Skip to main content

Voice Cloning

Clone any voice with just 3-5 seconds of reference audio. No fine-tuning needed (zero-shot).

How It Works​

  1. You provide a short audio clip + its transcript
  2. The codec encodes the audio into speech tokens
  3. These tokens are used as context for the LLM
  4. The LLM continues generating in the same voice style

Basic Cloning​

from vieneu import Vieneu

tts = Vieneu()

audio = tts.infer(
    text="Đây là giọng nói được clone.",
    ref_audio="path/to/speaker.wav",
    ref_text="Exact transcript of what the speaker says in the audio.",
)
tts.save(audio, "cloned_output.wav")

Tips for Best Results​

  • Audio quality: Use clean recordings without background noise
  • Duration: 3-5 seconds is ideal. Too short = poor quality, too long = wasted context
  • Transcript accuracy: The ref_text must accurately match what's spoken
  • Language: Reference audio should be in Vietnamese for best results

Using Encoded Codes Directly​

ref_codes = tts.encode_reference("speaker.wav")

for text in texts:
    audio = tts.infer(
        text=text,
        ref_codes=ref_codes,
        ref_text="Transcript of reference audio.",
    )

Preset Voices​

voices = tts.list_preset_voices()
for description, voice_id in voices:
    print(f"{voice_id}: {description}")

voice = tts.get_preset_voice("voice_name")
audio = tts.infer(text="Chào bạn!", voice=voice)