Suno发布首个与唱片业合作打造的AI音乐模型

qimuai 发布于 阅读:33 一手编译

Suno发布首个与唱片业合作打造的AI音乐模型

内容来源:https://www.theverge.com/ai-artificial-intelligence/991977/suno-releases-its-first-ai-music-model-made-with-record-industry-help

内容总结:

Suno发布首个与唱片行业合作打造的AI音乐模型v6。据Suno的Jack Brody向The Verge透露,v6“从头开始训练,使用了一套全新的数据,不再包含此前模型所依赖的相同数据”。新数据包括来自华纳音乐集团、BMG和Believe等合作伙伴的授权内容,以及“用户数据”。不过,这是否意味着v6的训练数据完全不含来源存疑的内容,目前尚不清楚。

v6的一大变化是实际上包含三个不同模型:v6、v6-wild和v6-mini。Mini面向所有用户免费开放,主打快速、轻量化的创作,但生成结果较为简单,且往往带有更多明显的AI痕迹。v6-wild则被设计得更加不可预测。Brody表示,v6-wild旨在制造“意外的惊喜和自然的不完美”。但在有限的测试中,笔者很难察觉到wild与标准版v6之间的区别。

三个模型在理解音乐流派方面都有显著提升。笔者用之前测试旧模型时用过的一些提示词重新尝试,比如要求生成“hyperpop”或“krautrock”歌曲,旧模型完全跑偏,而v6则准确抓住了笔者所提任何流派的所有表面特征。

然而,v6仍然无法真正实现所谓的“自然的不完美”。与v5一样,笔者尝试让它生成走调、跑调或略带不和谐效果的结果,每次都以失败告终。无论怎么调整提示,Suno v6都只能产出和声与节奏完美无瑕的作品。告诉它要“单调的人声”和“不要鼓”,它会直接无视(见下文)。反复要求“毫无章法、不顾调性和节奏的走调钢琴,与歌曲其余部分完全不合调”,得到的也完全不是那么回事。

它倒是保留了AI音乐那种不自然的不完美——人声部分尤其如此,相比v5,似乎产生了更多AI模型常见的刺耳痕迹。

Suno v6还在底层做了一些更新,可能从根本上改变用户与它的交互方式。现在用户可以在聊天中用自然语言编辑歌曲的局部,修改吉他线或某句歌词等单个元素时,不再需要重新生成整首歌曲。v6还能将你Suno库中的多个元素混搭在一起,把不同输出组合成新的创作。

提示词也不再局限于文字描述。Suno v6可以根据图片、视频或其他音频来生成曲目。

Suno v6从今天开始逐步推送,该公司表示最终将淘汰之前的模型。

中文翻译:

