TTS in your own voice
Read one fixed sentence into your microphone and type anything: it comes back spoken in your voice. Open-source zero-shot voice cloning (OmniVoice) running on a single RTX 3070.
Checking…
1Read this sentence out loud
This is my own voice, and I agree to let it be used for speech synthesis today.
A quiet room and your normal speaking voice give the closest match
2Type what it should say
Ground rules
- Use only your own voice. There is no file upload: the next step opens only after you read the consent sentence into the mic
- The result is synthetic speech. Do not use it to deceive anyone or to pass as someone else
- Your recording sits in a temporary folder on the server for 30 minutes and is then deleted. It is never used for training, and the text you type is not stored
- Rate limits count a hash of your IP, never the IP itself: 10 recordings and 40 generations an hour
Why OmniVoice
Five open-source models under licenses that allow commercial use (Apache 2.0, MIT) ran on the same RTX 3070. Each cloned the same reference voice saying the same Korean sentence; Whisper transcribed the result for pronunciation errors, and a speaker-verification model (ECAPA) scored how close it sounded to the reference.
| Model | Char. errors | Similarity | Time for 5 s | GPU memory |
|---|---|---|---|---|
| OmniVoice | 0% | 0.812 | 2.2 s | 2.2 GB |
| Fun-CosyVoice3 0.5B | 3.1% | 0.808 | 7.1 s | 3.6 GB |
| VoxCPM2 | 0% | 0.783 | 17 s | 5.9 GB |
| Chatterbox Multilingual | 0% | 0.749 | 11 s | 3.6 GB |
| GPT-SoVITS v2ProPlus | 0% | 0.687 | 7.5 s | 1.9 GB |
A small test: one sentence and one reference voice, a stock TTS voice rather than a real person. In recent papers (MiniMax multilingual test), OmniVoice and VoxCPM2 also sit at the top of open models for Korean speaker similarity, around 0.83. OmniVoice · multilingual comparison