app.py

import os
# 1. Force PyTorch to allow loading "unsafe" weights (The VAD models require this)
os.environ["TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD"] = "1"

import gradio as gr
import whisperx
import json
import torch
import gc

# --- NEW HELPER FUNCTION ---
def format_timestamp(ts):
    """Convert seconds (float) to mm:ss format"""
    minutes = int(ts // 60)
    seconds = int(ts % 60)
    return f"{minutes}:{seconds:02}"
# ---------------------------

# 2. Global Patch for torch.load (Backup fix for libraries that hardcode parameters)
_original_load = torch.load
def patched_load(*args, **kwargs):
    if 'weights_only' in kwargs:
        kwargs['weights_only'] = False
    return _original_load(*args, **kwargs)
torch.load = patched_load

# 1. Setup Device & Config
device = "cuda" if torch.cuda.is_available() else "cpu"
batch_size = 16 
compute_type = "float16" if device == "cuda" else "int8" # int8 is faster on CPU
DEFAULT_XLSR = "facebook/wav2vec2-large-xlsr-53"
ALIGN_MODEL_MAP = {
    # =====================
    # Built-in WhisperX aligner (no explicit model needed)
    # =====================
    "en": None,     # English
    "tl": None,     # Tagalog / Filipino
    "es": None,     # Spanish
    "fr": None,     # French
    "de": None,     # German
    "it": None,     # Italian
    "pt": None,     # Portuguese
    "ru": None,     # Russian
    "nl": None,     # Dutch

    # =====================
    # Explicit wav2vec2 models (needed)
    # =====================

    # Chinese
    "zh": "jonatasgrosman/wav2vec2-large-xlsr-53-chinese-zh-cn",

    # Japanese
    "ja": "jonatasgrosman/wav2vec2-large-xlsr-53-japanese",

    # Korean
    "ko": "kresnik/wav2vec2-large-xlsr-korean",

    # Thai
    "th": "airesearch/wav2vec2-large-xlsr-53-th",

    # Vietnamese
    "vi": "nguyenvulebinh/wav2vec2-large-xlsr-53-vietnamese",

    # Indonesian / Malay
    "id": "indonesian-nlp/wav2vec2-large-xlsr-indonesian",
    "ms": "indonesian-nlp/wav2vec2-large-xlsr-indonesian",

    # Arabic
    "ar": "elgeish/wav2vec2-large-xlsr-53-arabic",

    # Hindi
    "hi": "theainerd/Wav2Vec2-large-xlsr-hindi",

    # Turkish
    "tr": "savasy/wav2vec2-large-xlsr-turkish",

    # Polish
    "pl": "jonatasgrosman/wav2vec2-large-xlsr-53-polish",

    # Ukrainian
    "uk": "jonatasgrosman/wav2vec2-large-xlsr-53-ukrainian",

}

print(f"Loading WhisperX model on {device}...")
model = whisperx.load_model("medium", device, compute_type=compute_type)

def generate_lyrics(audio_file_path):
    if audio_file_path is None:
        return {"error": "No audio file provided"}

    try:
        audio = whisperx.load_audio(audio_file_path)
        result = model.transcribe(audio, batch_size=batch_size)

        lang = result["language"].lower()
        if lang.startswith("zh"): lang = "zh"
        
        align_model_name = ALIGN_MODEL_MAP.get(lang, DEFAULT_XLSR)

        try:
            if align_model_name is None:
                model_a, metadata = whisperx.load_align_model(language_code=lang, device=device)
            else:
                model_a, metadata = whisperx.load_align_model(language_code=lang, device=device, model_name=align_model_name)

            result = whisperx.align(result["segments"], model_a, metadata, audio, device, return_char_alignments=False)
            del model_a
        except Exception as align_err:
            print(f"[WARN] Alignment skipped: {align_err}")

        # --- UPDATED FORMATTING ---
        formatted_lyrics = [
            {
                "time": format_timestamp(seg["start"]), # Changed from round(seg["start"], 3)
                "text": seg["text"].strip(),
                "chords": []
            }
            for seg in result["segments"]
        ]
        # --------------------------

        gc.collect()
        if device == "cuda":
            torch.cuda.empty_cache()

        return {"lyrics": formatted_lyrics}

    except Exception as e:
        return {"error": str(e)}

demo = gr.Interface(
    fn=generate_lyrics,
    inputs=gr.Audio(type="filepath", label="Upload Vocals/Audio"),
    outputs=gr.JSON(label="JSON Result"),
    title="WhisperX Aligned Lyric Generator",
    description="Transcribes audio and provides mm:ss timestamps."
)

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