Fish Audio 完成 5200 万美元种子轮融资，打造面向创作者和企业的 AI 语音模型

AI 生成语音模型市场规模庞大。创意类应用需要 AI 语音模型具备更强的表现力，而希望自动化客服和销售运营的企业，则需要模型具有更高的可控性。

总部位于帕洛阿尔托的 Fish Audio 希望通过其包含超过 1.5 万项自然语言控制能力的库，满足上述所有应用场景。自去年推出以来，这家初创公司已有超过 800 万人使用其模型的开源版或托管版，并实现了 2100 万美元的年经常性收入（ARR）。

为延续这一增长势头，该公司周二宣布完成 5200 万美元种子轮融资，由 Coreline Ventures 和今日资本（Capital Today）领投。359 Capital、Parable、Play Time、Alphalist Partners、Bayhouse Ventures、Carya Venture Partners 和 HF0 也参与了本轮融资。

Fish Audio 最初是前 Nvidia 研究员廖世嘉（Shijia Liao）的一个小型项目。由于对市场上缺乏表现力的合成语音感到不满，他在一张 GPU 上训练了一个语音生成模型，随后将其开源。Fish Speech 在 GitHub 上的代码库目前已获得超过 3.1 万颗星，被独立开发者、电子游戏设计师和内容创作者使用。

该公司在过去一年发布了五个模型：四个语音生成模型和一个语音转文本模型。其已开源三个语音生成模型，但最新的 S2.1 Pro 模型仅可通过付费 API 使用。

Fish Audio 提供面向创作者和团队的付费月度套餐，用户可获得一定时长的生成额度以及语音克隆功能。该公司还提供企业版 API 和平台，并表示 HeyGen 和 Sanas 等机构已经在使用其服务。

“每家企业都有不同的使用场景和偏好。例如，像 HeyGen 这样使用我们语音为 AI 虚拟形象提供支持的公司，希望语音更具真实感；游戏工作室则希望其角色拥有富有表现力的声音；而像 LiveKit 这样的语音 Agent 公司，则希望获得更自然、低延迟且足够富有表现力的语音，以用于通话，”Fish Audio CEO 兼联合创始人 Rissa Cao 表示。

这家初创公司构建语音库的方式之一，是邀请用户提交自己的声音来训练模型，并在其声音被使用时向其提供报酬。不过，这一做法在几个月前引发了一些问题：一些创作者指称，他们的声音未经同意便被上传至 Fish Audio。该公司当时设有 DMCA 下架流程来处理这类问题，但实际下架所需时间很长。

Cao 向 TechCrunch 表示，公司目前已将下架流程自动化。她称，创作者可以提交一小段声音样本或合同，以证明已上传的声音属于自己，平台将在不到三分钟内将该声音下架。

不过，这仍无法阻止他人在艺术家不知情的情况下上传其声音。而在艺术家发现此事之前，除非其申请删除，否则其声音将继续在平台上被使用。

Coreline Ventures 合伙人 Osuke Honda 表示，社区驱动型模式只有在创作者信任平台时才能奏效。

“只有当创作者信任平台时，以社区为中心的方法才能成为一种持久优势。这意味着同意、透明度和署名必须被内置于产品中，而不是被当作事后补救措施。我认为，行业需要转向经验证的声音所有权、明确的许可条款、便捷的举报和下架流程，以及最终让创作者能够在其声音被授权或用于商业用途时获得经济回报的收益分成模式，”他说。

Cao 表示，在公司仅以开源项目和面向创作者的套餐提供产品时，其运营效率很高，并不需要外部资本。但随着投资者兴趣升温，公司希望开发更先进的模型并满足企业客户需求，因此决定寻求融资。

展望未来，Fish Audio 计划在今年发布音频理解模型，同时也正在开发语音到语音模型。

语音生成市场竞争激烈，包括 ElevenLabs、WellSaid、Cartesia、Speechify、Async（原 Podcastle）和 Krisp 等公司，都在争夺创作者和企业的预算。

359 Capital 合伙人 Rico Mallozzi 认为，面向开发者的细粒度控制能力和具备成本效率的模型训练，将帮助 Fish Audio 更好地与大型 AI 实验室竞争。

“与其他一些资金雄厚的 AI 实验室或公司相比，他们凭借现有团队所构建的最先进模型令人难以置信。这表明他们在缩小人工语音与类人语音之间差距方面具备出色的技术能力，”Mallozzi 在电话中向 TechCrunch 表示。

本文已更新，以反映该公司在种子轮融资中筹集了 5200 万美元。

The market for AI-generated voice models is massive. Creative use cases require AI voice models to be more expressive, while enterprises looking to automate customer support and sales ops need them to be more steerable.

