はじめに

この記事は、MIDIデータをCSVに書き出すために作ったプログラムです。
単に全データを書き出すだけではなく、

拍子・テンポ情報
ノート情報
マーカー情報
テキスト情報

などを分け、わかりやすく保存することを目的としています。MIDIデータを扱うアプリケーションを開発するときの仲介としての位置づけです。

特に、MIDIデルタ時間（プログラム内では「拍子時間」と呼んでいます）から絶対時間（～～秒）の計算をあらかじめ行うことで、アプリケーション側でその計算を行わなくてよいことが利点です。

実装

以下にプログラムを公開しました。

長いですが、メインの処理は以下のような実装になっています。

import py_midicsv as pm
import argparse
import pandas as pd
import os


def midi_to_csv(input_path):
    csv_string = pm.midi_to_csv(input_path)
    df = pd.DataFrame([data.strip().split(", ") for data in csv_string]).map(lambda x: x.replace('"', '') if type(x) == str else x)
    return df


def get_time(x):
    m = int(x // 60)
    s = x - (m * 60)
    s_f = int(s // 1)
    ss = int((s - s_f) * 100)
    return f"{m:0d}:{s_f:02d}:{ss:02d}"


def main(args):
    df = midi_to_csv(args.input)
    os.makedirs(args.output_dir, exist_ok=True)
    
    # TEMP
    df.to_csv(os.path.join(args.output_dir, "all.csv"))
    df[1] = df[1].astype(int)

    d = 0
    beat_n, beat_b = 0, 0
    tempo = 0.0

    t_now_d = 0 
    t_now_s = 0.0
    t_d_1bar = 0
    bar = 0

    bars_data = []

    # Extract data except notes
    for idx, row in df.iterrows():
        # ===== Header =====
        if row[2] == "Header":
            d = int(row[5])
            continue
        
        elif row[2] == "Title_t":
            title = str(row[3])
            continue
        
        # ===== Content =====
        if row[2] == "Tempo":
            tempo = float(60 / (int(row[3]) * (10 ** -6)))
        
        elif row[2] == "Time_signature":
            if t_d_1bar != 0:
                t_delta = int(row[1]) - t_now_d
                bars = t_delta // t_d_1bar
                for i in range(bars):
                    bar += 1
                    bars_data.append([bar, t_now_d + i * t_d_1bar, beat_n, beat_b, t_d_1bar])
            
            beat_n = int(row[3])
            beat_b = int(2 ** int(row[4]))
            t_d_1bar = int(d / (beat_b / 4) * beat_n)
            t_now_d = int(row[1])
            
        elif row[2] == "End_track":
            if t_d_1bar != 0:
                t_delta = int(row[1]) - t_now_d
                bars = t_delta // t_d_1bar
                for i in range(bars + 10):
                    bar += 1
                    bars_data.append([bar, t_now_d + i * t_d_1bar, beat_n, beat_b, t_d_1bar])
            
            t_now_d = int(row[1])
            break

    df_tempo = df[df[2] == "Tempo"].copy()
    df_tempo["v"] = df_tempo[3].apply(lambda x: round(60 / (int(x) * (10 ** -6)), 2))
    df_tempo = df_tempo[[1, "v"]].reset_index(drop=True)
    df_tempo["next_change"] = list(df_tempo[1].values[1:]) + [9999999]
    df_tempo["next_change"] = df_tempo["next_change"].astype(int)
    df_tempo[1] = df_tempo[1].astype(int)

    bars_tempo_data = []
    tempo = 0.0
    t_now_d = 0
    chou = 0
    major = 0

    for idx, tempo_data in df_tempo.iterrows():    
        t_now_d = int(tempo_data[1])
        tempo = tempo_data["v"]
        next_change = tempo_data["next_change"]
        
        if not t_now_d in [data[1] for data in bars_data]:
            for i in range(len(bars_data)):
                if bars_data[i][1] > t_now_d:
                    bar_t = i - 1
                    break
            
            d_tick = bars_data[bar_t][4] / bars_data[bar_t][2]
            delta = t_now_d - bars_data[bar_t][1]
            ticks = int(delta // d_tick) + 1
            d_ = int(delta - ((ticks - 1) * d_tick))
            b_0 = bars_data[bar_t][0]
            b_1 = ticks
            bars_tempo_data.append([f"{b_0}.{b_1}.{d_}", t_now_d, *bars_data[bar_t][2:], tempo, chou, major])
        
