project
ViiTorVoice - An AI-powered voice synthesis model developed by Cloud Curvature
ViiTorVoice, launched by Cloudwise, is the world's first AI speech synthesis model supporting partial editing, topping the Seed-TTS authoritative evaluation list. It employs a NAR non-autoregressive architecture, enabling speech editing as easy as editing a Word document...
What is ViiTorVoice?
ViiTorVoice, launched by Cloudwise, is the world's first AI speech synthesis model supporting partial editing, topping the Seed-TTS authoritative evaluation list. It employs a NAR non-autoregressive architecture, enabling "speech editing like editing Word"—it can selectively replace a word or segment in the audio while maintaining the rest of the timbre, rhythm, and emotion completely unchanged. It supports text-free cross-language cloning, precise emotion control, and ultra-low latency inference of 60ms. The 1B parameter model is open source.
ViiTorVoice's main functions
-
Fragment-level local editing: You can specify to replace a word or segment in the audio while leaving the rest completely unchanged.
-
Voice cloningSupports zero-shot cross-language cloning; simply upload plain audio to extract the timbre.
-
Emotion and Paralinguistic ControlBy inserting special tokens, precise word-level control is achieved, and CFG technology is introduced to enhance emotional expression.
-
Low-latency inferenceEnd-to-end first frame generation time is less than 60ms, and first block inference is supported.
-
Multilingual supportIt can generate content in multiple languages, including Chinese, English, Japanese, and Korean.
Follow us on WeChat and reply with "open source",join inAI open source project discussion group
How to use ViiTorVoice
-
Visit the Demo pageOpen the Hugging Face online experience address (https://huggingface.co/spaces/ZzWater/ViiTorVoice).
-
Select function modeSwitch between the "Voice Edit" and "Voice Clone" tabs at the top of the interface.
-
Upload source audioDrag the original audio file that you want to edit or clone into the "Source Audio" upload area on the left.
-
Input text commandsEnter the original transcribed text and the modified target text in the edit box, and the system will automatically locate the difference fragments.
-
Insert emotion tagTo control emotions, insert a special token into the text (e.g.,
<laugh>,<sigh>Mark sub-language information. -
Adjust generation parametersSet the inference steps (4 or 8 steps) and other configurations, then click the generate button and wait for the model to process.
-
Preview and DownloadPlay the audio in the "Edited Audio" area on the right, and download and save it to your local device after confirming that it is correct.
ViiTorVoice's core advantages
-
The world's number one in evaluation resultsSeed-TTS has an English word error rate of 1.32 and a Chinese word error rate of 0.99, leading both SIM-o and UTMOS.
-
Unique local editing capabilitiesIt is the only technology in the industry that supports segment-level targeted editing, solving the pain point of traditional TTS requiring the entire segment to be re-recorded when changing words.
-
Text-free dependency cloningIt can achieve cross-language timbre transfer without the need for accurate text transcription, and is suitable for minority language scenarios.
-
Speed ReasoningThe first frame latency is less than 60ms, which is far better than the industry average of 150-200ms.
ViiTorVoice project address
- GitHub repositoryhttps://github.com/viitor-ai/viitor-voice-nar
- HuggingFace model libraryhttps://huggingface.co/ZzWater/ViiTorVoice-NAR
Comparison of ViiTorVoice with similar products
| Comparison Dimensions | ViiTorVoice | Qwen3-TTS |
|---|---|---|
| Research and development | Cloud Curvature | Alibaba Tongyi Lab |
| Core Architecture | NAR Non-Autoregressive(Cloze test) | AR Self-Regression(Frame-by-frame chain generation) |
| Partial editing | Supports fragment-level targeted modificationChange words but not the whole paragraph | Not supportedModifying a single word requires regenerating the entire paragraph. |
| Seed-TTS English WER | 1.32(Lower = More accurate) | 1.54 |
| Seed-TTS Chinese WER | 0.99(Industry's first <1.0) | 1.15 |
| First frame delay | <60ms | ~150-200ms |
| No text clone | Cross-language cloning with only pure audio | Audio and accurate transcribed text are required. |
| Emotional control | Precise word-level token control (laughter/sigh, etc.) | Basic Natural Language Description Control |
Application scenarios of ViiTorVoice
- Film and television post-productionDialogue adjustments do not require recalling actors to the recording studio; directors can replace lines in real time on a non-linear timeline, with timbre and breathing rhythm remaining seamlessly consistent.
- Audiobooks and podcastsWhen a slip of the tongue or a mistake in proper nouns occurs during recording, only one or two seconds of audio need to be modified, without having to re-record the entire chapter. Post-production audio editing time has been reduced from several days to tens of minutes.
- Short drama localization for overseas marketsThere's no need to reassemble voice actors; simply replace specific words in the original recording to produce multilingual versions, each sounding like the original performance.
- Advertising and MarketingWhen temporarily adjusting product names or slogans, directly replace the audio clips in a partial way to maintain the original brand tone and emotional appeal.
- Game voice actingCross-language character voice cloning: Simply upload the character's original audio to generate voiceovers in multiple languages, including Chinese, English, Japanese, and Korean.