app.py

import numpy as np
import gradio as gr
from pypinyin import lazy_pinyin

from pinyin_dict import PINYIN_DICT

from espnet_model_zoo.downloader import ModelDownloader
from espnet2.fileio.read_text import read_label
from espnet2.bin.svs_inference import SingingGenerate


spks = {
    "singer1 (man)": 1,
    "singer2 (man)": 2,
    "singer3 (female)": 5,
    "singer4 (female)": 9,
    "singer5 (man)": 18,
    "singer6 (female)": 15,
    "singer7 (man)": 23,
    "singer8 (man)": 25,
    "singer9 (female)": 29,
    "singer10 (man)": 27,
}

def gen_song(lang, tempo, texts, durs, pitchs, spk):
    if lang == "zh":
        PRETRAIN_MODEL = "espnet/aceopencpop_svs_visinger2_40singer_pretrain"
        fs = 44100
        text_list = lazy_pinyin(texts)

    # preprocess
    if texts is None:
        return (fs, np.array([0.0])), "Error: No Text provided!"
    if durs is None:
        return (fs, np.array([0.0])), "Error: No Dur provided!"
    if pitchs is None:
        return (fs, np.array([0.0])), "Error: No Pitch provided!"

    dur_list = durs.strip().split()
    pitch_list = pitchs.strip().split()

    if len(text_list) != len(dur_list):
        return (fs, np.array([0.0])), f"Error: len in text({len(text_list)}) mismatch with duration({len(dur_list)})!"
    if len(text_list) != len(pitch_list):
        return (fs, np.array([0.0])), f"Error: len in text({len(text_list)}) mismatch with pitch({len(pitch_list)})!"

    ## text to phoneme
    sybs = []
    if lang == "zh":
        pinyin_dict = PINYIN_DICT
        for text in text_list:
            text = text.lower()
            if text not in pinyin_dict:
                return (fs, np.array([0.0])), f"Error: pinyin `{text}` is invalid!"
            phns = "_".join(pinyin_dict[text])
            sybs.append(phns)
            
    ## pitch
    pitch_dict = {}
    with open("./midi-note.scp", "r", encoding="utf-8") as f:
        for line in f:
            items = line.strip().split()
            pitch_dict[items[0]] = int(items[1])
            pitch_dict[items[1]] = int(items[1])

    labels = []
    notes = []
    st = 0
    for phns, dur, pitch in zip(sybs, dur_list, pitch_list):
        if pitch not in pitch_dict:
            return (fs, np.array([0.0])), f"Error: pitch `{pitch}` is invalid!"
        pitch = pitch_dict[pitch]
        dur = float(dur)
        phn_list = phns.split("_")
        lyric = "".join(phn_list)
        note = [st, st + dur, lyric, pitch, phns]
        st += dur
        notes.append(note)
        for phn in phn_list:
            labels.append(phn)

    phns_str = " ".join(labels)
    batch = {
        "score": (
            int(tempo),
            notes,
        ),
        "text": phns_str,
    }
    
    # Infer
    device = "cpu"
    # device = "cuda" if torch.cuda.is_available() else "cpu"
    d = ModelDownloader()
    pretrain_downloaded = d.download_and_unpack(PRETRAIN_MODEL)
    svs = SingingGenerate(
        train_config = pretrain_downloaded["train_config"],
        model_file = pretrain_downloaded["model_file"],
        device = device
    )
    sid = spks[spk]
    output_dict = svs(batch, sids=np.array([sid]))
    wav_info = output_dict["wav"].cpu().numpy()
    return (fs, wav_info), "success!"


title = "Demo of Singing Voice Synthesis in Muskits-ESPnet"

description = """
This is the demo page of our toolkit <b>Muskits-ESPnet: A Comprehensive Toolkit for Singing Voice Synthesis in New Paradigm<b>.


<p>How to use:</p>
<ol>
  <li> Choose language ID. Language id </li>
  <li> Input tempo in integer </li>
  <li> Input text, duration, pitch of equal length </li>
  <li> Choose ons singer </li>
  <li> Click submit button </li>
</ol>


"""

article = """
<div style='margin:20px auto;'>

<p>References: <a href="https://arxiv.org/abs/2409.07226">Muskits-ESPnet paper</a> |
<a href="https://github.com/espnet/espnet">espnet GitHub</a> |
<a href="https://huggingface.co/espnet/aceopencpop_svs_visinger2_40singer_pretrain">pretrained model</a></p>

<pre>
@inproceedings{wu2024muskits,
  title = {{Muskits-ESPnet}: A Comprehensive Toolkit for Singing Voice Synthesis in New Paradigm},
  author = {Yuning Wu and Jiatong Shi and Yifeng Yu and Yuxun Tang and Tao Qian and Yueqian Lin and Jionghao Han and Xinyi Bai and Shinji Watanabe and Qin Jin},
  booktitle={Proc. ACM Multimedia},
  year={2024},
}
</pre>

</div>
"""


# SP: silence, AP: aspirate.
examples = [
    ["zh", 89, "雨淋湿了SP天空AP", "0.23 0.16 0.36 0.16 0.07 0.28 0.50 0.21", "60 62 62 62 0 62 58 0", "singer1 (man)"],
    ["zh", 89, "雨淋湿了SP天空AP", "0.23 0.16 0.36 0.16 0.07 0.28 0.50 0.21", "C4 D4 D4 D4 rest D4 A#3 rest", "singer2 (man)"],
    ["zh", 89, "雨淋湿了SP天空AP", "0.23 0.16 0.36 0.16 0.07 0.28 0.50 0.21", "C4 D4 D4 D4 rest D4 Bb3 rest", "singer3 (female)"],
]

gr.Interface(
    fn=gen_song,
    inputs=[
        gr.Radio(label="language", choices=["zh"], value="zh"),
        gr.Textbox(label="Tempo"),
        gr.Textbox(label="Text"),
        gr.Textbox(label="Duration"),
        gr.Textbox(label="Pitch"),
        gr.Radio(
            label="Singer",
            choices=[
                "singer1 (man)",
                "singer2 (man)",
                "singer3 (female)",
                "singer4 (female)",
                "singer5 (man)",
                "singer6 (female)",
                "singer7 (man)",
                "singer8 (man)",
                "singer9 (female)",
                "singer10 (man)",
            ],
            value="singer1 (man)"
        ),
    ],
    outputs=[
        gr.Audio(label="Generated Song", type="numpy"),
        gr.Textbox(label="Running Status"),
    ],
    title=title,
    description=description,
    article=article,
    examples=examples,
).launch()