Suno 全新的 v6 AI 音乐模型,是它第一个在唱片行业支持下打造的作品。Suno 的 Jack Brody 对 The Verge 表示,v6 是“从零开始训练的,使用了一套全新的数据,不再包含我们此前模型训练时所用的同一批数据”。这些数据包括来自合作伙伴华纳音乐集团、BMG 和 Believe 的授权内容,以及“用户数据”。目前尚不清楚,这是否意味着 v6 的训练数据已经完全不含来源可疑的内容。
Suno 发布首个在唱片行业协助下打造的 AI 音乐模型
v6 仍然无法捕捉真实人类音乐中的“自然瑕疵”。
v6 仍然无法捕捉真实人类音乐中的“自然瑕疵”。
v6 的一大变化在于,它实际上包含三个不同的模型:v6、v6-wild 和 v6-mini。Mini 是向所有人免费开放的模型,主打快速、轻资源的创作。它生成的结果更简单,也往往带有更多一眼就能看出是 AI 提示词产物的痕迹。v6-wild 则应当更具不可预测性。Brody 说,v6-wild 是为“意外之喜和自然瑕疵”而设计的。不过,在我有限的测试中,我很难察觉到 wild 和标准版 v6 之间有什么差别。
这三个模型在理解音乐类型方面都有了显著提升。我拿以前用在旧模型上的一些提示词重新试了一遍,那些提示词要求生成“hyperpop”或“krautrock”歌曲,结果旧模型偏得离谱。相比之下,v6 把我抛出的每一种风格最表面、最典型的特征都拿捏住了。
然而,它依然做不到的,是真正呈现出那种“自然瑕疵”。和 v5 一样,我每次想让它做出跑调、走音或轻微不协和的效果,最后都失败了。不管你怎样逼 Suno 的 v6,它做出来的东西依然只有和声与节奏上的完美。你告诉它要“单调人声”和“不要鼓”,它会直接无视你(见下文)。反复要求它做出“毫无章法、完全不顾调性和节奏的走音钢琴”,而且还要“与歌曲其余部分完全不在一个调上”,结果也完全不是那么回事。
它倒是保留了 AI 音乐那种不自然的瑕疵,尤其是人声,似乎比 v5 更容易产生 AI 模型常见的那种边缘刺耳、痕迹明显的毛病。
Suno 的 v6 还有一些底层更新,可能会从根本上改变人们与它的互动方式。现在,用户可以在聊天中用自然语言编辑歌曲的某个部分,而修改吉他线或某一句歌词这样的单个元素,也不再需要重新生成整首曲目。v6 还可以把你 Suno 资料库中的多个元素混搭在一起,把不同的输出组合成新的作品。
提示词也不再局限于文字描述。Suno v6 可以根据图像、视频或其他音频来创作曲目。
Suno v6 从今天开始逐步推出,该公司表示,旧模型最终会退役。

英文来源:

Suno’s new v6 AI music model is its first made with support from the record industry. Suno’s Jack Brody told The Verge that v6 was “trained from the ground up, with a new set of data that does not include the same data that our previous models were trained on.” The data includes content licensed from partners Warner Music Group, BMG, and Believe, as well as “user data.” It’s unclear whether that means v6 training data is completely free of dubiously obtained content.
Suno releases its first AI music model made with record industry help
v6 still can’t capture the ‘natural imperfections’ of real human music.
v6 still can’t capture the ‘natural imperfections’ of real human music.
One of the big changes in v6 is that there are actually three different models: v6, v6-wild, and v6-mini. Mini is the model available for free to all. It’s focused on fast, resource-light creation. The results it spits out are simpler and often have more artifacts that betray it as the product of an AI prompt. v6-wild is supposed to be more unpredictable. Brody said that v6-wild is designed for “happy accidents and natural imperfections.” Though, in my limited testing, I found it hard to notice a difference between wild and bog-standard v6.
All three models are dramatically better at understanding genre. I recycled a handful of prompts that I used with older models calling for “hyperpop” or “krautrock” songs that missed the mark by a mile. v6, on the other hand, nailed all the superficial hallmarks of any genre I threw at it.
What it remains incapable of doing, however, is actually delivering on the “natural imperfections.” As with v5, my attempts to get off-key, out-of-tune, or slightly dissonant results failed every time. No matter how hard you push Suno’s v6, it remains incapable of making anything other than harmonic and rhythmic perfection. Tell it you want “monotone vocals” and “no drums,” and it straight-up ignores you (see below). Repeated demands for “out-of-tune piano played haphazardly, with no regard for key or rhythm” that was “fully out of tune with the rest of the song,” delivered nothing of the sort.
It does retain the unnatural imperfections of AI music, with vocals especially seeming to produce more of the harsh-edged artifacts common with AI models than v5 did.
Suno’s v6 also has some under-the-hood updates that can fundamentally change how people interact with it. Now users can edit parts of a song using plain language within a chat, and changing individual elements like a guitar line or a single lyric doesn’t require regenerating the entire track. v6 can also mash together multiple elements from your Suno library, combining outputs into new creations.
Prompts are also no longer limited to text descriptions. Suno v6 can create tracks based on images, video, or other audio.
Suno v6 starts rolling out today, and the company says it will eventually retire its previous models.

ThevergeAI大爆炸

文章目录


    扫描二维码,在手机上阅读