Palo Alto-based Fish Audio wants to cater to all of those use cases with its library of more than 15,000 natural language controls. Since launching last year, the startup now has more than 8 million people using the open source or hosted versions of its models, and generates annual recurring revenue of $21 million.

To continue building on that traction, the startup on Tuesday said it has raised $52 million in a seed round that was led by Coreline Ventures and Capital Today. The funding also saw participation from 359 Capital, Parable, Play Time, Alphalist Partners, Bayhouse Ventures, Carya Venture Partners, and HF0.

Fish Audio started as a small project by former Nvidia researcher Shijia Liao, who, frustrated by non-expressive synthetic voices available on the market, trained a voice-generation model on a single GPU, which he then open-sourced. The Fish Speech repository on GitHub now has more than 31,000 stars and is used by indie developers, video game designers, and creators.

The company has launched five models in the last year: four speech-generation models and one speech-to-text model. It has open-sourced three of its speech-generation models, but its latest S2.1 Pro model is available only through its paid API.

Fish Audio offers paid monthly plans suited for creators and teams that unlock a set number of minutes of generation plus voice-cloning features. The company also offers an enterprise version of its APIs and platform, and says organizations like HeyGen and Sanas are already using it.

“Every enterprise has different use cases and different preferences. For example, companies like HeyGen, which use our voices to power AI avatars, want realism in voices; a gaming studio would want expressive voices for their characters; and voice agent companies like LiveKit want more natural-sounding and low-latency voices that are expressive enough for calls,” Fish Audio’s CEO and co-founder Rissa Cao said.

One way the startup has built its library of voices is by asking users to submit their own voices for training its models, and compensating them if their voices are used. That resulted in some trouble a few months ago, however, as some creators alleged that their voices were uploaded to Fish Audio without their consent. The startup had a DMCA takedown process in place to address such concerns, but the takedowns themselves took a long time.

Cao told TechCrunch that the company has now automated the takedown process. Creators can submit a short voice sample or a contract to prove that an uploaded voice belongs to them, and their voice will be taken off the startup’s platform in less than three minutes, she said.

Still, that doesn’t prevent anyone from uploading an artist’s voice without their knowledge. And until the artist finds out, their voice will continue to be used on the platform unless they file for its removal.

Osuke Honda, a partner at Coreline Ventures, said a community-driven model only works when creators trust the platform.

“A community-centric approach can only become a durable advantage if creators trust the platform. That means consent, transparency, and attribution must be built into the product rather than treated as afterthoughts. I believe the industry needs to move toward verified voice ownership, clear licensing terms, easy reporting and takedown processes, and eventually revenue-sharing models where creators benefit financially when their voices are licensed or used commercially,” he said.

Cao said when the startup was only offering its product as an open source project with plans for creators, it was running efficiently and didn’t need outside capital. But it wanted to develop more advanced models and wanted to accommodate enterprises as investor interest was ramping up, which led it to seek capital.

Looking ahead, Fish Audio plans to release an audio understanding model this year. It’s also building a speech-to-speech model.

The speech-generation market is crowded, with companies like ElevenLabs, WellSaid, Cartesia, Speechify, Async (previously Podcastle), and Krisp competing for creators and enterprises’ budgets.

According to Rico Mallozzi, a partner at 359 Capital, fine-grained controls for developers and cost-efficient model training will help Fish Audio compete better with big AI labs.

“I think what they’ve been able to build, state-of-the-art models, with the team they have, compared to some of these other well-funded AI labs or companies, is incredible. It shows their technical acumen in closing the gap between artificial-sounding and human-like voices,” Mallozzi told TechCrunch over a call.

The story has been updated to reflect that the company raised $52 million in the seed round.