app.py

import gradio as gr
import numpy as np
import spaces
import torch
from cached_path import cached_path
from f5_tts.infer.utils_infer import (
    infer_process,
    load_model,
    load_vocoder,
    preprocess_ref_audio_text,
)
from f5_tts.model import DiT


vocoder = load_vocoder()

# common usage
v1_base_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
v1_small_cfg = dict(dim=768, depth=18, heads=12, ff_mult=2, text_dim=512, conv_layers=4)
alg_vocab_path = str(cached_path("hf://chenxie95/F5-TTS_v1_Small_Algerian/vocab.txt"))

tts_model_choice = "v1-base_zh-en"  # default
tts_model_collections = {
    "v1-base_zh-en": load_model(
        DiT,
        v1_base_cfg,
        str(cached_path("hf://SWivid/F5-TTS/F5TTS_v1_Base/model_1250000.safetensors")),
        vocab_file=str(cached_path("hf://SWivid/F5-TTS/F5TTS_v1_Base/vocab.txt")),
    ),
    "v1-small_alg-64h_300k": load_model(
        DiT,
        v1_small_cfg,
        str(cached_path("hf://chenxie95/F5-TTS_v1_Small_Algerian/64h_model_300000.safetensors")),
        vocab_file=alg_vocab_path,
    ),
    "v1-small_alg-64h_300k_no-ema": load_model(
        DiT,
        v1_small_cfg,
        str(cached_path("hf://chenxie95/F5-TTS_v1_Small_Algerian/64h_model_300000_no-ema.safetensors")),
        vocab_file=alg_vocab_path,
    ),
    "v1-small_alg-64h_200k": load_model(
        DiT,
        v1_small_cfg,
        str(cached_path("hf://chenxie95/F5-TTS_v1_Small_Algerian/64h_model_200000.safetensors")),
        vocab_file=alg_vocab_path,
    ),
    "v1-small_alg-64h_200k_no-ema": load_model(
        DiT,
        v1_small_cfg,
        str(cached_path("hf://chenxie95/F5-TTS_v1_Small_Algerian/64h_model_200000_no-ema.safetensors")),
        vocab_file=alg_vocab_path,
    ),
}


@spaces.GPU
def infer(
    ref_audio_orig,
    ref_text,
    gen_text,
    model,
    seed,
    show_info=gr.Info,
):
    if not ref_audio_orig or not ref_text.strip() or not gen_text.strip():
        gr.Warning("Please ensure [Reference Audio] [Reference Text] [Text to Generate] are all provided.")
        return gr.update(), ref_text, seed

    if seed < 0 or seed > 2**31 - 1:
        gr.Warning("Please set a seed in range 0 ~ 2**31 - 1.")
        seed = np.random.randint(0, 2**31 - 1)
    torch.manual_seed(seed)
    used_seed = seed

    ref_audio, ref_text = preprocess_ref_audio_text(ref_audio_orig, ref_text, show_info=show_info)

    final_wave, final_sample_rate, _ = infer_process(
        ref_audio,
        ref_text,
        gen_text,
        tts_model_collections[tts_model_choice],
        vocoder,
        show_info=show_info,
        progress=gr.Progress(),
    )

    return (final_sample_rate, final_wave), ref_text, used_seed


with gr.Blocks() as app_basic_tts:
    with gr.Row():
        with gr.Column():
            ref_wav_input = gr.Audio(label="Reference Audio", type="filepath")
            ref_txt_input = gr.Textbox(label="Reference Text")
            gen_txt_input = gr.Textbox(label="Text to Generate")
            generate_btn = gr.Button("Synthesize", variant="primary")
            with gr.Row():
                randomize_seed = gr.Checkbox(
                    label="Randomize Seed",
                    info="Check to use a random seed for each generation. Uncheck to use the seed specified.",
                    value=True,
                    scale=3,
                )
                seed_input = gr.Number(show_label=False, value=0, precision=0, scale=1)
        audio_output = gr.Audio(label="Synthesized Audio")

    def basic_tts(
        ref_wav_input,
        ref_txt_input,
        gen_txt_input,
        randomize_seed,
        seed_input,
    ):
        if randomize_seed:
            seed_input = np.random.randint(0, 2**31 - 1)

        audio_out, ref_text_out, used_seed = infer(
            ref_wav_input,
            ref_txt_input,
            gen_txt_input,
            tts_model_choice,
            seed_input,
        )
        return audio_out, ref_text_out, used_seed

    ref_wav_input.clear(
        lambda: gr.update(),
        None,
        ref_txt_input,
    )

    generate_btn.click(
        basic_tts,
        inputs=[
            ref_wav_input,
            ref_txt_input,
            gen_txt_input,
            randomize_seed,
            seed_input,
        ],
        outputs=[audio_output, ref_txt_input, seed_input],
    )


with gr.Blocks() as demo:
    gr.Markdown(
        """
        # 🗣️ F5-TTS Online Demo for Dev Test
        
        Upload or record a reference voice, give its transcription text, then order the text to generate and have fun!
        """
    )

    def switch_tts_model(new_choice):
        global tts_model_choice
        tts_model_choice = new_choice

    choose_tts_model = gr.Dropdown(
        choices=[t for t in tts_model_collections], label="Choose TTS Model", value=tts_model_choice
    )

    choose_tts_model.change(
        switch_tts_model,
        inputs=[choose_tts_model],
    )

    gr.TabbedInterface(
        [app_basic_tts],
        ["Basic-TTS"],
    )


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