app.py

import gradio as gr
import numpy as np
from f5_tts.api import F5TTS

# -----------------------
# Load model ONCE globally
# -----------------------
# Works with the older F5TTS API (no vocoder_name argument).
# It will use default base model + default vocoder.
f5 = F5TTS(device="cpu")  # or simply F5TTS() if you like


def clone_voice(
    ref_audio,
    ref_text,
    gen_text,
    nfe_step,
    speed,
    target_rms,
):
    if ref_audio is None:
        raise gr.Error("Please upload a 5–15 second reference audio.")

    if not ref_text.strip():
        raise gr.Error(
            "Please enter the EXACT transcript of your reference audio.\n"
            "This avoids using a slow ASR model on CPU."
        )

    if not gen_text.strip():
        raise gr.Error("Please enter the text you want to generate.")

    ref_path = ref_audio  # because we use type='filepath' in the Audio component

    wav, sr, _ = f5.infer(
        ref_file=ref_path,
        ref_text=ref_text,
        gen_text=gen_text,
        nfe_step=int(nfe_step),     # lower → faster
        speed=float(speed),         # speaking speed
        target_rms=float(target_rms)
        # other params use defaults (cfg_strength, sway_sampling_coef, etc.)
    )

    return sr, np.array(wav, dtype=np.float32)


# -----------------------
# Gradio UI
# -----------------------
with gr.Blocks() as demo:
    gr.Markdown(
        """
        # F5-TTS Voice Cloner — Optimized for Free CPU

        **Tips for best speed:**
        - Upload **5–15 seconds** of clean speech.
        - ALWAYS fill the **Reference Text** (do NOT rely on ASR).
        - Generate **1–2 sentences** at a time.
        - Lower **NFE Steps** → faster (start with 12–16).
        """
    )

    with gr.Row():
        with gr.Column():
            ref_audio = gr.Audio(
                label="Reference audio (5–15 seconds)",
                sources=["upload"],
                type="filepath"
            )

            ref_text = gr.Textbox(
                label="Reference text (transcription of the reference audio)",
                placeholder="Type EXACTLY what you said in the audio...",
                lines=2
            )

            gen_text = gr.Textbox(
                label="Text to synthesize",
                placeholder="Enter 1–2 sentences...",
                lines=3
            )

            nfe_step = gr.Slider(
                minimum=4,
                maximum=32,
                value=16,
                step=2,
                label="NFE steps (Lower = faster)",
            )

            speed = gr.Slider(
                minimum=0.7,
                maximum=1.4,
                value=1.0,
                step=0.05,
                label="Speaking speed"
            )

            target_rms = gr.Slider(
                minimum=0.05,
                maximum=0.3,
                value=0.1,
                step=0.01,
                label="Volume (RMS)"
            )

            generate_btn = gr.Button("Generate")

        with gr.Column():
            output_audio = gr.Audio(label="Output audio")

    generate_btn.click(
        fn=clone_voice,
        inputs=[ref_audio, ref_text, gen_text, nfe_step, speed, target_rms],
        outputs=output_audio
    )


if __name__ == "__main__":
    demo.launch()