import csv
import gradio as gr
import os
import time
import os
import pandas as pd
app = None
os.environ['GRADIO_ANALYTICS_ENABLED'] = 'False'
CURR_BASE_DIR = os.getcwd()
BASE_AUDIOS_DIR = os.path.join(os.getcwd(), "results")
BASE_VIDEO_DIR = BASE_AUDIOS_DIR
should_stop = False
current_process = None
if not os.path.exists(BASE_AUDIOS_DIR):
os.makedirs(BASE_AUDIOS_DIR)
custom_css = """
.gradio-container{
background-color: unset;
}
input, textarea {
font-family: 'Noto Naskh Arabic', 'Arial', sans-serif !important;
}
.large_text input{
font-size:20px !important
}
.tab-container button{
font-weight: bold !important;
}
.tabs{
gap:0px !important;
}
.tabitem{
padding:0px !important;
}
button.secondary:hover{
background-color:steelblue !important;
color:white;
}
.tab-container button {
font-size:20px !important;
}
.tabitem{
padding-top:0px !important;
}
#tool-header{
padding:20px 0px 20px 0px !important;
}
#tool-header h1{
display: flex;
align-items: center;
}
#tool-header img{
width: 80px;
display: inline-block;
margin-right: 10px;
border-radius: 5px;
}
"""
def natural_sort_key(file_name):
if ".DS_Store" in file_name:
return 0
return int(file_name.split("_")[1])
def get_data(folders_list, language="msa"):
results_csv_df = pd.read_csv(f"{CURR_BASE_DIR}/results/{language}/Ar_{language}_TTS_benchmark.csv")
data_map = {}
try:
for folder_name in folders_list:
full_audio_folder_path = os.path.join(BASE_AUDIOS_DIR,language ,folder_name)
if not os.path.exists(full_audio_folder_path):
print(f"Folder not found: {full_audio_folder_path}")
continue
if folder_name not in data_map:
data_map[folder_name] = []
##get all wavs and mp3s from full_audio_folder_path
for file in sorted(os.listdir(full_audio_folder_path), key=natural_sort_key):
if ".DS_Store" in file:
continue
if file.endswith(".wav") or file.endswith(".mp3"):
abs_path = os.path.join(full_audio_folder_path, file)
data_map[folder_name].append(abs_path)
final_df = pd.DataFrame(columns=["Text"].extend(folders_list))
final_df["Text"] = results_csv_df["Text"]
cache_random_hash = str(time.time())
for folder_name in data_map:
folder_files_list = data_map[folder_name]
formatted_column = pd.Series(folder_files_list).apply(
lambda file_path: f''
)
## append column to dataframe
final_df[folder_name] = formatted_column
return final_df
except Exception as e:
print(f"An error occurred: {e}")
gr.Warning(f"An error occurred: {e}")
raise e
with gr.Blocks(css=custom_css) as app:
gr.HTML(
f"""
Open-source Arabic TTS Benchmark
This benchmark represents a foundational step by SILMA.AI towards establishing a high-quality benchmark for open-source Arabic TTS models. We have found that standard quantitative metrics (such as WER, CER, SIM, and UTMOS) are often insufficient for accurately capturing the nuances of Arabic speech. Consequently, for this release we have prioritized direct auditory assessment, allowing listeners to evaluate the naturalness and quality of the models themselves. More advanced releases will follow as we develop the necessary technology.
You can also check our Commercial Arabic TTS Benchmark
Note: To make the benchmark more manageable, we have excluded models that require Tashkeel, as well as those based on old architectures
Open-source Arabic TTS Benchmark