silma-tts

SILMA TTS v1 is a high-performance, 150M-parameter bilingual (Arabic/English) TTS model developed by SILMA AI. Built on the cutting-edge F5-TTS diffusion architecture, the model was pretrained from scratch using tens of thousands of hours of high-quality public and proprietary data. To give back to the community, SILMA TTS is released under a highly permissive license, making state-of-the-art speech synthesis accessible for both research and commercial use.

If you’re looking for our commercial TTS models, please visit the page below: https://silma.ai/arabic-text-to-speech

Model Details

Installation

Using pip

# make sure ffmpeg is installed
apt-get update && apt install ffmpeg -y

# create and activate the environment
python -m venv silma-tts-env
source silma-tts-env/bin/activate

# install silma-tts library
pip install silma-tts

From source

# make sure ffmpeg is installed
apt-get update && apt install ffmpeg -y

# create and activate the environment
python -m venv silma-tts-env
source silma-tts-env/bin/activate

# clone repo and install
git clone https://github.com/SILMA-AI/silma-tts.git
cd silma-tts
pip install -e .

Usage & Inference

Using Gradio app

# Run the following command
silma-tts-app

Then open the following browser link http://127.0.0.1:7860/

Inference using Python

import time
from silma_tts.api import SilmaTTS

silma_tts = SilmaTTS()

## the voice/style you want to clone
reference_audio_file = "/root/silma-tts/src/silma_tts/infer/ref_audio_samples/ar.ref.24k.wav"
## the transcription of the reference_audio_file
reference_audio_text = "ويدقق النظر في القرآن الكريم وسائر الكتب السماوية ويتبع مسالك الرسل العظام عليهم الصلاة والسلام."

time_start = time.time()

wav, sr, spec = silma_tts.infer(
    ref_file=reference_audio_file,
    ref_text=reference_audio_text, # can also be left None - will be transcribed on the fly
    gen_text="""
    أنا نموذج جديد من سلمى لتحويل النص إلى كلام، يمكنني التحدث باللغة العربية مع أو بدون علامات التشكيل.
    I am the new SILMA model for converting text to speech, I can speak Arabic with or without diacritics.
    """.strip(),
    file_wave=str("generated_audio.wav"),
    seed=None,
    speed=1
)

time_end = time.time()
print(f"Time elapsed:{(time_end-time_start):.2f} seconds")

## Note 1: generated audio file (generated_audio.wav) will be saved in the current directory
## Note 2: You can also use the "wav" variable (raw waveform) to play the audio or return it via API

You can also run the example above directly using the following command, but only if you installed from source:

python src/silma_tts/infer/example.py

Training

Our model is 100% compatible with F5-TTS v1.1.7. This means you can make use of all the great resources, tools and community experience in the F5-TTS project.

Steps

## clone F5-TTS v1.1.7
cd /root
git clone --depth 1  --branch 1.1.7 https://github.com/SWivid/F5-TTS.git
cd F5-TTS
pip install -e .

## download silma-tts model weights, vocab.txt, patched finetuning script and config.yaml
hf download silma-ai/silma-tts  --local-dir /root/silma-tts-v1-weights

## create the project, then replace the default configuration, vocabulary, and fine-tuning Python file. Note that the patched finetune_cli.py overrides the F5TTS_v1_Base config with the silma-tts model config.
mkdir /root/F5-TTS/data/finetuning_project_char
cp /root/silma-tts-v1-weights/vocab.txt /root/F5-TTS/data/finetuning_project_char
cp /root/silma-tts-v1-weights/finetune_cli.py /root/F5-TTS/src/f5_tts/train/finetune_cli.py
echo /root/silma-tts-v1-weights/config.yaml > /root/F5-TTS/src/f5_tts/configs/F5TTS_v1_Base.yaml

## open F5-TTS UI training pipeline
f5-tts_finetune-gradio --port 7860 --host 0.0.0.0

## After preparing your data: go to "Train Model" tab -> "Path to the Pretrained Checkpoint" and enter the path to the silma-tts model weights file: /root/silma-tts-v1-weights/model.pt while leaving "Tokenizer File" empty

## For more information please follow the F5-TTS training guide below
## https://github.com/SWivid/F5-TTS/tree/main/src/f5_tts/train

Summary: you need to use the F5-TTS v1.1.7 training code, but use our config file, vocab, patched finetuning script and our pretrained weights

Acknowledgements

This repo builds directly upon the excellent foundation laid by the F5-TTS project. The core architecture and the majority of the code is derived from their work. Our work introduces new pretrained weights and significant optimizations to the inference code.

Other projects

1. We use CATT to enrich arabic text with Tashkeel in case text doesn't include tashkeel.
2. We use NeMo-text-processing to handle text normalization.

Citation

@article{silma-tts-v1,
  title   = {SILMA TTS: A Lightweight