diff --git a/experiment/bot.py b/experiment/bot.py
new file mode 100644
index 0000000..f5977b7
--- /dev/null
+++ b/experiment/bot.py
@@ -0,0 +1,115 @@
+import warnings
+from datetime import datetime
+from pathlib import Path
+
+import matplotlib
+from telegram import Update, Message
+from telegram.ext import Updater, CallbackContext, Dispatcher, CommandHandler, MessageHandler, \
+    Filters
+
+from ina_main import *
+
+warnings.filterwarnings("ignore")
+matplotlib.use('agg')
+
+
+def r(u: Update, msg: str, md=True):
+    updater.bot.sendMessage(chat_id=u.effective_chat.id, text=msg,
+                            parse_mode='Markdown' if md else None)
+
+
+def cmd_start(u: Update, c: CallbackContext):
+    r(u, '欢迎! 点下面的录音按钮就可以开始啦w')
+
+
+def process_audio(message: Message):
+    # Only when replying to voice or audio
+    audio = message.audio or message.voice
+    if not audio:
+        return
+
+    # Download audio file
+    date = datetime.now().strftime('%Y-%m-%d %H-%M')
+    try:
+        downloader = bot.getFile(audio.file_id)
+    except:
+        downloader = bot.getFile(audio.file_id)
+    file = Path(tmpdir).joinpath(f'{date} {message.from_user.name[1:]}.mp3')
+    print(downloader, '->', file)
+    downloader.download(file)
+
+    # Segment file
+    result = segment(file)
+
+    # Null case
+    print(result)
+    if len(result) == 0:
+        bot.send_message(message.chat_id, '分析失败, 大概是音量太小或者时长太短吧, 再试试w')
+        return
+
+    # Draw results
+    with draw_result(str(file), result) as buf:
+        f, m, o, pf = get_result_percentages(result)
+        msg = f"分析结果: {f*100:.0f}% 🙋‍♀️ | {m*100:.0f}% 🙋‍♂️ | {o*100:.0f}% 🚫\n" \
+              f"(结果仅供参考, 如果结果不是你想要的，那就是模型的问题，欢迎反馈)\n" \
+              f"" \
+              f"(因为这个模型基于法语数据, 和中文发音习惯有差异, 所以这个识别结果可能不准)"
+        bot.send_photo(message.chat_id, photo=buf, caption=msg,
+                       reply_to_message_id=message.message_id)
+
+
+def cmd_analyze(u: Update, c: CallbackContext):
+    reply = u.effective_message.reply_to_message
+
+    # Parse command
+    text = u.effective_message.text
+    if not text:
+        return
+    cmd = text.lower().split()[0].strip()
+
+    if cmd[0] not in '!/':
+        return
+    cmd = cmd[1:]
+
+    if cmd not in ['analyze', 'analyze-raw']:
+        return
+
+    if cmd == 'analyze-raw':
+        raw = True
+
+    if u.effective_user.id == reply.from_user.id:
+        process_audio(reply)
+    else:
+        r(u, '只有自己能分析自己的音频哦 👀')
+
+
+def on_audio(u: Update, c: CallbackContext):
+    process_audio(u.effective_message)
+
+
+if __name__ == '__main__':
+    tmpdir = Path('audio_tmp')
+    tmpdir.mkdir(exist_ok=True, parents=True)
+
+    # Find telegram token
+    path = Path(os.path.abspath(__file__)).parent
+    db_path = path.joinpath('voice-bot-db.json')
+    if 'tg_token' in os.environ:
+        tg_token = os.environ['tg_token']
+    else:
+        with open(path.joinpath('voice-bot-token.txt'), 'r', encoding='utf-8') as f:
+            tg_token = f.read().strip()
+
+    # Telegram login
+    updater = Updater(token=tg_token, use_context=True)
+    dispatcher: Dispatcher = updater.dispatcher
+    bot = updater.bot
+
+    dispatcher.add_handler(CommandHandler('start', cmd_start, filters=Filters.chat_type.private))
+    dispatcher.add_handler(CommandHandler('analyze', cmd_analyze, filters=Filters.reply))
+    dispatcher.add_handler(MessageHandler(Filters.reply, cmd_analyze))
+    dispatcher.add_handler(MessageHandler(Filters.voice & Filters.chat_type.private, on_audio))
+    dispatcher.add_handler(MessageHandler(Filters.audio & Filters.chat_type.private, on_audio))
+
+    print('Starting bot...')
