ഒരു പാട്ട് ഉണ്ടാക്കാം
Describe a song in Malayalam or English. RaagaAI writes lyrics, composes it in a raga, and mixes it on this device.
Tempo, key, raga and instruments
Each singer is a different synthetic voice — tract length, breathiness and vibrato all differ.
Instruments (leave all off for the style's own band)
Use my own lyrics
Tag parts with [Verse], [Pre-chorus], [Chorus] and [Bridge]. The song follows your lyrics, so Length is ignored.
No song is open. Create one, or open a song from your library.
Lyrics
How this song was made
Library
Songs you create are saved here on this device.
Settings
AI lyric writer
Without a provider, RaagaAI writes draft lyrics offline in Malayalam and English. Tamil, Hindi, Telugu and Kannada lyrics need a provider. Keys are stored only in this browser.
Model names change often. If a test fails with an unknown model, check your provider's current model list.
Singing voice
The built-in synthesizer always works offline. The other two sing with a real recorded-voice model and need no API key: RaagaAI speaks each syllable, then puts it on the melody in tune and in time.
Two real Malayalam voices from the Piper project, trained on the IIT-Madras Indic TTS corpus: Meera (female) and Arjun (male), so a duet uses two different singers. Run tools/fetch-voices.sh once to put them in voices/ (61 MB each). RaagaAI converts your Malayalam text to phonemes itself, so no espeak install is needed.
First use downloads a speech model (roughly 40–70 MB) from Hugging Face and keeps it cached. Models are MMS-TTS, licensed CC BY-NC 4.0 — personal, non-commercial use. One speaker per language, so a duet is the same voice at two pitches.
Any server that returns WAV or MP3 audio works. Run the page from http://localhost if your server is plain HTTP.
Real vocals with ElevenLabs
Optional and paid. Sends the song description and lyrics to ElevenLabs Eleven Music and saves the returned audio. Malayalam pronunciation isn't guaranteed. If your browser blocks the request, it will say so here.
Audio
Film score keeps the quiet parts quiet: light compression, gentle saturation, more air. Balanced is the usual streaming master.
What runs here, and what doesn't
Everything is generated on this device: prompt reading, draft lyrics, composition, instrument synthesis, mixing and loudness mastering measured to ITU BS.1770.
The voice is a synthesized guide that sings the melody on the vowels of your lyrics. It is not a trained singing model and it won't pronounce words clearly. Export MIDI with embedded lyrics to record a real singer, or use ElevenLabs above.
Not included: voice cloning, music video, collaboration, FLAC export, and cloud accounts. The content filter is a basic keyword check, not a trained moderation model.
You are responsible for how you use songs you make. Generated lyrics are drafts; check them before publishing.