ViiTorVoice-NAR Real-World Test - Open-Source and Free Voice Cloning Tool

Today I'm sharing some things suitable for sound reproduction.

To be honest, doing it now AI The most annoying thing about voice editing is that you only want to change a few words in the generated audio, but you have to re-record, re-dialogue, and re-edit it.

Today's ViiTorVoice-NAR project on GitHub is perfect for this scenario: given an original audio clip and corresponding text, you can directly modify certain parts of it while preserving the original feel of the voice, tone, and rhythm.

ViiTorVoice-NAR is viitor-ai open sourceThe non-autoregressive speech generation system is primarily designed for speech cloning, local speech editing, emotion control, and paralinguistic control.

Project address: https://github.com/viitor-ai/viitor-voice-nar

The core idea of ViiTorVoice-NAR is to change speech generation from a token-by-token chain to a discrete audio token filling method.

Traditional autoregressive TTS often generates audio tokens sequentially, and the next token can only be generated after the previous one is completed; ViiTorVoice-NAR uses a masked language model to complete the masked audio segments in a discrete codebook space.

This structure is well-suited for partial speech editing. For example, changing "Friday" to "Monday" in a sentence only requires locating and recombining the modified local area, without having to regenerate the entire audio segment.

ViiTorVoice-NAR's core capabilities:

• Voice cloning: Input a prompt audio, the model generates the corresponding speech of the target text, and tries to preserve the speaker's timbre in the prompt audio.
• Local audio editing: Input the original audio, the original text, and the edited complete text. The system will first perform a diff on the text, then combine the forced alignment results to find the audio area that needs to be replaced, and finally only reconstitute that small segment of audio.
• Emotion and paralinguistic control: Emotional or paralinguistic tags, such as mood, laughter, pauses, tone, etc., can be inserted into text conditions.
• Low-latency inference: The project supports first-block inference mode. The model can generate the first audio token first to shorten the first frame latency.

The project requires Git, UV, and an NVIDIA graphics card. 8GB of VRAM is sufficient, but 12GB would provide smoother performance.

Once the environment is set up, simply clone the project to an English path:

1 cd D:\ 2 git clone https://github.com/viitor-ai/viitor-voice-nar.git viitor-voice-nar 3 cd D:\viitor-voice-nar

Then create a Python 3.12 virtual environment and install the dependencies:

1 uv venv --python 3.12 2 uv pip install --python .venv\Scripts\python.exe -r requirements-grpc.txt 3 uv pip install --python .venv\Scripts\python.exe -r requirements-alone.txt 4 uv pip install --python .venv\Scripts\python.exe protobuf==4.25.3

Download the model; it's over 10 GB in size, so please allow enough space.

1 New-Item -ItemType Directory -Force local_models 2 $env:PYTHONIOENCODING="utf-8" 3 .venv\Scripts\hf.exe download ZzWater/ViiTorVoice-NAR --local-dir local_models

Finally, you can start using it after launching it:

1 cd D:\viitor-voice-nar 2 .\run_grpc_v2.ps1 start all -- --no-warmup

Case 1: Voice Cloning

1 Copy-Item "被克隆的语音位置" "D:\viitor-voice-nar\input_voice_clone.mp3" -Force
2 
3 curl.exe -X POST http://127.0.0.1:7861/v1/voice-clone `
4 -F "ref_audio=@D:\viitor-voice-nar\input_voice_clone.mp3" `
5 -F "text=你好，这是一次语音克隆测试。" `
6 -F "language=zh" `
7 -F "allow_missing_ref_text=true" `
8 -F "output_format=wav" `
9 -F "num_steps=32" `
10 --output "语音克隆输出位置" `
11 --write-out "HTTP %{http_code} size %{size_download}`n"

Original version:

It can be seen that the clone not only has a similar timbre, but also has a bit of a rap flavor.

Case 2: Partial word modification

1 Copy-Item "被克隆的语音位置""D:\viitor-voice-nar\input_liziming.mp3" -Force 2 3 curl.exe -X POST http://127.0.0.1:7861/v1/text-local-edit ` 4 -F "source_audio=@D:\viitor-voice-nar\input_liziming.mp3" ` 5 -F "original_text=请注意三年六班李子明，李子明同学，你妈妈拿了两罐旺仔牛奶要给你。" ` 6 -F "edited_text=请注意三年六班克劳德，克劳德同学，你妈妈拿了两份迪普斯科要给你。" ` 7 -F "language=zh" ` 8 -F "input_format=wav" ` 9 -F "output_format=wav" ` 10 -F "align_granularity=word" ` 11 -F "expand_mask_ratio=1.5" ` 12 -F "num_steps=32" ` 13 --output "语音克隆输出的位置" ` 14 --write-out "HTTP %{http_code} size %{size_download}`n"

Original version:

Apart from the parts that needed modification, there were basically no changes elsewhere.

Case 3 Emotional Control

1 Copy-Item "被克隆的语音位置""D:\viitor-voice-nar\input_sad_ref.wav" -Force 2 3 $Text = "你好，这是一次悲伤情绪的语音生成测试。" 4 5 curl.exe -X POST http://127.0.0.1:7861/v1/voice-clone ` 6 -F "ref_audio=@D:\viitor-voice-nar\input_sad_ref.wav" ` 7 --form-string"text=<|emotion-sad|>$Text" ` 8 -F "language=zh" ` 9 -F "allow_missing_ref_text=true" ` 10 -F "emotion_guidance_scale=6.0" ` 11 -F "nvv_guidance_scale=2.0" ` 12 -F "output_format=wav" ` 13 -F "num_steps=32" ` 14 --output "语音克隆输出的位置" ` 15 --write-out "HTTP %{http_code} size %{size_download}`n"

Original version:

The emotion was spot-on, and the timbre didn't change with the emotional shift.

AI Voiceover is now entering the most time-consuming and expensive part of content production: editing.

Monotype surveyed 1,008 creative professionals in 2025 and found that 57% of creative teams spend more than a quarter of their time each week on non-creative tasks, including asset management, compliance checks, and process bottlenecks.

Short video voiceovers, advertising materials, course narration, game voiceovers, and customer service announcements often only require changing a person's name, date, brand name, or price, while the traditional process requires re-recording, re-editing, and re-reviewing.

ViiTorVoice-NAR's voice cloning is responsible for generating new content, local voice editing is responsible for changing a few words in the old audio, and emotion control is responsible for adding tone changes, making the redesign easier.

What's saved isn't just the dubbing fee, but also the time spent on back-and-forth communication, waiting for scheduling, and repeated editing.

Individual creators can use it for audio creation and script revision, while companies can use it as an internal voice-over tool and build a lower-cost production process around licensed sound assets.

AI When voice technology truly enters production, it doesn't rely on generating something amazing the first time, but on minimizing revisions each time.

Original link:Low-starred gems on GitHubfreeSound reproduction project