build_svp.py

import json
import uuid
import os
import math
from tqdm import tqdm


def build_svp(template: dict, midis: list, f0: list, tempo: int, basename: str, extract_pitch: bool = True) -> str:
    notes = [] # 用于保存的音符数据
    datas = [] # 用于记录的音符数据

    per_dur = 705600000 # 每拍在sv的时长
    per_time = 60 / tempo # 每拍的时间
    template["time"]["tempo"] = [{"position": 0, "bpm": tempo}]

    index = 0
    for midi in tqdm(midis, desc="Generate midi notes"):
        offset = int(f0[index]["offset"] / per_time * per_dur) # 音符的起始时间在sv的时长

        dur = midi["note_dur"] # 音符的时长
        pitch = midi["note_midi"] # 音符的音高
        rest = midi["note_rest"] # 是否为休止符
        midi_duration = 0 # 该段音符的总时长

        for i in range(len(pitch)):

            current_duration = dur[i] / per_time * per_dur # 当前音符在sv的时长
            onset = midi_duration + offset # 音符的起始时间
            midi_duration += int(current_duration)
            
            if rest[i]: # 休止符
                continue
            
            current_pitch = round(pitch[i])
            
            note = {
                "musicalType": "singing",
                "onset": int(onset),
                "duration": int(current_duration),
                "lyrics": "la",
                "phonemes": "",
                "accent": "",
                "pitch": int(current_pitch),
                "detune": 0,
                "instantMode": False,
                "attributes": {"evenSyllableDuration": True},
                "systemAttributes": {"evenSyllableDuration": True},
                "pitchTakes": {"activeTakeId": 0,"takes": [{"id": 0,"expr": 0,"liked": False}]},
                "timbreTakes": {"activeTakeId": 0,"takes": [{"id": 0,"expr": 0,"liked": False}]}
            }

            data = {
                "start": int(onset),
                "finish": int(current_duration + onset),
                "pitch": int(current_pitch)
            }
            notes.append(note)
            datas.append(data)
        index += 1

    template["tracks"][0]["mainGroup"]["notes"] = notes
    template["tracks"][0]["mainGroup"]["uuid"] = str(uuid.uuid4()).lower()
    template["tracks"][0]["mainRef"]["groupID"] = template["tracks"][0]["mainGroup"]["uuid"]

    if extract_pitch:
        pitch = build_pitch(datas, f0, tempo)
    else:
        print("Pitch data is not extracted.")
        pitch = []
    template["tracks"][0]["mainGroup"]["parameters"]["pitchDelta"]["points"] = pitch

    file_path = os.path.join("results", f"{basename}.svp")
    with open(file_path, "w", encoding="utf-8") as f:
        json.dump(template, f)

    return file_path

def build_pitch(datas: list, f0s: list, tempo: int) -> list:
    pitch = [] # 用于保存的音高数据

    per_dur = 705600000 # 每拍在sv的时长
    per_time = 60 / tempo # 每拍的时间
    time_per_frame = 0.01 # 每帧的时间 hop_size / sample_rate
    
    for f0 in tqdm(f0s, desc="Generate Pitchs"):
        offset = f0["offset"]
        f0_data = f0["f0"]
        for i in range(len(f0_data)):
            pitch_onset, pitch_cents = None, None
            f0_value = f0_data[i] # 当前帧的f0值
            if f0_value == 0.0:
                continue
            onset_time = offset + i * time_per_frame # 当前帧的起始时间
            onset = (onset_time / per_time) * per_dur # 当前帧的起始时间在sv的时长
            pitch_onset = onset
            for data in datas:
                if data["start"] <= onset and onset < data["finish"]: # 当前帧在音符的时间范围内
                    pitch_cents = calculate_cents_difference(data["pitch"], f0_value)
                    break
            if pitch_onset is None or pitch_cents is None:
                continue
            pitch.append(pitch_onset)
            pitch.append(pitch_cents)
    return pitch

def calculate_cents_difference(midi_note, f0):
    def midi_to_freq(midi_note):
        A4 = 440.0
        return A4 * (2 ** ((midi_note - 69) / 12))

    def cents_difference(f0, midi_note):
        midi_freq = midi_to_freq(midi_note)
        return 1200 * math.log2(f0 / midi_freq)

    cents_diff = cents_difference(f0, midi_note)
    return round(cents_diff, 5)