I Tested Zero-Shot Voice Cloning with Emotion Control in OpenVoice — 8 Styles from a 14-Second Reference

#!/usr/bin/env python3
import argparse
import os
import sys

import torch

ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
if ROOT not in sys.path:
    sys.path.insert(0, ROOT)

from openvoice import se_extractor
from openvoice.api import BaseSpeakerTTS, ToneColorConverter


ENGLISH_STYLES = [
    "default",
    "whispering",
    "shouting",
    "excited",
    "cheerful",
    "terrified",
    "angry",
    "sad",
    "friendly",
]

DEFAULT_TEXTS = {
    "English": "This audio is generated by OpenVoice.",
    "Chinese": "今天天气真好，我们一起出去吃饭吧。",
}


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description="Run OpenVoice V1 style-controlled voice cloning."
    )
    parser.add_argument(
        "--reference",
        required=True,
        help="Reference audio path used for tone color extraction.",
    )
    parser.add_argument(
        "--text", default=None, help="Text to synthesize before tone color conversion."
    )
    parser.add_argument(
        "--language",
        default="English",
        choices=["English", "Chinese"],
        help="Base speaker language used by OpenVoice V1.",
    )
    parser.add_argument(
        "--style",
        default="default",
        help="Style preset. English supports angry/sad/friendly/etc. Chinese supports default only.",
    )
    parser.add_argument(
        "--output", default="outputs/output_v1.wav", help="Output wav path."
    )
    parser.add_argument("--speed", type=float, default=1.0, help="TTS speed.")
    parser.add_argument(
        "--device", default=None, help="Torch device, e.g. cuda:0 or cpu."
    )
    parser.add_argument(
        "--tau",
        type=float,
        default=0.3,
        help="Tone color conversion strength used by the converter.",
    )
    return parser.parse_args()


def validate_args(args: argparse.Namespace) -> None:
    if args.language == "English":
        if args.style not in ENGLISH_STYLES:
            raise SystemExit(
                f"Unsupported English style '{args.style}'. Available: {', '.join(ENGLISH_STYLES)}"
            )
    elif args.style != "default":
        raise SystemExit("Chinese in OpenVoice V1 supports only --style default.")


def main() -> None:
    args = parse_args()
    validate_args(args)

    device = args.device or ("cuda:0" if torch.cuda.is_available() else "cpu")
    text = args.text or DEFAULT_TEXTS[args.language]

    ckpt_base = (
        "checkpoints/base_speakers/EN"
        if args.language == "English"
        else "checkpoints/base_speakers/ZH"
    )
    ckpt_converter = "checkpoints/converter"

    base_speaker_tts = BaseSpeakerTTS(f"{ckpt_base}/config.json", device=device)
    base_speaker_tts.load_ckpt(f"{ckpt_base}/checkpoint.pth")

    tone_color_converter = ToneColorConverter(
        f"{ckpt_converter}/config.json", device=device
    )
    tone_color_converter.load_ckpt(f"{ckpt_converter}/checkpoint.pth")

    # Select Speaker Embedding based on style
    if args.language == "English":
        se_name = "en_style_se.pth" if args.style != "default" else "en_default_se.pth"
    else:
        se_name = "zh_default_se.pth"

    source_se = torch.load(f"{ckpt_base}/{se_name}", map_location=device).to(device)
    target_se, _ = se_extractor.get_se(
        args.reference, tone_color_converter, target_dir="processed_v1", vad=True
    )

    out_dir = os.path.dirname(args.output) or "."
    os.makedirs(out_dir, exist_ok=True)
    src_path = os.path.join(out_dir, "tmp_openvoice_v1.wav")

    base_speaker_tts.tts(
        text, src_path, speaker=args.style, language=args.language, speed=args.speed
    )
    tone_color_converter.convert(
        audio_src_path=src_path,
        src_se=source_se,
        tgt_se=target_se,
        output_path=args.output,
        tau=args.tau,
        message="@MyShell",
    )
    print(args.output)


if __name__ == "__main__":
    main()

# 1. Clone the repo
git clone https://github.com/myshell-ai/OpenVoice.git
cd OpenVoice

# 2. Install dependencies
pip install -e .

# 3. Download checkpoints (see official README)
# Place them under checkpoints/base_speakers/EN/ and checkpoints/converter/

# 4. Run a single style
python scripts/run_v1.py \
  --reference path/to/reference.wav \
  --style angry \
  --output outputs/angry_output.wav

for style in whispering shouting excited cheerful terrified angry sad friendly; do
  python scripts/run_v1.py \
    --reference reference.wav \
    --style "$style" \
    --output "outputs/${style}.wav"
done

[1] BaseSpeakerTTS
    Text → styled audio (in the base model's voice / source_se)
         ↓
[2] se_extractor
    Reference audio → target_se (speaker embedding vector)
         ↓
[3] ToneColorConverter
    source_se × target_se → final audio (reference voice × target emotion)