        for bar_data in bars_data:
            if t_now_d <= bar_data[1] < next_change:
                if len(df[(df[1] == int(bar_data[1])) & (df[2] == "Key_signature")]) == 1:
                    chou = int(df[(df[1] == int(bar_data[1])) & (df[2] == "Key_signature")][3].values[0])
                    major = df[(df[1] == int(bar_data[1])) & (df[2] == "Key_signature")][4].values[0]
                bars_tempo_data.append([f"{bar_data[0]}.1.0", *bar_data[1:], tempo, chou, major])
        
    bars_tempo_time_data = []
    t_now_s = 0.0

    for i in range(len(bars_tempo_data)):
        if i == 0:
            bars_tempo_time_data.append([*bars_tempo_data[i], 0.0])
            continue
        
        d_delta = bars_tempo_data[i][1] - bars_tempo_data[i - 1][1]
        tempo = bars_tempo_data[i - 1][5]
        
        haku = d_delta / d
        t_s_per_4 = 60.0 / tempo
        time = t_s_per_4 * haku
        t_now_s += time
        
        bars_tempo_time_data.append([*bars_tempo_data[i], t_now_s])
            
    df_hyousi_and_tempo = pd.DataFrame(bars_tempo_time_data, columns=["拍子位置", "時間", "拍子分子", "拍子分母", "1小節長さ", "テンポ", "調", "長調/短調", "絶対時間"])
    df_hyousi_and_tempo.to_csv(os.path.join(args.output_dir, "beats_and_tempo.csv"))

    # Calc time
    def get_times_t_d(t_d):
        j_bar = 0
        j = 0
        for j in range(len(bars_tempo_time_data)):
            if bars_tempo_time_data[j][1] > t_d:
                j -= 1
                j_bar = j
                while not ".1.0" in bars_tempo_time_data[j_bar][0]:
                    j_bar -= 1
                break

        d_delta = t_d - bars_tempo_time_data[j][1]
        tempo = bars_tempo_time_data[j][5]
        haku = d_delta / d
        t_s_per_4 = 60.0 / tempo
        time_s = t_s_per_4 * haku

        t_zettai =  bars_tempo_time_data[j][8] + time_s
        
        bar = bars_tempo_time_data[j_bar][0].replace(".1.0", "")
        d_delta = t_d - bars_tempo_time_data[j_bar][1]
        d_haku = bars_tempo_time_data[j_bar][4] / bars_tempo_time_data[j_bar][2]
        haku = int(d_delta // d_haku)
        remain = int(d_delta - (d_haku * haku))
        
        t_hyousi = f"{bar}.{haku+1}.{remain}"
        return t_zettai, t_hyousi


    # Extract notes data
    notes_data = []

    for _, row in df[df[2] == "Note_on_c"].iterrows():
        t_note_on = int(row[1])
        channel = int(row[3])
        pitch = int(row[4])

        for _, row2 in df[df[1] > t_note_on].iterrows():
            if int(row2[1]) <= t_note_on:
                continue
            elif row2[2] == "Note_off_c" and int(row2[3]) == channel and int(row2[4]) == pitch:
                t_note_off = int(row2[1])
                break

        t_zettai_note_on, hyousi_note_on = get_times_t_d(t_note_on)
        t_zettai_note_off, hyousi_note_off = get_times_t_d(t_note_off)

        # Define pitch display text
        pi_dict_normal = {0: "C", 1: "C#", 2: "D", 3: "Eb", 4: "E", 5: "F", 6: "F#", 7: "G", 8: "G#", 9: "A", 10: "Bb", 11: "B"}
        pi_dict_sharp = {0: "C", 1: "C#", 2: "D", 3: "D#", 4: "E", 5: "F", 6: "F#", 7: "G", 8: "G#", 9: "A", 10: "A#", 11: "B"}
        pi_dict_flat = {0: "C", 1: "Db", 2: "D", 3: "Eb", 4: "E", 5: "F", 6: "Gb", 7: "G", 8: "Ab", 9: "A", 10: "Bb", 11: "B"}
        oc = pitch // 12 - 1

        _chou = df_hyousi_and_tempo[df_hyousi_and_tempo["時間"] < t_note_on].iloc[-1]["調"]
        if _chou > 0:
            pi_dict = pi_dict_sharp
        elif _chou < 0:
            pi_dict = pi_dict_flat
        else:
            pi_dict = pi_dict_normal

        pi = pi_dict[pitch % 12]

        notes_data.append([t_note_on, t_note_off, hyousi_note_on, hyousi_note_off, t_zettai_note_on, t_zettai_note_off, oc, pi, pitch, channel])
    
    df_notes = pd.DataFrame(notes_data)
    channels = sorted(df_notes[9].unique()[:4])

    df_notes[9] = df_notes[9].apply(lambda c: channels.index(c))
    df_notes.to_csv(os.path.join(args.output_dir, "notes.csv"))


