README.md

metadata

title: Sangeet Studio Dev
emoji: 🎵
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.2.0
python_version: '3.12'
app_file: app.py
license: gpl-3.0
suggested_hardware: l4x1
short_description: Suno-style song generation + instant voice cloning
tags:
  - music-generation
  - voice-cloning
  - ace-step
  - seed-vc
  - indian-music

🎵 Sangeet Studio

Suno-style full-song generation with vocals, finetuned toward Indian/Pakistani music styles, plus instant artist-voice cloning — all open-source, chained in one app:

Stage	Model	License
Song generation (lyrics + style → full song, vocals + instruments)	ACE-Step v1.5 turbo (2.4B DiT, 8-step turbo, LM CoT planner)	MIT
Indian music styles (qawwali, ghazal, Bollywood, Hindustani, Sufi rock…)	RagaLoRA (adapter on the turbo DiT)	Apache-2.0
Stem split (vocals / instrumental)	Demucs htdemucs	MIT
Voice clone → song vocals (SVC, zero-shot from 5–25 s reference)	Seed-VC (f0-conditioned 44 kHz checkpoint)	GPL-3.0

⚙️ Hardware

This app requires a GPU. In Space Settings → Hardware, pick:

• L4 (24 GB) — recommended, comfortable headroom for the whole chain.
• A10G (24 GB) also works. T4 (16 GB) is tight but should fit (models are loaded sequentially).
• ZeroGPU is not recommended: models must be re-loaded on every request and long conversions hit duration limits.

First boot downloads ~10 GB of checkpoints (ACE-Step main model + LM) — allow 5–10 minutes before the first generation. First generation also loads Seed-VC and Demucs weights.

🎤 Languages

Vocal language is passed to ACE-Step as an ISO 639-1 code: Hindi (hi), Urdu (ur), Punjabi (pa), Tamil (ta), Telugu (te), Bengali (bn), Marathi (mr), Gujarati (gu), Kannada (kn), Malayalam (ml), Odia (or), Assamese (as), Nepali (ne), Sinhala (si), plus English and auto-detect.

Hindi and English are the strongest; write lyrics in the native script (Devanagari, Shahmukhi, Gurmukhi, Tamil…) — transliteration noticeably degrades pronunciation. Long-tail languages are supported but rougher. For Indian styles keep RagaLoRA enabled (strength ~0.8).

🧬 Voice cloning

Upload a clean 5–25 s clip of the target artist's voice (a cappella or dry vocals work best). The vocals stem is converted with Seed-VC SVC (30 diffusion steps, auto-F0 adjust on) and remixed with the instrumental stem. "Pitch shift" corrects big register mismatches (e.g. −5 for a male clone singing a female melody).

🎙 Live Rap Studio

Record your rap directly in the browser (mic) or upload a dry vocal. The app estimates the recording's tempo, generates a matching rap instrumental in the same style, keeps your own voice as the artist (or converts it to a cloned artist voice via Seed-VC), then polishes and masters the mix: high-pass filtering, silence trimming, gain staging, soft limiting and fades.

📜 Licenses & credits

• ACE-Step 1.5 — MIT (repo, paper)
• RagaLoRA — Apache-2.0
• Seed-VC — inference code adapted from app_svc.py, GPL-3.0. Because of this, this Space is distributed under GPL-3.0.
• Demucs — MIT.

Please respect the ACE-Step usage principles: verify originality of generated works, disclose AI involvement, and obtain permission when cloning real artists' voices — cloning someone's voice without consent is not an intended use.