IT之家 10 月 25 日消息,Twitter 联合创始人杰克 · 多尔西(Jack Dorsey)旗下的 Block 公司推出了一款全新语音模型 Dessa,兼容本地运行,能模拟笑声、停顿、哽咽等情绪表达,听感高度拟人。官方介绍,Dessa 完全基于人类语音数据训练,训练过程仅使用了一组英伟达 GPU 机架。
Dessa 属于全双工模型,和 OpenAI 今年早些时候发布的 GPT-Live 一样,可以一边说话一边收听使用者语音。主流语音模型一般需要搭配大语言模型来调取知识;Block 称,Dessa 自身的知识来源于其训练所用的对话数据。
Block 表示,现有语音 AI 感受经常生硬、违和,根源在于生成链路要先转成文本,再由文本驱动语音合成。ElevenLabs v4、Fish Audio Drama 3 等不少模型依靠语音标签来控制情绪。而这家多尔西的公司宣称,情绪表达是模型原生内置的,不过当下尚不清楚它底层是否仍然依靠某种标签来调控情绪。
Block 在博客中写道:“现在的语音 AI 交互感受实在糟糕。它先把你的语音转成文字,大量文字承载不了的信息直接丢失。即便是全球最强的 AI,也听不出你的情绪。笔者认为这很不合理,不仅感受差,还存在潜在风险。”
研究团队放出了多段和 Dessa 的对话样例,这个智能体可以切换音色、语速与情绪。例如有一段对话里,AI 说自己工作很累,发声状态也同步带上疲惫感。
还有一段示例中,AI 聊到对方带给自己安心的感受,语气带着情绪,近乎哽咽。另外一段演示是 Block 的员工和这个语音助手玩游戏。
官方说明,该模型可以在消费级电脑本地运行,设备门槛为 M4 Max 及更高规格的 MacBook。但现阶段 Dessa 暂不对外开放试用,Block 正向安全研究者发出测试邀请,重点测试这类语音智能体欺骗人类的风险情形。
据IT之家了解,Block 的主营业务是金融服务,面向小微企业给出应对思路,旗下 Square 已经在用语音 AI 处理下单业务。自研这款可端侧运行的语音模型之后,Block 有望以更低成本,落地更多面向商户与普通使用者的语音机型。
本文来源:https://www.ithome.com/1/017/557.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
According to IT Home on October 25, Block, the company founded by Twitter co-founder Jack Dorsey, has launched a new voice model called Dessa. The model supports local deployment and can simulate emotional expressions such as laughter, pauses, and sobbing, sounding highly lifelike. According to the official introduction, Dessa was trained entirely on human voice data, with the training process utilizing only a single rack of NVIDIA GPUs.
Dessa belongs to a full-duplex model, just like GPT-Live released by OpenAI earlier this year, which allows it to listen to users while speaking. Mainstream voice models generally require large language models to access knowledge; Block states that Dessa’s own knowledge comes from the conversational data used to train it.
Block states that current voice AI often feels stiff and unnatural because the generation process first converts speech to text, which then drives the speech synthesis. Many models, like ElevenLabs v4 and Fish Audio Drama 3, rely on voice tags to control emotion. Dorsey's company, however, claims that emotional expression is native to its model, although it's not yet clear if its underlying system still relies on some kind of tags to regulate emotion.
Block wrote on his blog: "Current voice-based AI interactions feel terrible. It converts your speech into text first, and a significant amount of information that text cannot carry is lost. Even the world's best AI cannot discern your emotions. I believe this is unreasonable; it not only feels bad but also poses potential risks.
The research team has released several conversation samples featuring Dessa, an AI agent capable of modulating its voice, speech rate, and emotional tone. For example, in one conversation, the AI mentioned being exhausted from work, and its vocal tone simultaneously reflected that fatigue.
In another example, the AI spoke about the sense of peace the other person brought, sounding emotional and almost choking up. Another segment showed Block employees playing games with the voice assistant.
According to official specifications, the model can run locally on consumer-grade computers, requiring an M4 Max MacBook or higher. However, Dessa is not currently available for public trial. Block is inviting security researchers to test the system, focusing on the risks of these voice AI agents deceiving humans.
According to IT Home, Block’s core business is financial services, providing solutions for small and micro-businesses. Its subsidiary, Square, already utilizes voice AI to process orders. With the development of this on-device voice model, Block is poised to deploy more voice-based solutions for merchants and general consumers at a lower cost.
Source: https://www.ithome.com/1/017/557.htm. For copyright issues, please email us at 13951431913@139.com or leave a comment on the right.