+    updater.start_polling()
diff --git a/experiment/ina_main.py b/experiment/ina_main.py
new file mode 100644
index 0000000..5032e82
--- /dev/null
+++ b/experiment/ina_main.py
@@ -0,0 +1,191 @@
+from __future__ import annotations
+
+import io
+import os
+import subprocess
+import tempfile
+import time
+import warnings
+from subprocess import Popen, PIPE
+from typing import NamedTuple, Callable
+
+import matplotlib.pyplot as plt
+import numpy as np
+import scipy.io.wavfile
+from PIL import Image
+from inaSpeechSegmenter import *
+from matplotlib.axes import Axes
+from matplotlib.figure import Figure
+import tensorflow as tf
+
+
+gpus = tf.config.experimental.list_physical_devices('GPU')
+for gpu in gpus:
+    tf.config.experimental.set_memory_growth(gpu, True)
+
+seg = Segmenter()
+
+
+class ResultFrame(NamedTuple):
+    gender: str
+    start: float
+    end: float
+    prob: float
+
+
+class Result(NamedTuple):
+    frames: list[ResultFrame]
+    file: str
+
+
+def segment(file) -> list[ResultFrame]:
+    return [ResultFrame(*s) for s in seg(file)]
+
+
+def to_wav(file: str, callback: Callable, start_sec: float = 0, stop_sec: float = 0):
+    """
+    Convert media to temp wav 16k file and return features
+    """
+    base, _ = os.path.splitext(os.path.basename(file))
+
+    with tempfile.TemporaryDirectory() as tmpdir_name:
+        # build ffmpeg command line
+        tmp_wav = tmpdir_name + '/' + base + '.wav'
+        args = ['ffmpeg', '-y', '-i', file, '-ar', '16000', '-ac', '1']
+
+        if start_sec != 0:
+            args += ['-ss', '%f' % start_sec]
+        if stop_sec != 0:
+            args += ['-to', '%f' % stop_sec]
+
+        args += [tmp_wav]
+
+        # launch ffmpeg
+        p = Popen(args, stdout=PIPE, stderr=PIPE)
+        output, error = p.communicate()
+        assert p.returncode == 0, error
+
+        return callback(tmp_wav)
+
+
+def show_image_buffer(buf):
+    im = Image.open(buf)
+    im.show()
+    buf.close()
+
+
+def draw_result(file: str, result: list[ResultFrame]):
+    """
+    Draw segmentation result
+
+    :param file: Audio file
+    :param result: Segmentation result
+    :return: Result image in bytes (please close it after use)
+    """
+    def wav_callback(wavfile: str):
+        sample_rate, audio = scipy.io.wavfile.read(wavfile)
+        _time = np.linspace(0, len(audio) / sample_rate, num=len(audio))
+
+        fig: Figure = plt.gcf()
+        ax: Axes = plt.gca()
+
+        # Plot audio
+        plt.plot(_time, audio, color='white')
+
+        # Set size
+        # fig.set_dpi(400)
+        fig.set_size_inches(18, 6)
+
+        # Cutoff frequency so that the plot looks centered
+        cutoff = min(abs(min(audio)), abs(max(audio)))
+        ax.set_ylim([-cutoff, cutoff])
+        ax.set_xlim([result[0].start, result[-1].end])
+
+        # Draw segmentation areas
+        colors = {'female': '#F5A9B8', 'male': '#5BCEFA', 'default': 'gray'}
+        for r in result:
+            color = colors[r.gender] if r.gender in colors else colors['default']
+            ax.axvspan(r.start, r.end - 0.01, alpha=.5, color=color)
+
+        # Savefig to bytes
+        buf = io.BytesIO()
+        plt.axis('off')
+        plt.savefig(buf, bbox_inches='tight', pad_inches=0, transparent=False)
+        buf.seek(0)
+        plt.clf()
+        plt.close()
+        return buf
+
+    return to_wav(file, wav_callback)
+
+
+def get_result_percentages(result: list[ResultFrame]) -> tuple[float, float, float, float]:
+    """
+    Get percentages
+
+    :param result: Result
+    :return: %female, %male, %other, %female-vs-female+male
+    """
+    # Count total and categorical durations
+    total_dur = 0
+    durations: dict[str, int] = {f.gender: 0 for f in result}
