# From a backing track to a guitar solo / 根据伴奏生成吉他主奏

Reusable case-study presentation guide · case archived 10 September 2026 · approximately 5–6 minutes including brief playback.

Purpose: establish a reusable way to judge generated musical parts: show relatively good and bad examples, identify what succeeds and fails, and trace the exact input-to-output path. The Base/XL pair is the first case, rather than the limit of the method. “Good/bad” below means relatively more/less convincing **within this pair and this guitar task**. It is not a formal adoption score or universal model ranking.

## English oral version

**Page 1 — Listen to the result**

The central question is whether a model can produce a convincing new part over an existing backing, and what makes a result good or bad. This case uses a more convincing example and a less convincing reference, made from the same backing and the same explicit generation settings.

The backing is thirty seconds of bass and drums, with the guitar stems excluded. The task is to create a new foreground guitar part on that same timeline. We are adding a part over the music, rather than continuing after the music ends.

[Play backing only briefly. Select Base, then guitar only, then together. Use the same position when comparing with XL. The first ten seconds are very quiet in both candidates: show this opening before seeking to the more active passage around thirteen seconds.]

For the Base example, I recognise the distorted electric-guitar tone and playing character from my own experience of playing guitar. That is the aspect I consider successful. In the paired XL example, the earlier problems remain, and I find Base more convincing. The XL output still contains an audible melodic line; its being a single-note line is not, by itself, a failure for a guitar solo.

Neither example is a complete success. Both have a very quiet opening, with the main energy arriving around thirteen seconds. We have not yet established whether the isolated part is clean, whether its notes and phrases fit this backing, or whether this works consistently on other inputs. So the positive example shows progress in guitar character; it does not close the accompaniment problem.

The signal analysis adds a second kind of evidence. It confirms the quiet opening and shows different spectral concentration between the outputs. Both files are technically intact. These checks help describe the sounds, but a spectral percentage cannot tell us whether the guitar playing is musically convincing.

**Page 2 — Follow the exact input**

Here is what I actually asked the model to do. The task instruction was: “Generate the GUITAR track based on the audio context.” The caption specified solo lead electric guitar, foreground melodic phrases with pauses, and guitar only, without drums, bass, vocals or other instruments. The full original wording is on the page.

There was no recovery or moderate-energy prompt, and no tempo or key value was supplied. The musical context was the backing audio itself. This run used the direct source-audio route, not codes generated by a music language model.

Inside the backend, a VAE converts the backing waveform into a continuous audio representation. A separate text encoder processes the task and description. The LEGO task gives that source context and the guitar instruction to the Base diffusion transformer, which generates the new candidate representation. The VAE then decodes it to audio. Outside the model, I mix half-gain backing with half-gain candidate for listening. I did not add distortion or edit notes to create the guitar impression.

**Page 3 — What changed, and what remains**

Getting here required checking our earlier assumptions. We retained direct audio, source-derived codes and text-generated codes, but found that non-empty codes take precedence over source audio. The text-generated route had not actually heard the backing. We also removed conflicting mood wording.

Another setting, DCW, had been requested on but skipped because of a dependency failure. Repairing it made the correction run, but the effective ON example was still noisy. That fixed an execution problem without solving the musical task.

Finally, I checked the supported LEGO checkpoints and tried ordinary Base with the matched direct-audio settings. The positive example came from this supported route. A separate audio-understanding call produced uncertain tempo, key and instrument estimates; I did not automatically feed them into generation.

The next work separates three questions: the quality of the isolated part, its fit to the specific backing, and the reliability of the method. Prompting, context length and multi-stage generation can then be compared against these audible criteria. The cause of the shared delayed entrance, and the full reason for the Base–XL quality difference, remain open.

## 中文口播版

**第 1 页：先听结果**

核心问题是：模型能否根据已有伴奏生成可信的新声部？什么样的例子是好的，什么样的是不好的？这个案例展示一个较有说服力的例子和一个较不理想的参照，两者使用相同伴奏与显式生成设置。

输入是 30 秒贝斯与鼓，构建时排除了吉他 stem。目标是在同一时间轴上加入前景吉他声部，是给伴奏配一个新声部。

【先短听伴奏。选 Base，依次听独轨和混音，再切换到 XL 的同一位置。两例开头都很安静，应先展示这一点，再跳到约 13 秒以后的主要乐段。】

普通 Base 的这个例子里，我能辨认出电吉他的失真音色和可信的演奏特征。这个判断来自我自己的吉他演奏经验，也是目前最明确的成功之处。对应的 XL 参照仍存在之前的问题；如果比较这两个，我更认可 Base。但 XL 至少生成了可以听见的旋律线，单音线条本身并不能被直接判为吉他 solo 的失败。

两者都还不是完整成功的结果。开头约 10 秒能量很低，主要声部到约 13 秒才出现。独轨是否干净、音符与乐句是否贴合伴奏、换输入后是否还能成功，都还没得到确立。因此，好例子展示的是吉他特征的进展，尚不能结束伴奏生成问题。

信号分析补充了另一种证据：它确认了低能量开头，也显示两份输出的频谱集中程度不同。两个文件都完整，没有满刻度样本。但是一个频带占比不能代替对吉他音色或音乐性的判断。

**第 2 页：看清输入与后端处理**

这里展示实际输入。任务指令是根据音频上下文生成 GUITAR 音轨；caption 则要求主音电吉他、带停顿的前景旋律乐句、只有吉他，不带鼓、贝斯、人声或其他乐器。完整英文原文直接放在页面上。

