How to Clone a Voice with CosyVoice v3 Plus (Free, Step-by-Step)

CosyVoice v3 Plus clones a voice from a short recording and then reads any script in that voice. The part that surprises people is that enrollment is free. You are not billed for creating the cloned voice. You are billed only for the speech you generate afterward, per character, and the cost shows before you hit generate. This makes it a cheap way to lock a narrator or a recurring host and reuse it across videos.

This is a real workflow guide for the CosyVoice Voice Clone tool on Moosky, not a general explainer. If you want a voice invented from a text description rather than cloned from a recording, the Fish Audio voice design guide covers that instead. Design invents a persona from nothing; cloning preserves a voice you already have. If you have a recording of the person you want, cloning is usually the better path, and our voice cloning 5-step guide frames the whole decision.

What CosyVoice actually does

CosyVoice takes a reference clip with clear speech and produces a voice profile that carries the speaker's timbre. You then feed it a script and it synthesizes natural-sounding narration in that timbre. It is a fast, API-based path: enrollment is near-instant and synthesis comes back quickly, which is why it suits people iterating on scripts rather than waiting on long renders. (Moosky also has a GPU-based option, Vibe Voice, if you want that route; CosyVoice is the API one.)

The model supports seven languages as a hint for both the reference and the output: English, Chinese, French, German, Japanese, Korean, and Russian. Match your script to the language you set, and the prosody holds together best. One limitation worth knowing up front: for cloned voices on this endpoint you cannot steer delivery with style instructions, only with the volume, speed, and pitch sliders described below.

Step 1: Prepare the reference clip

This is where most failed clones die, so read it before you record. CosyVoice wants a clip with clear, continuous speech. The requirements the tool enforces:

• Length: at least 5 seconds of real speech. The sweet spot the model was tuned on is 10 to 20 seconds. Anything up to 60 seconds is accepted by the API.
• Size: under 10 MB.
• Source: audio or video. If you upload a video that has an audio track, the tool extracts the speech for you, so a clip pulled from a podcast or a recorded intro works.
• Quality: one voice, no music bed, no overlapping speakers, no loud room echo. Background music and second voices are the top cause of a muddy clone.

You do not have to trim it yourself. If your upload is longer than 20 seconds or larger than 10 MB, CosyVoice trims it automatically before enrollment. If your clip is shorter than 5 seconds of speech, it is rejected outright, so plan for at least a sentence or two of continuous talking.

Record the person saying natural connected sentences, not a word list. The model is learning how they move between sounds, so a real phrase teaches it far more than isolated words.

Step 2: Open the CosyVoice Voice Clone tool

Go to the CosyVoice voice clone page and launch the tool. The compose screen has three parts: an Audio input for your reference clip, a Script field for the text the cloned voice will speak, and an Advanced panel with the controls that matter for a clone: Language, Voice Name Prefix, Volume, Speech Rate, and Pitch. You need to be signed in to generate.

Step 3: Upload the sample and enroll

Upload your reference clip into the Audio input. On the first run, the tool enrolls the voice: it submits the sample to Alibaba's voice-enrollment, waits for the profile to become ready, and returns a voice ID. This enrollment step costs nothing. The clone either comes back ready or tells you the sample failed, in which case the fix is almost always in your clip: too short, too noisy, or more than one speaker.

The Voice Name Prefix field in Advanced is optional. It labels the enrolled voice with a short alphanumeric tag (up to 10 characters) so you can recognize it later. Leave it blank and the system auto-generates one. The Language dropdown sets the hint for both your sample and your script, default English.

Once enrolled, the voice ID is saved on your project so later runs reuse it. You do not re-upload the sample or re-enroll every time; the profile persists.

Step 4: Write the script and set delivery

Type the text you want the voice to speak into the Script field. Then decide on delivery using the three Advanced sliders:

• Volume: output loudness, 0 to 100, default 50.
• Speech Rate: speed multiplier, default 1.0 for normal pace.
• Pitch: global pitch shift, default 1.0 (unchanged).

Those sliders are your only delivery levers for a cloned voice on this endpoint. If the clone sounds too fast or flat, adjust speed and pitch rather than re-recording, because the timbre is locked in and pacing is a synthesis choice. Generate a take, listen, then tweak and re-run. The cost is shown before you generate, and it is billed by the number of characters in your script, so short audition lines cost fractions of a cent.

Step 5: Turn the voice into narrated video

A cloned voice on its own is just audio. The payoff is using it to narrate video. Generate your narration in the cloned voice, then feed that audio into lip-synced video scenes, and an on-screen character speaks in that exact voice. This is how a recurring host, a brand narrator, or a faceless-channel voice stays identical across every video you publish. The same voice ID also drives straight text-to-speech, so you can produce voiceover and explainer tracks with no video at all. For the end-to-end version of this, see the faceless YouTube workflow guide, where cloned narration is the spine of the whole pipeline.

Troubleshooting the clone

A few recurring problems and their actual fixes:

• The sample is rejected as too short. It has fewer than 5 seconds of continuous speech. Re-record a longer phrase, or pick a segment of the video that has more talking.
• The voice sounds muffled or wrong. Your clip had music, echo, or a second voice. The clone is only as clean as its source, so give it isolated dry speech.
• The output language is off. Your script and the Language setting disagree. Set the dropdown to match the script.
• It sounds robotic on long paragraphs. Adjust Speech Rate and Pitch. Cloned delivery cannot be style-directed, so pacing changes happen through the sliders.

The workflow in one paragraph

Record or extract 10 to 20 seconds of clean single-speaker speech, under 10 MB. Upload it to the CosyVoice Voice Clone tool on Moosky; enrollment is free and the voice ID persists on your project. Write your script, pick the language, and tune volume, speed, and pitch with the Advanced sliders. Generate audio, listen, and adjust. Then use that voice to narrate video through lip-synced scenes, or for plain text-to-speech voiceover. Enrollment free, synthesis billed per character with the cost shown first, no subscription. That is the whole loop.