TART：支持演奏技巧与把位指法的模块化吉他音频制谱工具

My Perspective on the Research

It's clear that automatic music transcription (AMT) for guitar still faces significant hurdles. Current systems struggle with capturing nuanced techniques like slides, bends, and percussive hits. Incorrect string-fret assignments are also a recurring issue, and the reliance on clean recordings limits generalizability to the real-world noise. Recognizing these limitations, I see the value in proposing TART, a modular, four-stage approach. This pipeline is quite elegant: an audio-to-MIDI transcription model, a dedicated expressive technique classifier, an audio-conditioned T5 encoder-decoder for precise string-fret allocation, and a final automated tablature generator.

I'm particularly interested in the evaluation approach. Testing TART in a zero-shot setting across datasets like GuitarSet, EGDB, and their noisy counterparts (Noisy GuitarSet and Noisy EGDB) provides a rigorous assessment. The performance metrics are promising; achieving 81.35% audio-to-MIDI F50 (a 6.67-point improvement), 71.8% string-fret Tab F1 (an 8.5-point boost), and 54.08% end-to-end Tab F1 are impressive.

Essentially, I see TART as a pioneering framework. It's the first that I'm aware of to directly generate guitar tablature that includes both fingering and expressive technique annotations, all from raw audio input. That's a valuable contribution to the field.

吉他自动音乐转录（AMT）目前仍受制于三大挑战：现有系统往往难以捕捉滑音、推弦和打板敲击等表现力演奏技巧；经常将音符分配到错误的弦品组合；且通常仅在纯净录音上进行训练，限制了其向带噪真实音频的泛化能力。为应对这些挑战，我们提出了 TART——一个模块化的四阶段“音频到吉他谱”处理流程，包括：(1) 音频到 MIDI 转录模型，(2) 演奏技巧分类器，(3) 用于弦品分配的音频条件化 T5 编码器-解码器，以及 (4) 自动化吉他谱生成器。我们在零样本（zero-shot）设置下，于 GuitarSet、EGDB 以及两个增强基准 Noisy GuitarSet 和 Noisy EGDB 上对 TART 进行了评估。在这四个基准上的平均结果显示，TART 取得了 81.35% 的音频到 MIDI F50（比先前最佳基线高出 6.67 个百分点）、71.8% 的弦品 Tab F1（比先前最佳基线高出 8.5 个百分点）以及 54.08% 的端到端 Tab F1。据我们所知，TART 是首个能够直接从吉他音频生成同时包含指法和演奏技巧标注的吉他谱框架。