# CosyVoice3: Clone a Voice, Then Generate the Narration
Voice cloning plus text-to-speech in ComfyUI. You feed a short reference clip, the workflow captures
the timbre, and then it generates the narration from text — which is the audio track that will drive
everything downstream in a digital-human pipeline.
## Why voice first
In a talking-head pipeline the audio is the driver. The mouth movement, the head motion and the eye
line are all timed from the waveform. Get the voice right before you generate anything visual,
because changing the narration later means regenerating the head.
## What this workflow does
1. **Clone** — load a short reference clip (a few seconds of clean speech is enough) and extract the
speaker's voice characteristics.
2. **Speak** — feed the target text; the workflow synthesises speech in the cloned voice.
3. **Hand off** — the resulting audio track becomes the input for the OmniRef or Pika digital-human
workflow.
## Tips from the episode
- Keep the reference clip clean: no music, no room echo, one speaker only.
- Match the reference language to the target language, otherwise the prosody drifts.
- For long narration, generate in sentence-sized chunks and join them — long single passes tend to
flatten the intonation.
- Always listen to the first 5 seconds and the last 5 seconds; that is where artefacts show up.
---
**About this workflow**
Part of AI-KSK episode 329, which builds a talking-head digital human from a single image plus an audio track and finishes it at 2K. This is a ComfyUI workflow export: it references models that must be present in your ComfyUI environment. On RunningHub the cloud instance already has them installed.
It is a tutorial workflow, not a packaged product. Node parameters are the ones used in the episode; adjust frame count, sample steps and resolution for your own hardware.
**License / credit:** independent AI-KSK tutorial release, not affiliated with or endorsed by MiniMax, Alibaba, ByteDance or RunningHub. Model licences remain with their respective authors — check each model's own licence before commercial use.
Voice cloning plus text-to-speech in ComfyUI. You feed a short reference clip, the workflow captures
the timbre, and then it generates the narration from text — which is the audio track that will drive
everything downstream in a digital-human pipeline.
## Why voice first
In a talking-head pipeline the audio is the driver. The mouth movement, the head motion and the eye
line are all timed from the waveform. Get the voice right before you generate anything visual,
because changing the narration later means regenerating the head.
## What this workflow does
1. **Clone** — load a short reference clip (a few seconds of clean speech is enough) and extract the
speaker's voice characteristics.
2. **Speak** — feed the target text; the workflow synthesises speech in the cloned voice.
3. **Hand off** — the resulting audio track becomes the input for the OmniRef or Pika digital-human
workflow.
## Tips from the episode
- Keep the reference clip clean: no music, no room echo, one speaker only.
- Match the reference language to the target language, otherwise the prosody drifts.
- For long narration, generate in sentence-sized chunks and join them — long single passes tend to
flatten the intonation.
- Always listen to the first 5 seconds and the last 5 seconds; that is where artefacts show up.
---
**About this workflow**
Part of AI-KSK episode 329, which builds a talking-head digital human from a single image plus an audio track and finishes it at 2K. This is a ComfyUI workflow export: it references models that must be present in your ComfyUI environment. On RunningHub the cloud instance already has them installed.
It is a tutorial workflow, not a packaged product. Node parameters are the ones used in the episode; adjust frame count, sample steps and resolution for your own hardware.
**License / credit:** independent AI-KSK tutorial release, not affiliated with or endorsed by MiniMax, Alibaba, ByteDance or RunningHub. Model licences remain with their respective authors — check each model's own licence before commercial use.
Description
# CosyVoice3: Clone a Voice, Then Generate the Narration
Voice cloning plus text-to-speech in ComfyUI. You feed a short reference clip, the workflow captures
the timbre, and then it generates the narration from text — which is the audio track that will drive
everything downstream in a digital-human pipeline.
## Why voice first
In a talking-head pipeline the audio is the driver. The mouth movement, the head motion and the eye
line are all timed from the waveform. Get the voice right before you generate anything visual,
because changing the narration later means regenerating the head.
## What this workflow does
1. **Clone** — load a short reference clip (a few seconds of clean speech is enough) and extract the
speaker's voice characteristics.
2. **Speak** — feed the target text; the workflow synthesises speech in the cloned voice.
3. **Hand off** — the resulting audio track becomes the input for the OmniRef or Pika digital-human
workflow.
## Tips from the episode
- Keep the reference clip clean: no music, no room echo, one speaker only.
- Match the reference language to the target language, otherwise the prosody drifts.
- For long narration, generate in sentence-sized chunks and join them — long single passes tend to
flatten the intonation.
- Always listen to the first 5 seconds and the last 5 seconds; that is where artefacts show up.
---
**About this workflow**
Part of AI-KSK episode 329, which builds a talking-head digital human from a single image plus an audio track and finishes it at 2K. This is a ComfyUI workflow export: it references models that must be present in your ComfyUI environment. On RunningHub the cloud instance already has them installed.
It is a tutorial workflow, not a packaged product. Node parameters are the ones used in the episode; adjust frame count, sample steps and resolution for your own hardware.
**License / credit:** independent AI-KSK tutorial release, not affiliated with or endorsed by MiniMax, Alibaba, ByteDance or RunningHub. Model licences remain with their respective authors — check each model's own licence before commercial use.
Voice cloning plus text-to-speech in ComfyUI. You feed a short reference clip, the workflow captures
the timbre, and then it generates the narration from text — which is the audio track that will drive
everything downstream in a digital-human pipeline.
## Why voice first
In a talking-head pipeline the audio is the driver. The mouth movement, the head motion and the eye
line are all timed from the waveform. Get the voice right before you generate anything visual,
because changing the narration later means regenerating the head.
## What this workflow does
1. **Clone** — load a short reference clip (a few seconds of clean speech is enough) and extract the
speaker's voice characteristics.
2. **Speak** — feed the target text; the workflow synthesises speech in the cloned voice.
3. **Hand off** — the resulting audio track becomes the input for the OmniRef or Pika digital-human
workflow.
## Tips from the episode
- Keep the reference clip clean: no music, no room echo, one speaker only.
- Match the reference language to the target language, otherwise the prosody drifts.
- For long narration, generate in sentence-sized chunks and join them — long single passes tend to
flatten the intonation.
- Always listen to the first 5 seconds and the last 5 seconds; that is where artefacts show up.
---
**About this workflow**
Part of AI-KSK episode 329, which builds a talking-head digital human from a single image plus an audio track and finishes it at 2K. This is a ComfyUI workflow export: it references models that must be present in your ComfyUI environment. On RunningHub the cloud instance already has them installed.
It is a tutorial workflow, not a packaged product. Node parameters are the ones used in the episode; adjust frame count, sample steps and resolution for your own hardware.
**License / credit:** independent AI-KSK tutorial release, not affiliated with or endorsed by MiniMax, Alibaba, ByteDance or RunningHub. Model licences remain with their respective authors — check each model's own licence before commercial use.