ice-012-audio

Overview

ICE-012 Audio is a multilingual TTS model with streaming output and reference-based voice cloning. It introduces an active acoustic adapter — conditioning codec embeddings before the backbone and refining hidden states after it.

Quick Start

from transformers import AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained(
    "darkps/ice-012-audio",
    trust_remote_code=True,
    device_map="auto",
)

Usage

Text-to-Speech

python 02.py --language "arabic" --gender female --style "young adult" --speed 1.0 --output out.wav
python 02.py --text-file input.txt --language "en" --gender male --style "elderly, low pitch"
python 02.py --language "egyptian lahgtna" --gender female --auto-voice

Voice Cloning

python 03.py --reference ref.wav --output clone.wav
python 03.py --reference ref.wav --text-file input.txt --reference-text "transcript"

Use only audio you own or have explicit permission to process.

Voice Options

Gender

Age · Pitch · Style (--style)

Combining modifiers:

--gender female --style "young adult, high pitch"
--gender male   --style "elderly, low pitch"
--gender female --style "teenager, british accent"

Speed (--speed)

0.5 · 1.0 (default) · 1.5 · 2.0

Language Parameter

Accepts a full name (case-insensitive) or a 2–3-letter ID. Unrecognized values fall back to language-agnostic mode.

# All three are equivalent:
language = "palestinian lahgtna"
language = "Palestinian Lahgtna"
language = "ps"

ICE Model Family

Audio Samples

🎙️  Sample 01 🎙️  Sample 02 🎙️  Sample 03 🎙️  Sample 04

Language Coverage

Supported Languages

Arabic Dialects

Full List (590 languages)