Swara Studio
Pick a band, write the words, generate via your own Suno or Mureka key. Nothing runs on our servers except a stateless relay to reach the provider - your key and your songs never leave your control.
?How this page works
- Band is optional - pick one to reuse a saved style prompt, or leave it for a one-off song.
- Pick a provider first - it decides what you need next: a key for Suno/Mureka, or a running pod for ACE-Step.
- Suno/Mureka finish in the cloud, usually within a few minutes. ACE-Step needs your own RunPod pod already running.
- Finished songs save to your Library automatically - no separate save step.
Band optional
Ties songs to a reusable style prompt and mastering target - skip this and generate a one-off song without one.
Song
Generate with
Sing or upload a raw take. ACE-Step builds a brand new song locked to its melody and timing, then your actual recording - not a clone, not a re-sing - goes back on top of the new band.
?How this page works
- Record or upload one take - the new song's length locks to it, there's no separate duration setting.
- This needs the Combined RunPod image (Settings → Deploy & manage your RunPod pod) - one pod that both composes the new band and keeps your real voice on it.
- Two steps run behind the scenes on that one pod: it composes a new band around your take, then puts your actual voice back on top of it - not a re-sing, not a clone.
- Takes a few minutes - GPU rendering, then GPU stem separation.
Your recording
A cappella works best - just your voice, no backing track already on it.
Song
Combined pod
This page needs the Combined RunPod image - one pod that composes the new band and keeps your real voice on it. If Create already has a plain ACE-Step pod running, deploy Combined here instead (Settings → Deploy & manage your RunPod pod).
Your artists, saved to your own Google Drive (hidden app-data folder) - not our database. Every song made for a band reuses its style prompt and mastering target.
?How this page works
- A band is a style prompt plus a mastering target (LUFS/crest) - every song made for it reuses both.
- Bands save to a hidden folder in your own Google Drive - sign in with Google first, or changes only last until you close the tab.
- Bands are optional at generation time - skip them on Create for a one-off song.
Your bands
Everything you've generated - stored in the local folder you picked on Settings. Nothing is uploaded anywhere.
?How this page works
- "This computer" is fastest but Chromium-only, and the folder is remembered - on a later visit it reconnects by itself, or with one "Allow" click (pick "Allow on every visit" and it stops asking).
- "Google Drive" is slower per file but works in any browser and follows you to any signed-in device.
- Refresh reconciles the list against what's actually in the folder, so files added another way (like the desktop app) still show up.
Pick a song
Press play on any song above
Style prompt
Pick a song to see its details
Lyrics
Pick a song to see its details
Storage
Where these songs actually live - the same setting as Settings.
Chrome/Edge only. Songs stay on this device.
?How this page works
- Needs the exact columns from the blank template below - build your own sheet from scratch and column matching will likely fail.
- Each row can set its own Duration/Language, or leave them blank to use the defaults in the side panel.
- Rows run in order, unattended - select just the pending ones, or stop mid-run if you need to.
- When it's done, download the updated sheet - re-running it later only touches rows that actually failed.
Tracks
| # | Title | Album / Artist | Status |
|---|
Generate with
Train a voice character on a free Google Colab GPU - no server of ours involved, nothing to install.
?How this page works
- Training runs on a free Google Colab GPU, not on Swara Sound's servers or your own machine - just a free Google account, no install.
- How long it takes depends on which GPU Colab happens to hand you that session - treat it as "leave it running," not instant.
- You need to be signed in with Google for a trained voice to show up here afterward.
The whole flow, 4 steps: upload your recordings to Drive → open the free notebook in Colab and run it (~1-3 hours, nothing runs on our servers) → bring the two result files back → your voice is ready. Steps 1-3 below stay on screen the whole time so you can see where you are.
Step 1 - Upload your recordings
10-15 minutes of clean, solo recordings (no music/background noise) works best - .wav/.mp3/.flac/.ogg/.m4a, any number of files.
Step 2 - Train it on Colab (free, ~1-3 hours)
Once step 1 is uploaded, it's waiting at
My Drive/SwaraStudioVoices/MyVoice_dataset.zip.
Saves train_voice_colab.ipynb to your computer.
colab.research.google.com → Upload notebook → pick the file you just downloaded.
Open colab.research.google.comIn the notebook's settings cell: MODEL_NAME = "MyVoice" - the same name you used in step 1.
Runtime → Change runtime type → T4 GPU, then Runtime → Run all.
Takes ~1-3 hours - the last cell downloads two result files straight to your browser when it's done.
Step 3 - Bring the result back
The notebook's last cell downloads two files to your browser (or grab them
from My Drive/SwaraStudioVoices/<name>_model/ if you closed the tab):
the .index file and one _<N>e.pth file (start with
150e). Pick both here.
Your voices
Sign in with Google to keep your bands and API keys in your own Drive (a hidden app-data folder only this app can see). Pick a folder on this computer for the actual audio files.
Appearance
Scales the whole app's text and controls up or down - saved on this device only.
Google account
Sign in via the account button at the top of the sidebar. This app's OAuth Client ID is set by default - only edit it if you're running your own deployment.
External APIs
Suno (via kie.ai) and Mureka - your own keys, saved to your Drive app-data folder once signed in. Plain paid HTTPS calls; this app relays the request so the browser's CORS rules don't get in the way, but never stores your key server-side.
ACE-Step (your RunPod)
Where a song generated with ACE-Step actually runs: a RunPod GPU pod you deploy and pay for yourself (RunPod bills for GPU time, not a fixed price per song). Use the card below to have this app deploy and manage that pod for you - it fills in the Pod URL here automatically. Only fill this in by hand if you'd rather run and manage the pod yourself (RunPod console, or the desktop app). Only using Vocal to Song? Pick the Combined image below instead - it also runs that page's processing step on this same pod, at no extra idle-GPU cost. See the Vocal to Song guide.
Deploy & manage your RunPod pod
Needs a RunPod account API key (different from the pod's own optional key above) from your RunPod account (Settings → API Keys).
Pod settings
Good habit: never Start a stopped pod - on RunPod Secure Cloud it can lose its GPU to someone else and then refuse to restart. Press Redeploy to begin a session (it terminates any pod at the URL above and finds a free GPU fresh every time).
Applies after every ACE-Step Create or Batch run (not Suno/Mureka, which never touch the pod) - so a forgotten pod doesn't keep billing. "Terminate" is the safest choice if you're not watching closely; you'll need to Redeploy before the next ACE-Step run either way.
Vocal to Song processing
Filled in automatically when you deploy the Combined image above - same
pod as ACE-Step, port 8001, its own API key generated fresh each redeploy. Only edit these by
hand if you're deliberately running the worker on a genuinely separate pod (see
vocal_worker/README.md in the repo for that manual setup).
Song library storage
Where the actual generated audio is kept. Bands and API keys always live in your Drive app-data folder either way - this only affects the songs themselves.
Chrome/Edge only - the File System Access API isn't available in Firefox/Safari. Songs stay on this device only.
Help
There's no support desk behind this app - the FAQ & troubleshooting page is it, and it's written to actually answer things. The Privacy & Terms page covers exactly what does (and doesn't) leave your own Drive and computer.