app.py

import pandas as pd
import gradio as gr

def compare_csv_files():
    df1 = pd.read_csv("fish-speech-1.5.csv")
    df2 = pd.read_csv("fish-speech-1.4.csv")
    
    merged_df = pd.merge(df1, df2, on="SourceText", suffixes=("_1.5", "_1.4"))
    
    merged_df["WordErrorRate_Diff"] = merged_df["WordErrorRate_1.5"] - merged_df["WordErrorRate_1.4"]
    merged_df["CharacterErrorRate_Diff"] = merged_df["CharacterErrorRate_1.5"] - merged_df["CharacterErrorRate_1.4"]
    
    merged_df["WordErrorRate_Comparison"] = merged_df["WordErrorRate_Diff"].apply(
        lambda x: "1.4 is the same as 1.5 (Ignored due to large diff)" if abs(x) > 10 else (
            f"1.5 is stronger than 1.4 ({x:.8f})" if x < 0 else (
                f"1.4 is stronger than 1.5 ({-x:.8f})" if x > 0 else "1.4 is the same as 1.5 (0)"
            )
        )
    )
    merged_df["CharacterErrorRate_Comparison"] = merged_df["CharacterErrorRate_Diff"].apply(
        lambda x: "1.4 is the same as 1.5 (Ignored due to large diff)" if abs(x) > 10 else (
            f"1.5 is stronger than 1.4 ({x:.8f})" if x < 0 else (
                f"1.4 is stronger than 1.5 ({-x:.8f})" if x > 0 else "1.4 is the same as 1.5 (0)"
            )
        )
    )
    
    avg_word_diff = merged_df["WordErrorRate_Diff"].loc[merged_df["WordErrorRate_Diff"].abs() <= 10].mean()
    avg_char_diff = merged_df["CharacterErrorRate_Diff"].loc[merged_df["CharacterErrorRate_Diff"].abs() <= 1].mean()
    overall_summary = f"""
    <h3>Overall Comparison:</h3>
    <p>Average WordErrorRate Difference (excluding large diffs): {f'1.5 is stronger ({avg_word_diff:.8f})' if avg_word_diff < 0 else f'1.4 is stronger ({0 - avg_word_diff:.8f})'}</p>
    <p>Average CharacterErrorRate Difference (excluding large diffs): {f'1.5 is stronger ({avg_char_diff:.8f})' if avg_char_diff < 0 else f'1.4 is stronger ({0 - avg_char_diff:.8f})'}</p>
    """
    
    result = merged_df[[
        "SourceText",
        "WordErrorRate_1.5", "WordErrorRate_1.4", "WordErrorRate_Comparison",
        "CharacterErrorRate_1.5", "CharacterErrorRate_1.4", "CharacterErrorRate_Comparison",
        "WhisperText_1.5", "WhisperText_1.4"
    ]]
    
    return overall_summary + result.to_html(index=False)

gr.Interface(
    fn=compare_csv_files,
    inputs=None,
    outputs="html",
    title="Fish Speech Benchmark",
    description="This is a non official model performance test from Fish Speech / More data will be added later (not too much)"
).launch()