这是一个基于深度学习的音频音高检测系统，可以将音频文件转换为乐谱。该项目复现了CREPE音高检测模型，并集成了音乐理论分析功能。
- 音频音高实时检测
- MIDI文件生成
- 乐谱可视化
- 支持多种音频格式
- 集成MuseScore显示功能
- Python 3.7+
- PyTorch 1.7+
- music21
- librosa
- numpy
- scipy
- MuseScore 4 (用于乐谱显示)
- 克隆仓库:
git clone https://github.com/yourusername/pitch-detection.git
cd pitch-detection- 安装依赖:
pip install -r requirements.txt- 安装MuseScore 4 并配置环境变量
- 基础音高检测:
from mymodel import create_score
# 创建乐谱
score = create_score("input.wav", "model.pt")
# 保存MIDI文件
score.write('midi', fp='output.mid')- 使用演示脚本:
python demo.py --input audio.wav --output score.midpitch-detection/
├── mymodel/
│ ├── __init__.py
│ ├── data_load.py # 数据加载模块
│ └── test_demo.py # 测试演示模块
├── CREPE.py # CREPE模型实现
├── demo.py # 演示脚本
└── README.md
音高检测核心功能:
def create_score(audio_path, model_path, confidence_threshold=0.5, min_note_length=0.1):
"""使用训练好的模型创建乐谱"""
print("加载模型...")
model, device = load_model(model_path, model_size='tiny') # 确保使用tiny模型
print("加载音频文件...")
# 进行音高检测
print("正在进行音高检测...")
time, frequency, confidence = process_audio(audio_path, model, device)
# 创建音乐流
s = stream.Stream()
s.insert(0, tempo.MetronomeMark(number=120))
s.append(meter.TimeSignature('4/4'))
# 处理音高序列
current_note = None
current_start = None
pending_notes = []
print("处理检测到的音高...")
print(f"检测到的频率范围: {frequency.min():.2f}Hz - {frequency.max():.2f}Hz")
print(f"置信度范围: {confidence.min():.2f} - {confidence.max():.2f}")
# 应用中值滤波来平滑频率序列
window_size = 5
frequency = np.pad(frequency, (window_size//2, window_size//2), mode='edge')
frequency = np.array([np.median(frequency[i:i+window_size])
for i in range(len(frequency)-window_size+1)])
for i in range(len(time)):
if confidence[i] >= confidence_threshold:
midi_note = hz_to_midi_note(frequency[i])
if midi_note is not None and 21 <= midi_note <= 108: # 限制在钢琴音域内
if current_note is None:
# 开始新音符
current_note = midi_note
current_start = time[i]
elif abs(current_note - midi_note) >= 0.5: # 音高变化超过半音
# 结束当前音符
duration = time[i] - current_start
if duration >= min_note_length:
pending_notes.append({
'pitch': int(round(current_note)),
'start_time': current_start,
'duration': duration
})
# 开始新音符
current_note = midi_note
current_start = time[i]
elif current_note is not None:
# 结束当前音符
duration = time[i] - current_start
if duration >= min_note_length:
pending_notes.append({
'pitch': int(round(current_note)),
'start_time': current_start,
'duration': duration
})
current_note = None数据处理部分:
if num_segments > max_segments_per_file:
step = num_segments // max_segments_per_file
segment_indices = range(0, num_segments, step)
else:
segment_indices = range(num_segments)
# 添加每个片段的索引信息
for i in segment_indices:
total_segments += 1
start_sample = i * self.hop_length
start_frame = start_sample // self.hop_length
num_frames = self.segment_length // self.hop_length
# 检查这个片段是否有有效的音高标签
if start_frame + num_frames <= len(pitch_labels):
pitch_segment = pitch_labels[start_frame:start_frame + num_frames]
pitch_stats['debug_info']['total_pitch_segments'] += 1
