README.md

metadata

title: Karaoke Generator
emoji: 🎤
colorFrom: red
colorTo: purple
sdk: gradio
sdk_version: 6.19.0
python_version: '3.13'
app_file: app.py
pinned: false
license: mit

Karaoke Generator

Generate karaoke videos from any audio file with synced, styled lyrics.

Workflow

Step 1 — Transcribe

Upload your audio file. The app runs faster-whisper (int8 quantized, 4–8× faster than standard Whisper on CPU) to produce a word-timestamped JSON file you can download and edit.

For Gujarati audio, selecting Language = Gujarati + Model = small automatically uses the fine-tuned vasista22/whisper-gujarati-small model — trained on 430+ hours of Gujarati speech data for much better accuracy.

Edit the JSON externally (paste into Gemini/ChatGPT to fix wrong words, or edit timestamps manually). The format is easy to work with:

[
  {
    "start": 1.23,
    "end": 3.45,
    "text": "He Karunanidhi Swami",
    "words": [
      {"word": "He",           "start": 1.23, "end": 1.50},
      {"word": "Karunanidhi", "start": 1.50, "end": 2.20},
      {"word": "Swami",       "start": 2.20, "end": 3.45}
    ]
  }
]

Step 2 — Generate Video

Paste your refined JSON, upload audio + optional background image, choose styling (font, colors, outline), and click Generate. The app renders a 1280×720 MP4 with:

• Left-to-right color sweep on each word as it's sung (\kf ASS karaoke tag)
• Your chosen background image (scaled/cropped to fill)
• Custom font, font size, and three independent color pickers

Vocal Removal Options

Method	Speed	Notes
Skip	Instant	Lyrics over original audio
Fast (ffmpeg)	~5 sec	Phase-cancellation, stereo files only
AI (MDX-Net)	5–10 min	Best quality, CPU only on free tier

Tech Stack

• faster-whisper — transcription
• vasista22/whisper-gujarati-small — Gujarati fine-tune
• audio-separator — MDX-Net vocal removal
• ffmpeg — video rendering with ASS karaoke subtitles
• Gradio — UI