评论
打印
LifeScore, says Tom Gruber, a co-founder, is "literally swamped with requests" from clients including Sony Music, Universal Music Group and Warner Music Group.
联合创始人汤姆·格鲁伯表示，LifeScore被索尼音乐、环球音乐集团和华纳音乐集团等客户的请求淹没，不堪重负。
An original release is typically turned into anywhere from a handful to a dozen remixes. But one client aims to release a dizzying 6,000 or so AI versions of an original track, each targeting a different market.
原始曲目通常会混几次到十几次音。但一位客户打算发布一首原创曲目的6000多个AI版本，每个版本都针对不同的市场，令人眼花缭乱。
Artists including Pink Floyd's David Gilmour and Tom Gaebel, a German pop singer, use LifeScore's AI to power websites that allow fans to generate, with a few clicks, new remixes adapted to personal tastes.
摇滚乐团平克·弗洛伊德的成员大卫·吉尔摩和德国流行歌手Tom Gaebel等艺术家使用LifeScore的AI功能为网站提供支持，粉丝只需点击几下即可生成适合个人品味的新混音。
If this seems like dizzying progress, it's worth noting that AI's impact on music is still in its early days. Legal uncertainties over the use of copyrighted recordings to train models have slowed development.
如果这种进步看起来令人眼花缭乱，那么值得注意的是，AI对音乐的影响仍处于早期阶段。使用受版权保护的录音来训练模型带有法律不确定性，从而减缓了开发速度。
Outfits that have coughed up for licensing fees note that this can get expensive. To save on that cost, MusicGen's training set mostly sidestepped hits, says Dr Adi.
支付版权费的公司指出，开发可能会变得昂贵。阿迪博士表示，为了节省成本，MusicGen的训练样本大多避开了热门歌曲。
Though output is pretty good, he adds, the model is not yet "artistic enough" to generate narratively complete songs. Harmonic misalignments are common.
他补充道，虽然输出的歌曲相当不错，但该模型还不够“艺术化”，无法生成叙事完整的歌曲。 谐波失调很常见。
OpenAI, a San Francisco firm, for its part, says its MuseNet model struggles to pull off "odd pairings", such as a Chopin style that incorporates bass and drums.
总部位于旧金山的公司OpenAI表示，其MuseNet模型很难实现“奇怪的配对”，例如融合贝斯和鼓的肖邦风格。
In time, bigger training sets of better music will largely overcome such shortcomings, developers reckon.
开发人员认为，随着时间的推移，增加优质音乐的训练样本将在很大程度上克服这些缺点。
A Stability AI spokesperson says that while Stable Audio's top duration for coherently structured music—"intro, development and outro"—is now about 90 seconds, upgrades will produce longer pieces with "full musicality".
Stability AI发言人表示，虽然现在Stable Audio结构连贯的音乐（“前奏、展开和尾奏”）的最高持续时间约为90秒，但升级后将生成更长的具有“完整音乐性”的作品。
But judging music AI by its ability to crank out polished tracks mostly misses the point. The technology's greatest promise, for now at least, lies elsewhere.
但从能否生成优美的曲目方面来评估AI大多没有抓住要点。至少就目前而言，该技术最大的前景在于其他地方。
Part of it is the empowerment of amateurs. AI handles technical tasks beyond many people's capabilities and means. As a result, AI is drawing legions of newbies into music-making.
其中一部分是业余爱好者的赋权。AI处理的技术任务超出了许多人的能力和手段。因此，AI正在吸引大批新手进击音乐制作领域。
This is a boon for experimentation by what Simon Cross, head of products at Native Instruments, a firm based in Berlin, calls "bedroom producers".
这对于被柏林公司Native Instruments的产品主管Simon Cross称为“卧室制作人”的实验来说是一个福音。
Consider rx, a Native Instruments AI "assistant" that corrects errors in things like pitch and timing.
以rx为例，它是Native Instruments的 AI“助手”，可以纠正音高和计时等方面的错误。
For the latter, software time-shifts notes by cutting out or inserting slivers of sound with matching timbre, a process called "dynamic time-warping".
软件通过剪切或插入匹配音色的声音片段来对音符进行时移，这一过程称为“动态时间归整”。
The company's AI also determines what mixing and mastering processes were performed on a song of a user's choosing.
rx还可以决定对用户选择的歌曲进行哪些混音以及母带处理过程。
It then replicates, or at least approximates, the same expensive processing on the user's own creations.
然后，它会在用户的作品上复制或至少接近相同的昂贵处理。
Boomy, an online "music automation" platform for what Alex Mitchell, its ceo, describes as "low-friction" song production with text prompts, has more than 2m users.
Boomy是一个“音乐自动化”在线平台，拥有200多万用户，其首席执行官亚历克斯·米切尔称其为带有文本提示的“低摩擦”歌曲制作。
The company, based in Berkeley, California, uploads users' (vetted) creations to streaming services and collects a cut of revenues.
该公司总部位于加利福尼亚州伯克利市，通过将用户（经过审核的）创作上传到流媒体服务盈利。
AI serves professionals, too. The soundtracks to "Barbie" and "Oppenheimer" were cleaned up in post-production with rx, for example.
AI也为专业人士服务。 例如，《芭比》和《奥本海默》的配乐在后期制作中用rx进行了处理。
Another application area is "style transfer", in which models transform music recorded with one instrument into sounds that seem to come from a different one, often with a twist or two requested by the user.
另一个应用领域是“风格迁移”，模型将一种乐器录制的音乐转换为不同乐器的声音，通常会根据用户需求进行一两次调整。
Style transfers are also used for voice. A model developed by a startup in London called Voice-Swap slices up sounds sung by (remunerated) professional singers and rearranges the slivers into lyrics written by the service's users, who pay licensing fees for the rights to sell the resulting tracks.
风格迁移也用于语音。伦敦一家名为Voice-Swap的初创公司开发了一种模型，将（付费的）专业歌手演唱的歌曲进行切片，并将这些碎片重新排列成由该服务的用户编写的歌词，要想销售所得曲目，用户必须支付版权费。
And AI tools already exist to recreate singers' voices in other languages.
AI工具已经可以用其他语言重现歌手的声音。
Vocaloid, a voice-synthesising tool from Yamaha, a Japanese instrument manufacturer, is one of many that can use a translation sung by a native speaker as a template for an AI to imitate as it rearranges, modifies and stitches together tiny snippets of the original singer's voice.
Vocaloid是日本乐器制造商雅马哈开发的语音合成软件，它可以使用母语人士演唱的翻译作为模板供人工智能模仿，因为它可以重新排列、修改和缝合原歌手声音的微小片段。
Accomplished musicians now widely tap MusicGen and its competitors as sources of "infinite inspirations", the better to alight upon promising composition ideas, says Meta's Dr Adi.
Meta的阿迪博士表示，技艺高超的音乐家现在普遍将MusicGen及其竞争对手视为“无限灵感”的来源，以便发现更好的作曲创意。
Whether such inspiration pays off will, ultimately, be up to the listener to decide.
这种灵感是否有效，最终将由听众决定。
