Attachment 2_Voice Cloning Collection Standards and Reference
Voice Cloning Collection Standards and Reference
Voice Cloning - Audio Collection Standards:
1. Audio Duration
Effective audio duration: 1-3 minutes;
2. Recording Preparation
Prepare text content according to different use cases, ensuring the content matches the intended context (reading aloud, livestreaming, everyday speech, etc.);
3. Recording Environment
The recording environment must be quiet, with background noise below 40dB. You can use a "decibel meter" app to test the ambient noise level;
Keep the recording free of echo, reverberation, and noise;
4. Recording Equipment and Parameters
A noise-cancelling microphone or wireless lavalier mic ("little bee") is recommended for recording equipment; a windscreen or pop filter is recommended if possible;
A 48kHz sample rate is recommended for audio recording;
Save audio in a lossless format such as WAV, AIFF, or M4A; do not save audio in a compressed format such as MP3;
5. The Speaker
The speaker should not be too close to the microphone, to avoid popping sounds and recording breathing noises; each sentence should be spoken smoothly in one go;
Pronunciation should be clear and articulate, with clear pauses between sentences;
Pause for 1-2 seconds between sentences;
Keep a consistent tone and style throughout, avoiding a mix of different emotions.
6. Differences Between Voice and Avatar Requirements
V2 avatars have high requirements for the source footage — you must speak normally, wear a noise-cancelling microphone, keep the environment quiet, and record a video of at least three minutes;
V5 avatars have lower requirements for the source footage — the source video can be recorded with or without normal speech, minimum 3 minutes; GPU acceleration is required before training a V5 avatar, but not for V2;
V9 voice uses less computer performance; V6 voice uses slightly more than V9 but does not cause livestream lag. Requirements for all voice recordings: speak normally (do not read out numbers), wear a noise-cancelling microphone, keep the environment quiet with no noise at all. If the source audio has noise, use CapCut (Jianying) to apply noise reduction.