没有 recovery、moderate energy 等方向性情绪文字，也没有填入 BPM 或调性。音乐语境来自伴奏音频。本例使用直接源音频路线，没有通过音乐语言模型生成 codes。

后端先用 VAE 将伴奏波形变成连续的音频表示，再用独立文本编码器处理指令和描述。LEGO 把源音频上下文和吉他任务交给 Base 的扩散 Transformer，由它生成候选声部，再通过 VAE 解码成波形。最后，在模型之外，将 0.5 倍伴奏和 0.5 倍候选声部相加用于试听。我没有另外加失真效果或修改音符来制造吉他印象。

**第 3 页：排查过程与剩余问题**

得到这个例子之前，我们先核实了几个理解偏差。三条路线都保留了，但它们并不等价：非空 codes 会优先于 source audio，文本生成 codes 的路线实际上没有听伴奏。我们也去掉了早期冲突的情绪文字。

另一处是 DCW：当时请求开启，但因为依赖导入失败，实际没有执行。修好以后，校正确实运行并改变输出，但有效 ON 的那个例子仍然很吵杂。执行问题得到修复，不意味着音乐质量问题也解决了。

随后核实 LEGO 支持的 checkpoint，用相同显式设置尝试普通 Base，得到当前正面例子。独立的音频理解调用返回过不确定的速度、调性和乐器描述，我们没有将它们自动填回生成。

接下来分开研究三个问题：独轨质量、对这段伴奏的适配、方法的稳定性。然后把 prompt、上下文长度和多阶段生成作为方法来比较。两例共同延迟进入的原因，以及 Base 与 XL 音质差异的完整原因，都还需要调查。

## Useful answers during discussion / 互动问答

- **Did the model actually receive the backing? / 模型真的听到了伴奏吗？** The saved run logs source-audio processing; audio_codes is empty and the music LM is not initialized. This supports actual backing consumption through the VAE/LEGO context path. It does not yet isolate the causal benefit of that particular backing.
- **Why is Base better? / Base 为什么更好？** We can describe the recognised guitar character and measured acoustic differences. We have not established the full mechanism. The checkpoints differ in architecture, weights and training; the reused XL run was in an earlier process. Equal seeds do not imply equal internal noise.
- **Is the bad example completely unsuccessful? / 较差例子是否一无是处？** No. XL produces a melodic line and a technically valid file. It is less convincing in this pair’s user feedback, while the exact new detailed failure score is absent. Do not call monophony itself a failure.
- **Did DCW cause this pair’s difference? / 这对差异是 DCW 导致的吗？** Both selected examples are DCW OFF. The separate OFF/ON experiment demonstrated an effective correction after repair, but that ON sample’s noisy feedback is not the selected XL OFF audio.
- **Did the text describe recovery or moderate energy? / 是否有与伴奏冲突的文字？** Not in this pair. That was an earlier setup issue; the actual strings shown here only describe the target part and operation.
- **Why not claim clean guitar stems yet? / 为什么暂时不称为纯净吉他 stem？** “Guitar only” is the instruction, while current positive feedback was on the mix. Isolated cleanliness, harmonic fit and adoption still need explicit listening.
- **What should the professor listen for? / 请导师重点听什么？** Instrument identity, attack/sustain and phrasing, how the part enters, and how notes sit against the backing. Ask for the observed passage and criterion; do not equate louder with better.

## Playback and scope notes / 播放与范围

- Case switch: Base (new e06) versus XL-Base (reused e04 OFF). The three track controls always mean backing / candidate / mixture for the selected case.
- Preserve identical playback positions. Include the quiet opening; do not present only a favourable excerpt as the entire result.
- Backing-only = 0.5 × input backing; solo = 0.5 × native candidate; mix = their sum. Fixed gain does not equal perceptual loudness matching.
- Video illustrates the saved **Base mix only**, not the currently selected XL case and not live generation. WAV players retain the original prepared WAV bytes; video uses an AAC derivative.
- No new model inference, music editing, listening judgement or formal score was produced to create this briefing. This is a presentation draft grounded in existing records.

## Reuse with another model or target part / 换模型或声部时如何沿用

Keep the same reporting sequence: **target → actual inputs → executed path → isolated and mixed examples → strengths and failures → evidence → next discriminating test**.

| Field to replace for a new case | Keep this question explicit |
|---|---|
| Input and intended part / 输入与目标声部 | What musical job should the new part perform? / 新声部要完成什么音乐任务？ |
| Prompt, audio, model and effective settings / 实际输入与设置 | What reached the model, rather than merely appearing in a request? / 哪些条件实际进入了模型？ |
| More/less convincing examples / 较好与较差样例 | What does each achieve, and where does each fall short? / 两例分别做到了什么、哪里不足？ |
| Same-position backing, solo and mix / 同位置三种试听对象 | Can the listener distinguish the candidate from the backing? / 能否分清候选声部与伴奏？ |
| Listening and quantitative observations / 听感与量化观察 | Which claim does each observation support, and which does it not? / 每条证据能支持哪一个结论？ |
| Failure, unknown and cause / 失败、未知与原因 | Is this a demonstrated failure, an untested property, or an explanation still needing a test? / 是已观察到的失败、未验证的能力，还是待验证的原因？ |
| Next comparison / 下一项比较 | Which change would distinguish competing explanations? / 下一项变化能区分哪些解释？ |

The current page is a worked example of this method. Reusing the framework does not mean reusing this case’s scores, prompt assumptions or musical conclusions.