The short version
TTS, text to speech, starts from text. You type a sentence and a voice model speaks it. RVC, Retrieval-based Voice Conversion, starts from audio. You give it a recording and it returns the same performance in a different voice.
That one difference decides almost everything else: what you need to feed each one, what they are good at, and why they work so well together.
What TTS is good at
Text to speech is the fastest way to get spoken audio when nobody is around to record it. Scripts, narration, notifications, placeholder dialogue: type it, generate it, done. Modern neural voices sound natural, keep a steady pace and never get tired of retakes.
Its limit is the voice list. You get the voices that come with the engine, and making a new one usually means a lot of data and training. Most text to speech voices also cannot sing; they are built for speech.
What RVC is good at
RVC is about the voice, not the words. Because it keeps the timing, pitch and phrasing of the input, it can turn a sung vocal into another singer or a spoken take into another speaker, with the performance intact. That is why it took over AI covers and demo vocals.
RVC models are also easy to make. A clean ten minutes of someone's voice can be enough to train a usable model, and there are many shared models in the community. Its limit is the mirror image of text to speech: it cannot invent a performance. Somebody, or something, has to provide the audio.
Side by side
Input: text to speech takes text, RVC takes audio. Output: both give you audio, but text to speech gives a voice from its own list while RVC gives the voice of whatever model you load.
Performance: text to speech decides the pacing and intonation itself. RVC keeps the pacing, pitch and emotion of the audio you feed it, which is why a great input performance matters so much more for RVC.
Singing: most text to speech voices are built for speech only. RVC handles singing well, as long as the model was trained on enough sung material.
Making new voices: building a new text to speech voice is a big job. Training a new RVC model takes a clean set of recordings and some time on your computer, which is why there are so many RVC voices around.
Common mistakes when combining them
Mismatched base voice. Feeding a deep male text to speech voice into a high female RVC model forces the model to stretch, and the result can sound strained. Start from a base voice in the same range as the target.
Writing for the page instead of the ear. Long sentences with no commas come out breathless. Write short lines and punctuate where a person would pause, because RVC keeps whatever pacing the text to speech step gives it.
Ignoring the index file. If your model came with a .index file, use it. It helps the converted voice keep the target's character instead of drifting toward the base voice.
Why chaining them works
Put text to speech first and RVC second and you get the best of both: text goes in, speech in any RVC voice comes out. The text to speech step supplies the performance, and the RVC step supplies the voice. That turns every RVC model you have into a voice that can read.
The catch has always been setup, since the two usually live in different tools. Xport Studio runs the chain in one step: type a line, pick an RVC model from your Library and press Generate. It reads the line with a built-in neural voice matched to your model's gender, converts it, and gives you a WAV, all offline on your computer. You can hear examples on the RVC text to speech page.
Which one should you use?
Use text to speech alone when any good voice will do and you just need words spoken. Our free in-browser text to speech tool is enough for quick narration and tests.
Use RVC alone when you already have a performance, sung or spoken, and want it in another voice. That is the vocal converter job: covers, demo vocals, swapping a guide vocal.
Use both when you need a specific voice to say something nobody recorded: voiceovers in a custom voice, intros and drops, character lines, or your own cloned voice reading a script.