> ## Documentation Index > Fetch the complete documentation index at: https://docs.rawchunks.space/llms.txt > Use this file to discover all available pages before exploring further. # Use a custom voice > Upload a vocal reference, save a song's voice, and reuse a voice when its mode is ready. Open Studio → **GLOBAL PARAMETERS** → **VOICE**. **CUSTOM VOCAL** offers an uploaded reference sample or a trained voice from **SAVED VOICES**. ## Use a vocal reference 1. Select **UPLOAD SAMPLE** and choose or drop an audio file. The dialog lists WAV, MP3, FLAC, OGG, and WebM. 2. Review the filename and size. Use **REPLACE** or **REMOVE** to change it. 3. Set **VOCAL RESEMBLANCE**, then **DONE**. 4. Confirm your generation. The app uploads and saves the sample at this point, then uses it as a reference for the track. Custom Vocal dialog with Upload Sample and an audio drop area listing supported formats. Choose an isolated **mono or stereo** vocal recording of no more than **four minutes** if you want it prepared for trained use. Uploads must be under **50 MiB**. A sample can be saved before preparation finishes; a long recording can save and still fail preparation. Selecting a file alone does not upload it. The upload and generation have separate [service charges](/app/payments#service-prices). ## Save vocals from a generated song On a **READY** Studio variation, select **SAVE VOICE**. Rawchunks keeps a voice profile named from the song and starts preparing its vocals for reuse. Check it in **MY-PAGE → MY VOICES**, or reopen **VOICE → SAVED VOICES**. You may see **SAVED**, **PREPARING**, **READY**, or **ERROR**, with separate **Fast** and **Master** status. ## Reuse a trained voice Choose **FAST** or **STUDIO MASTER** in Studio first. Open **SAVED VOICES**, select a voice marked **Ready** for that mode, then **DONE**. Unready voices are disabled for the selected mode. Saved voice selected in Custom Vocal, with Fast Ready, Master Ready and a Vocal Resemblance slider. Preparation trains a small custom voice adaptation, called **LoRA**, for each generation mode. You don't set training parameters or start a separate training job in the app. A reference sample can be usable before the trained voice is ready. ## Adjust resemblance and check the result **VOCAL RESEMBLANCE** covers **10–100%**, starting at **65%**. It controls how strongly you request the voice's character. Even at its highest setting, listen to the result before treating it as a match. In **MY VOICES**, **PLAY SAMPLE** plays the saved reference. **RETRY** refreshes voice status after a loading error; it does not restart training. Use **REMOVE VOICE** in the Studio dialog to clear the generation selection. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.