app.py

import gradio as gr

LAST_UPDATED = "Nov 25th 2024"

####################################
# Datos estáticos del leaderboard
####################################
leaderboard_data = [
    {'name': 'StyleTTS 2', 'STOI': 0.998, 'PESQ': 3.921, 'WER': 0.162, 'UTMOS': 3.47},
    {'name': 'Matxa-TTS', 'STOI': 0.996, 'PESQ': 3.539, 'WER': 0.179, 'UTMOS': 3.50},
    {'name': 'Matxa-TTS-multiaccent', 'STOI': 0.996, 'PESQ': 3.415, 'WER': 0.242, 'UTMOS': 2.98},
    {'name': 'StableTTS', 'STOI': 0.997, 'PESQ': 3.643, 'WER': 0.164, 'UTMOS': 2.62},
]


# Texto para la pestaña de métricas
METRICS_TAB_TEXT = """
## Metrics
Models in the leaderboard are evaluated using several key metrics: 
* **UTMOS** (UTokyo-SaruLab Mean Opinion Score),
* **WER** (Word Error Rate),
* **STOI** (Short-Time Objective Intelligibility),
* **PESQ** (Perceptual Evaluation of Speech Quality).
These metrics help evaluate both the accuracy and quality of the model.
### UTMOS (UTokyo-SaruLab Mean Opinion Score)[[Paper](https://arxiv.org/abs/2204.02152)]
UTMOS is a MOS prediction system. **A higher UTMOS indicates better quality** of the generated voice.
### WER (Word Error Rate)
WER is a common metric for evaluating speech recognition systems. It measures the percentage of words in the generated transcript that differ from the reference (correct) transcript. **A lower WER value indicates higher accuracy**.
Example:
| Reference   | the  | cat | sat     | on  | the  | mat |
|-------------|------|-----|---------|-----|------|-----|
| Prediction  | the  | cat | **sit** | on  | the  |     |
| Label       | ✅   | ✅  | S       | ✅  | ✅   | D   |
The WER calculation is done as follows:
```
WER = (S + I + D) / N = (1 + 0 + 1) / 6 = 0.333
```
### STOI (Short-Time Objective Intelligibility)[[Paper](https://ieeexplore.ieee.org/abstract/document/5495701?casa_token=PLtqLc8KNAgAAAAA:FOLuZ4dgMYsnGb1dQHgqVOouQzRJ3vA5yqj-sbwf8gs9Q-AIDCLkMZzAgzRrAogwwxULK9zsYeE)]
STOI measures the intelligibility of the synthesized speech signal compared to the original signal. **A higher STOI indicates better intelligibility**.
### PESQ (Perceptual Evaluation of Speech Quality)[[Paper](https://ieeexplore.ieee.org/abstract/document/941023?casa_token=jdtHy84_KhQAAAAA:qHN3WbT6cNdufj6OOn_fn0Je0RedMv-WJCmhQ_3CWy4nMTuDvFMF3KstAmKqLx5suQwdPgGByoY)]
PESQ is a perceptual metric that evaluates the quality of speech in a similar manner to how a human listener would. **A higher PESQ indicates better voice quality**.
## Benchmark Datasets
Model performance is evaluated using [our test datasets](https://huggingface.co/spaces/projecte-aina/catalan_tts_arena/blob/main/bsc.txt). These datasets cover a variety of domains and acoustic conditions, ensuring a robust evaluation.
"""

CITATION_TEXT = """@misc{catalan-tts-arena,
	title        = {Catalan Texto-to-Speech Leaderboard},
	author       = {Rodolfo Zevallos, José Giraldo, Alex Peiró-Lilja, Carme Armentano-Oller},
	year         = 2024,
	publisher    = {Hugging Face},
	howpublished = "\\url{https://huggingface.co/spaces/projecte-aina/catalan_tts_arena}"
}
"""

DESCR = """
# 🏆 Catalan TTS Arena: Benchmarking TTS Models
\nThe Catalan TTS Leaderboard ranks and evaluates TTS models in Catalan.
\nThe leaderboard currently focuses on Catalan TTS, and will be expanded to multilingual evaluation in later versions.
""".strip()

####################################
# Functions (static version)
####################################

def get_leaderboard():
    """
    Retorna el leaderboard en orden descendente por PESQ y luego por UTMOS.
    """
    # Ordenar primero por PESQ (calidad del habla) y luego por UTMOS (calidad percibida)
    sorted_leaderboard = sorted(leaderboard_data, key=lambda x: (x['UTMOS']), reverse=True)
    
    # Asignar el rank basado en el orden por PESQ
    for rank, model in enumerate(sorted_leaderboard):
        model['rank'] = rank + 1  # rank es la posición en la lista (1-indexed)
    
    return [[model['rank'], model['name'], model['UTMOS'], model['WER'], model['STOI'], model['PESQ']] for model in sorted_leaderboard]

####################################
# Interfaz con Gradio
####################################

theme = gr.themes.Base(
    font=[gr.themes.GoogleFont('Libre Franklin'), gr.themes.GoogleFont('Public Sans'), 'system-ui', 'sans-serif'],
)

with gr.Blocks(theme=theme) as demo:
    gr.Markdown(DESCR, elem_classes="markdown-text")
    
    with gr.Tabs(elem_classes="tab-buttons") as tabs:
        with gr.TabItem("🏅 Leaderboard", elem_id="od-benchmark-tab-table", id=0):
            leaderboard_table = gr.DataFrame(
                headers=["Rank", "Model", "UTMOS ⬆️️", "WER ⬇️", "STOI", "PESQ"], 
                datatype=["str", "str", "str", "str", "str", "str"], 
                value=get_leaderboard()  # Carga los datos iniciales de la tabla
            )

        with gr.TabItem("📈 Metrics", elem_id="od-benchmark-tab-table", id=1):
            gr.Markdown(METRICS_TAB_TEXT, elem_classes="markdown-text")

        
    gr.Markdown(f"Last updated on **{LAST_UPDATED}**", elem_classes="markdown-text")

    with gr.Row():
        with gr.Accordion("📙 Citation", open=False):
            gr.Textbox(
                value=CITATION_TEXT, lines=7,
                label="Copy the BibTeX snippet to cite this source",
                elem_id="citation-button",
                show_copy_button=True,
            )


# Lanzar la aplicación
demo.queue(api_open=False, default_concurrency_limit=40).launch(show_api=False)