# 确保片段长度足够长（至少32帧）
if len(pitch_segment) >= 32: # 降低最小帧数要求
valid_pitches = pitch_segment[pitch_segment > 0]
valid_ratio = len(valid_pitches) / len(pitch_segment)
# 更新有效片段计数
pitch_stats['debug_info']['valid_pitch_segments'] += 1
# 放宽筛选条件
if valid_ratio >= 0.1: # 降低到10%
pitch_stats['debug_info']['ratio_check_passed'] += 1
if len(valid_pitches) > 0:
pitch_mean = np.mean(valid_pitches)
if 40 <= pitch_mean <= 1000:
pitch_stats['debug_info']['range_check_passed'] += 1
self.segments.append({
'wav_file': wav_file,
'pitch_file': pitch_file,
'start_sample': start_sample,
'waveform_length': waveform.shape[1],
'valid_ratio': valid_ratio,
'pitch_mean': pitch_mean
})
valid_segments += 1
else:
pitch_stats['filtered_by_range'] += 1
else:
pitch_stats['filtered_by_ratio'] += 1
# 更新音高统计信息
if len(valid_pitches) > 0:
pitch_stats['max_pitch'] = max(pitch_stats['max_pitch'], valid_pitches.max())
pitch_stats['min_pitch'] = min(pitch_stats['min_pitch'], valid_pitches.min())
pitch_stats['pitch_histogram'].extend(valid_pitches.tolist())
pitch_stats['total_frames'] += len(pitch_segment)
pitch_stats['valid_frames'] += len(valid_pitches)
except Exception as e:
print(f"处理文件 {wav_file} 时出错: {str(e)}")
print(f"错误详情: ", e.__class__.__name__)
import traceback
print(traceback.format_exc())
continue
# 打印详细的统计信息
print(f"\n数据集统计信息:")
print(f"总片段数: {total_segments}")
print(f"有效片段数: {valid_segments} ({valid_segments/max(1,total_segments)*100:.1f}%)")
print(f"被有效帧比例过滤掉的片段数: {pitch_stats['filtered_by_ratio']}")
print(f"被音高范围过滤掉的片段数: {pitch_stats['filtered_by_range']}")
print(f"总帧数: {pitch_stats['total_frames']}")
print(f"有效帧数: {pitch_stats['valid_frames']} ({pitch_stats['valid_frames']/max(1,pitch_stats['total_frames'])*100:.1f}%)")
print("\n调试信息:")
print(f"总音高片段数: {pitch_stats['debug_info']['total_pitch_segments']}")
print(f"有效音高片段数: {pitch_stats['debug_info']['valid_pitch_segments']}")
print(f"通过比例检查的片段数: {pitch_stats['debug_info']['ratio_check_passed']}")
print(f"通过范围检查的片段数: {pitch_stats['debug_info']['range_check_passed']}")
if pitch_stats['valid_frames'] > 0:
print(f"音高范围: {pitch_stats['min_pitch']:.1f}Hz - {pitch_stats['max_pitch']:.1f}Hz")
# 打印音高分布统计
hist = np.histogram(pitch_stats['pitch_histogram'],
bins=np.linspace(0, 1000, 21)) # 0-1000Hz分20个区间
print("\n音高分布:")
for i in range(len(hist[0])):
if hist[0][i] > 0:
print(f"{hist[1][i]:.0f}-{hist[1][i+1]:.0f}Hz: {hist[0][i]} 帧")请参考以下编码规范:
<p>There are the following guidelines regarding the code style in files:
</p>
<ul class="itemize mark-bullet">
<li>Indent size is 4.
</li><li>The TAB character is forbidden outside of Makefiles as is any
form of trailing whitespace. Commits containing either will be
rejected by the git repository.
</li><li>You should try to limit your code lines to 80 characters; however, do so if
and only if this improves readability.本项目采用 GNU General Public License v3.0 许可证。详情请参见 LICENSE 文件。
- Fork 项目
- 创建特性分支
- 提交更改
- 推送到分支
- 创建 Pull Request
如有问题或建议，请通过以下方式联系：
- 提交 Issue
- 发送邮件至: your.email@example.com
- CREPE项目团队
- music21库开发者
- FFmpeg社区