app.py

"""Gradio web UI for the phonetic-reduction analyzer.

Run with:
    .venv/bin/python app.py
Then open the URL Gradio prints (defaults to http://127.0.0.1:7860).
"""
from __future__ import annotations

import os
import tempfile

import gradio as gr
import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt
import pandas as pd

from reduction import pipeline, stt, tts


# --- model cache ---------------------------------------------------------
_models: dict[str, object] = {}
_voice = None


def _get_model(size: str):
    if size not in _models:
        _models[size] = stt.load_model(size=size)
    return _models[size]


def _get_voice():
    global _voice
    if _voice is None:
        _voice = tts.load_voice()
    return _voice


# --- plot ----------------------------------------------------------------
def _make_plot(df: pd.DataFrame) -> plt.Figure:
    n = len(df)
    fig, ax = plt.subplots(figsize=(max(8, n * 0.55), 4.5))
    x = list(range(n))
    labels = df["word"].tolist()

    # Colour bars by STT confidence: low-confidence → red, normal → steelblue
    probs = df["stt_probability"].fillna(1.0).tolist()
    bar_colors = ["#cc3333" if p < 0.6 else "#4a90c4" for p in probs]
    ax.bar(x, df["combined_score"], color=bar_colors, alpha=0.55, label="combined")
    ax.plot(x, df["duration_score"], "o-", color="#2c7a2c", lw=1.4, label="duration", alpha=0.85)
    ax.plot(x, df["spectral_score"], "s-", color="#d97726", lw=1.4, label="spectral", alpha=0.85)
    ax.axhline(1.0, color="gray", linestyle="--", linewidth=0.8)

    ax.set_xticks(x)
    ax.set_xticklabels(labels, rotation=40, ha="right", fontsize=10)
    ax.set_ylabel("skóre  (1 = průměr,  <1 redukováno)")
    ax.set_xlabel("token")
    ax.set_title("Per-word reduction scores")
    ax.legend(loc="upper right", framealpha=0.9)
    ax.grid(True, axis="y", alpha=0.25)
    fig.tight_layout()
    return fig


# --- main analyze callback ----------------------------------------------
def analyze_audio(audio_path: str | None, whisper_size: str, progress=gr.Progress()):
    if not audio_path:
        return "❌ Nahraj nebo nahrej zvuk.", None, None, None

    progress(0.1, desc=f"Načítám Whisper {whisper_size}…")
    model = _get_model(whisper_size)
    progress(0.3, desc="Načítám Piper hlas…")
    voice = _get_voice()

    progress(0.5, desc="Analyzuji…")
    rows, whisper_text, _ = pipeline.analyze(
        audio_path, model, voice, language="cs", groundtruth=None, trim=True
    )

    progress(0.9, desc="Renderuji graf a CSV…")
    df = pd.DataFrame(rows)

    # Pretty display dataframe (round numerics, drop internal cols)
    display_cols = [
        "word",
        "stt_probability",
        "orig_duration",
        "tts_duration",
        "duration_ratio",
        "spectral_distance",
        "duration_score",
        "spectral_score",
        "combined_score",
    ]
    display_df = df[display_cols].copy()
    for c in display_cols[1:]:
        display_df[c] = display_df[c].round(3)

    fig = _make_plot(df)

    # CSV download (full row, not abbreviated)
    csv_path = os.path.join(tempfile.gettempdir(), "reduction_result.csv")
    df.to_csv(csv_path, index=False)

    return whisper_text or "(žádný transkript)", display_df, fig, csv_path


# --- UI -----------------------------------------------------------------
HEADER_MD = """
# Měření fonetické redukce — prototyp
Per-word reduction scores via STT → per-word TTS canonical → duration + spectral DTW.
Skóre **< 1** = redukováno oproti průměru nahrávky, **> 1** = plně artikulováno.
Červené sloupce = nízká STT confidence (pravděpodobně přeslech / halucinace).
"""

with gr.Blocks(title="Phonetic reduction analyzer") as demo:
    gr.Markdown(HEADER_MD)
    with gr.Row():
        with gr.Column(scale=2):
            audio_in = gr.Audio(
                sources=["upload", "microphone"],
                type="filepath",
                label="Nahrávka (drag & drop wav, nebo klikni na mikrofon a namluv)",
            )
        with gr.Column(scale=1):
            whisper_dd = gr.Dropdown(
                choices=["small", "medium", "large-v3"],
                value="large-v3",
                label="Whisper model",
                info="large-v3 = nejlepší kvalita, ~6s/věta na CPU",
            )
            run_btn = gr.Button("Analyzuj", variant="primary", size="lg")

    transcript_out = gr.Textbox(label="Whisper transkript", interactive=False, lines=2)
    plot_out = gr.Plot(label="Skóre redukce")
    table_out = gr.Dataframe(label="Per-word detaily", interactive=False)
    csv_out = gr.File(label="Stáhnout CSV", interactive=False)

    run_btn.click(
        analyze_audio,
        inputs=[audio_in, whisper_dd],
        outputs=[transcript_out, table_out, plot_out, csv_out],
    )


if __name__ == "__main__":
    # On HuggingFace Spaces, $SPACE_ID is set; on local dev it isn't.
    on_hf = bool(os.environ.get("SPACE_ID"))
    if on_hf:
        # HF provides PORT and expects 0.0.0.0; let HF's launcher own model preloading
        demo.launch(server_name="0.0.0.0")
    else:
        print("[startup] preloading large-v3 + Piper voice (one-time, ~10s)…")
        _get_model("large-v3")
        _get_voice()
        print("[startup] ready, launching Gradio…")
        demo.launch(server_name="127.0.0.1", server_port=7860, inbrowser=False)