    # Extract text data
    text_data = []
    for _, row in df[df[2] == "Text_t"].iterrows():
        t = int(row[1])
        text = row[3]
        t_zettai, hyousi = get_times_t_d(t)
        text_data.append([t, t_zettai, get_time(t_zettai), hyousi, text])
    pd.DataFrame(text_data).to_csv(os.path.join(args.output_dir, "texts.csv"))

    # Extract marker data
    text_data = []
    for _, row in df[df[2] == "Marker_t"].iterrows():
        t = int(row[1])
        text = row[3]
        t_zettai, hyousi = get_times_t_d(t)
        text_data.append([t, t_zettai, get_time(t_zettai), hyousi, text])
    pd.DataFrame(text_data).to_csv(os.path.join(args.output_dir, "markers.csv"))


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description='MIDI to CSV utility based on py_midicsv library')
    parser.add_argument('-i', '--input', required=True, help="Input path")
    parser.add_argument('-o', '--output_dir', default="output", help="Output directory")
    args = parser.parse_args()
    try:
        main(args)
        print("Converted.")
    except Exception as e:
        print(f"Error has occered: {e}")

実行結果のサンプル

以下のようなCSVデータが得られます。

beats_and_tempo.csv：各小節の始めの位置、変更したテンポの位置などが格納されます。

,拍子位置,時間(秒),拍子分子,拍子分母,1小節長さ,テンポ,調,長調/短調,絶対時間(秒)
0,1.1.0,0,4,4,1920,175.0,-6,major,0.0
1,2.1.0,1920,4,4,1920,175.0,-6,major,1.3714285714285714
2,3.1.0,3840,4,4,1920,175.0,-6,major,2.742857142857143
3,4.1.0,5760,4,4,1920,175.0,-6,major,4.114285714285714
4,5.1.0,7680,4,4,1920,175.0,-6,major,5.485714285714286
5,6.1.0,9600,4,4,1920,175.0,-6,major,6.857142857142858
6,7.1.0,11520,4,4,1920,175.0,-6,major,8.22857142857143
7,8.1.0,13440,4,4,1920,175.0,-6,major,9.600000000000001
8,9.1.0,15360,4,4,1920,175.0,-6,major,10.971428571428573

markers.csv：小節単位のマーカー情報が格納されます。

,拍子位置,時間(秒),時間,拍子位置,マーカー
0,5760,4.114285714285714,0:04:11,4.1.0,I
1,13440,9.600000000000001,0:09:60,8.1.0,I-A1
2,28800,20.57142857142857,0:20:57,16.1.0,A1
3,59520,42.51428571428572,0:42:51,32.1.0,B1
4,74880,53.48571428571431,0:53:48,40.1.0,S1
5,117120,83.65714285714283,1:23:65,62.1.0,S1-A2
6,128640,91.88571428571423,1:31:88,68.1.0,A2
7,144000,102.85714285714276,1:42:85,76.1.0,B2

notes.csv：ノート情報が格納されます。

,Note ON拍子位置,Note OFF拍子位置,Note ON時間(秒),Note OFF時間(秒),オクターブ,音階,音高,チャンネル
0,5760,6235,4.1.0,4.1.475,4.114285714285714,4.453571428571428,5,Bb,82,0
1,6240,6478,4.2.0,4.2.238,4.457142857142856,4.627142857142857,5,Eb,75,0
2,6720,7195,4.3.0,4.3.475,4.8,5.139285714285714,5,Ab,80,0
3,7200,7438,4.4.0,4.4.238,5.142857142857142,5.312857142857142,5,Db,73,0
4,7440,7915,4.4.240,5.1.235,5.314285714285714,5.6535714285714285,5,Ab,80,0
5,7920,8158,5.1.240,5.1.478,5.6571428571428575,5.827142857142857,5,Db,73,0

texts.csv：コード等任意の位置に設定したテキスト情報が格納されます。

,拍子位置,時間(秒),時間,拍子位置,テキスト
0,5760,4.114285714285714,0:04:11,4.1.0,Cdim
1,6480,4.628571428571428,0:04:62,4.2.240,B
2,7440,5.314285714285714,0:05:31,4.4.240,Bbm
3,8400,6.0,0:06:00,5.2.240,Gbm/A
4,9600,6.857142857142858,0:06:85,6.1.0,Abm7