A3S — Multilingual Text-to-Speech

A3S is a multilingual Text-to-Speech (TTS) model built upon Qwen3-TTS-12Hz-0.6B-Base, extending its original 10-language support with Arabic.

The model supports 11 languages (Arabic, Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian)

Arabic speech generation supports 3 dialects:

• 🇸🇦 Saudi Arabic (Saudi Arabia)
• 🇦🇪 Emirati Arabic (United Arab Emirates)
• 🌍 Modern Standard Arabic (MSA)

Available Arabic Speakers

A3S provides three Arabic speaker identities:

Speaker
omar
hanouf
fahad

Speaker selection is performed using the speaker parameter during inference.

---

Arabic Voice Samples

Listen to the model's Arabic speech generation across different dialects.

Language / Dialect	Audio Sample
Saudi Arabic (Saudi Arabia)
Emirati Arabic (United Arab Emirates)
Modern Standard Arabic (MSA)

---

Multilingual Voice Samples

A3S supports speech generation in the following 11 languages.

Language	Audio Sample
English (English)
Chinese (Mandarin Chinese)
Japanese (Japanese)
Korean (Korean)
German (German)
French (French)
Russian (Russian)
Portuguese (Portuguese)
Spanish (Spanish)
Italian (Italian)

---

Quick Start

Get started with A3S using our ready-to-run Google Colab notebook, or follow the instructions below to run the model locally.

Local Installation

Install the required dependencies:

pip install qwen_tts soundfile

For Ubuntu or Google Colab:

apt-get update && apt-get install -y sox libsox-fmt-all

Optional: FlashAttention 2

For compatible NVIDIA GPUs (Ampere or newer):

pip install flash-attn --no-build-isolation

FlashAttention is optional and requires a compatible CUDA environment.

Load the Model

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel
from transformers import logging as hf_logging
from IPython.display import Audio, display

hf_logging.set_verbosity_error()

device = "cuda:0"

tts = Qwen3TTSModel.from_pretrained(
    "MahmoudIbrahim/A3S",
    device_map=device,
    dtype=torch.bfloat16,
    # attn_implementation="flash_attention_2",
)

Generate Speech

wavs, sr = tts.generate_custom_voice(
    text="وش علومك؟ طال عمرك كنت مسرع ومستعجل بروح الدوام، بس الزحمة هدت حيلنا وتأخرت شوي، وقفت على الطاير أخذت لي قهوة وكملت يومي.",
    speaker="hanouf",
    temperature=0.7,
    subtalker_temperature=0.3
)

sf.write("output.wav", wavs[0], sr)

display(Audio("output.wav"))

Change the speaker parameter to select a different voice.

---

Model Performance Comparison (NVIDIA A10 GPU)

The following table presents reported inference performance for A3S and other TTS models.

Model	Model Size	TTFC (Time to First Chunk)	TTFA (Time to First Audio)	RTF (Real-Time Factor)	Generation Speed	Total Latency (3.8s Audio)
A3S	0.6B	~60 ms	200–400 ms	~0.18 (5.5x faster)	~110+ tokens/sec	~650 ms
OmniVoice (Non-Streaming)	0.6B	~1,410 ms	~1,410 ms	~0.37 (2.7x faster)	N/A	~1,410 ms
Higgs	4B	~60 ms	356–550 ms	~0.70 (1.43x faster)	~39.3 tokens/sec	~2,692 ms

---

Authors & Collaborators

• Nada Saeed Mohamed — Collaborator
• Mahmoud Ibrahim — Collaborator