lrc(synced_words)
    srt_text = format_srt(synced_words)
    json_data = format_json(synced_words)
    ass_text = format_ass(synced_words)

    def write_tmp(content, suffix):
        f = tempfile.NamedTemporaryFile(suffix=suffix, delete=False, mode="w", encoding="utf-8")
        f.write(content); f.close(); return f.name

    lrc_file = write_tmp(lrc_enhanced, ".lrc")
    srt_file = write_tmp(srt_text, ".srt")
    json_file = write_tmp(json.dumps(json_data, indent=2, ensure_ascii=False), ".json")
    ass_file = write_tmp(ass_text, ".ass")

    elapsed = time.time() - start_time
    log(f"\n✅ Done in {elapsed:.1f}s — {len(synced_words)} words synchronized")
    progress(1.0, desc="✅ Done!")

    return ("\n".join(status_log), lrc_enhanced, lrc_file, lrc_standard, srt_text, srt_file, json_data, json_file, ass_text, ass_file)


# ═══════════════════════════════════════════════════════════════════════════════
# GRADIO UI
# ═══════════════════════════════════════════════════════════════════════════════

with gr.Blocks(title="🎵 Lyric Sync", css=".main-title{text-align:center;margin-bottom:.5em}.subtitle{text-align:center;color:#666;margin-bottom:1.5em}") as demo:
    gr.HTML("<h1 class='main-title'>🎵 Lyric Sync</h1>")
    gr.HTML("<p class='subtitle'>Automatic perfect song lyric acquisition and synchronization</p>")
    with gr.Row():
        with gr.Column(scale=1):
            audio_input = gr.Audio(sources=["upload"], type="filepath", label="Upload Audio")
            gr.Markdown("### Song Metadata (optional)\n*Provide artist & title to skip identification, or leave blank for auto-detection.*")
            artist_input = gr.Textbox(label="Artist", placeholder="e.g. Radiohead")
            title_input = gr.Textbox(label="Song Title", placeholder="e.g. Creep")
            run_btn = gr.Button("▶ Sync Lyrics", variant="primary", size="lg")
            status_output = gr.Textbox(label="Pipeline Status", lines=12, interactive=False, show_copy_button=True)
        with gr.Column(scale=2):
            with gr.Tabs():
                with gr.Tab("📝 Enhanced LRC (word-level)"):
                    lrc_enhanced_output = gr.Textbox(label="Enhanced LRC", lines=20, show_copy_button=True, interactive=False)
                    lrc_download = gr.File(label="⬇ Download .lrc")
                with gr.Tab("📄 Standard LRC (line-level)"):
                    lrc_standard_output = gr.Textbox(label="Standard LRC", lines=20, show_copy_button=True, interactive=False)
                with gr.Tab("🎬 SRT Subtitles"):
                    srt_output = gr.Textbox(label="SRT", lines=20, show_copy_button=True, interactive=False)
                    srt_download = gr.File(label="⬇ Download .srt")
                with gr.Tab("📊 JSON"):
                    json_output = gr.JSON(label="Word-level JSON")
                    json_download = gr.File(label="⬇ Download .json")
                with gr.Tab("🎤 ASS Karaoke"):
                    ass_output = gr.Textbox(label="ASS (karaoke \\k tags)", lines=20, show_copy_button=True, interactive=False)
                    ass_download = gr.File(label="⬇ Download .ass")
    run_btn.click(fn=run_pipeline, inputs=[audio_input, artist_input, title_input], outputs=[status_output, lrc_enhanced_output, lrc_download, lrc_standard_output, srt_output, srt_download, json_output, json_download, ass_output, ass_download])
    gr.Markdown("---\n### How it works\n| Step | What | Technology |\n|------|------|---|\n| 1 | Vocal separation | Demucs htdemucs (~9.2 dB SDR) |\n| 2 | Word-level transcription | Whisper large-v3 |\n| 3 | Song identification | AcoustID / LRCLIB text search |\n| 4 | Lyrics acquisition | LRCLIB (free, no auth) |\n| 5 | Alignment | Sequence alignment (LCS + fuzzy matching) |\n| 6 | Timing refinement | Onset/offset detection (librosa, 5.8ms resolution) |\n\n**Tips:** Provide artist & title for best results. Supported: MP3, WAV, FLAC, OGG, M4A. Processing: ~30-90s.")

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