+    for f in result:
+        dur = f.end - f.start
+        durations[f.gender] += dur
+        total_dur += dur
+
+    # Convert durations to ratios
+    for d in durations:
+        durations[d] /= total_dur
+
+    # Return results
+    f = durations.get('female', 0)
+    m = durations.get('male', 0)
+
+    fm_total = f + m
+    pf = 0 if fm_total == 0 else f / fm_total
+
+    return f, m, 1 - f - m, pf
+
+
+def test():
+    # results: BatchResults = BatchResults(
+    #     [Result([ResultFrame('female', 0.0, 10.48), ResultFrame('male', 10.48, 12.780000000000001)],
+    #             '../test.csv')],
+    #     1.7032792568206787, 1.7032792568206787, 1,
+    #     [('../test.csv', 0)])
+
+    warnings.filterwarnings("ignore")
+    audio_file = '../test.flac'
+
+    # Warmup run
+    results = segment(audio_file)
+    print(results)
+
+    # # Actual run
+    # results = process(seg, ['../test.flac'])
+    # print(results)
+
+    # Benchmark
+    # iterations = 60
+    # total_time = 0
+    # audio_len = float(subprocess.getoutput(f'ffprobe -i {audio_file} -show_entries format=duration -v quiet -of csv="p=0"'))
+    # print(f'Audio length: {audio_len}')
+    #
+    # for i in range(iterations):
+    #     results = process(seg, [audio_file])
+    #     total_time += results.time_full
+    #
+    # time_per_second = total_time / iterations / audio_len
+    # print(f'Benchmark result: {total_time}s / {iterations} iterations = {time_per_second} seconds of processing per second in audio')
+    # print(f'Score: {1 / time_per_second}')
+
+    # Draw results
+    # with draw_result(audio_file, results.results[0]) as buf:
+    #     show_image_buffer(buf)
+    #     print(get_result_percentages(results.results[0]))
+
+
+if __name__ == '__main__':
+    test()
+    pass
diff --git a/experiment/utils.py b/experiment/utils.py
new file mode 100644
index 0000000..bbd421f
--- /dev/null
+++ b/experiment/utils.py
@@ -0,0 +1,49 @@
+def ansi_rgb(r: int, g: int, b: int, foreground: bool = True) -> str:
+    """
+    Convert rgb color into ANSI escape code format
+
+    :param r:
+    :param g:
+    :param b:
+    :param foreground: Whether the color applies to forground
+    :return: Escape code
+    """
+    c = '38' if foreground else '48'
+    return f'\033[{c};2;{r};{g};{b}m'
+
+
+def color(msg: str) -> str:
+    """
+    Replace extended minecraft color codes in string
+
+    :param msg: Message with minecraft color codes
+    :return: Message with escape codes
+    """
+    replacements = ["&0/\033[0;30m", "&1/\033[0;34m", "&2/\033[0;32m", "&3/\033[0;36m", "&4/\033[0;31m", "&5/\033[0;35m", "&6/\033[0;33m", "&7/\033[0;37m", "&8/\033[1;30m", "&9/\033[1;34m", "&a/\033[1;32m", "&b/\033[1;36m", "&c/\033[1;31m", "&d/\033[1;35m", "&e/\033[1;33m", "&f/\033[1;37m", "&r/\033[0m", "&n/\n"]
+    for r in replacements:
+        msg = msg.replace(r[:2], r[3:])
+
+    while '&gf(' in msg or '&gb(' in msg:
+        i = msg.index('&gf(') if '&gf(' in msg else msg.index('&gb(')
+        end = msg.index(')', i)
+        code = msg[i + 4:end]
+        fore = msg[i + 2] == 'f'
+
+        if code.startswith('#'):
+            rgb = tuple(int(code.lstrip('#')[i:i+2], 16) for i in (0, 2, 4))
+        else:
+            code = code.replace(',', ' ').replace(';', ' ').replace('  ', ' ')
+            rgb = tuple(int(c) for c in code.split(' '))
+
+        msg = msg[:i] + ansi_rgb(*rgb, foreground=fore) + msg[end + 1:]
+
+    return msg
+
+
+def printc(msg: str):
+    """
+    Print with color
+
+    :param msg: Message with minecraft color codes
+    """
+    print(color(msg + '&r'))
diff --git a/test.flac b/test.flac
deleted file mode 100644
index 1efcdb2..0000000
Binary files a/test.flac and /dev/null differ