大模型正在从聊天框进入真实世界，而语音可能成为其中最重要的入口之一。
10 月 22 日，科大讯飞集中披露了新一代人机交互技术进展。其中，正式发布的星火语音基座大模型 SparkAudio 1.0，将语音识别、同传和交互进一步扩展至复杂音频理解、情绪识别、实时说话人分离和环境声理解；与此同时，新一代端云一体通用大模型讯飞星火 X2.6*及端侧模型 Spark-X2.6-4B 也同步亮相。
讯飞开放平台公布的数据显示，2026 年平台大模型相关硬件装机量、激活量以及单台设备的人机交互次数均快速增长，其中语音大模型相关调用量同比增长超过 20 倍。
这意味着，智能硬件对大模型的要求已经不只是“能够对话”。当模型进入机器人、汽车、智能家居和公共服务设备，AI 需要面对多人同时讲话、复杂环境噪声、不同声源位置、连续上下文以及多台设备协同等现实问题。
SparkAudio 1.0：从“听清一句话”走向“听懂现场发生了什么”
此次发布的核心模型之一，是星火语音基座大模型 SparkAudio 1.0。与传统主要围绕自动语音识别（ASR）和语音合成（TTS）展开的语音系统不同，SparkAudio 1.0 试图把更多声音信息纳入统一理解体系。
根据科大讯飞公布的信息，SparkAudio 1.0 不仅支持语音识别、同传和语音交互，还可以处理音频理解、情绪识别、实时说话人分离以及环境声理解等任务。
这也是语音大模型与传统语音识别模型之间一个重要的能力分界。过去，一套完整语音交互系统通常采用级联架构：首先通过声学算法完成降噪，再进行语音识别，将语音转换成文本，交给自然语言理解或大语言模型，最后再通过语音合成输出回答。这种结构每一层都承担明确任务，但也天然存在信息损失。说话人的语气、情绪、环境声音、空间位置等信息，在语音被转换成纯文本以后很容易消失。
讯飞此次强调的方向，是将过去分散处理的部分能力进一步统一，让模型不仅处理语言内容，同时理解声音本身携带的信息。在多人对话环境中，这一点尤其重要。系统需要知道现在是谁在说话、用户有没有打断、不同声音来自什么方向，以及究竟应该回应哪一个人。
SparkAudio 1.0 由此开始从“语音转文字”的传统能力，进一步向完整声音场景理解扩展。目前，该模型相关能力已经进入讯飞输入法 16.0、讯飞听见、讯飞智能办公本等产品。
星火 X2.6-4B：把更多任务搬到本地
除了专门面向声音世界的 SparkAudio 1.0，讯飞此次另一条重要模型路线是端云一体的讯飞星火 X2.6。其中，云端使用星火 X2.6 提供更强的复杂任务处理能力，端侧则由 Spark-X2.6-4B 承担本地任务。
这款 4B 级端侧模型延续了此前 Spark-X2.5 系列的百万 Token 长上下文架构，同时通过新的后训练进一步强化长文本理解、工具调用和智能体任务能力。公开模型信息显示，Spark-X2.6-4B 拥有约 41 亿参数，最长上下文达到约 100 万 Token。
它的上一代是今年 9 月发布的 Spark-X2.5-4B 和 Spark-X2.5-1.7B。两款模型同样原生支持最长 100 万 Token 上下文，其中 1.7B 模型进一步降低了端侧部署所需要的计算资源。
按照讯飞公布的数据，目前部分端侧方案已经能够达到约 70%的云端能力。过去，端侧模型受到参数规模、内存和算力约束，更多只能承担唤醒、固定指令以及简单问答等任务。但随着 4B 甚至更小模型开始具备百万上下文、工具调用和 Agent 能力，部分原本必须上传云端完成的工作开始能够直接留在设备本地。
讯飞表示，简单、高频、涉及敏感数据的任务可以直接由端侧完成；更加复杂的任务调用云端星火 X2.6，还有一部分任务则由两者共同执行。
科大讯飞董事长刘庆峰判断：未来约 60%的应用可以在本地完成，30%通过端云协同解决，剩余约 10%的高复杂度任务交给云端大模型。
讯飞正在把“声学空间”也交给模型理解
仅仅提高语言模型能力，还解决不了真实世界里的交互问题。一个人在办公室里与 AI 讲话，和五个人在嘈杂展厅里同时交谈，对系统而言是完全不同的问题。
因此，在 SparkAudio 1.0 之外，讯飞还公布了新的多模交互和具身多模交互技术，把声源定位、多人交互、视听觉信息和空间环境进一步结合。
此前，讯飞已经尝试将唇动视觉信息与麦克风阵列结合，用多模态方式解决多人嘈杂环境中的“鸡尾酒会问题”。但这类方案也存在局限。例如，不同光线、角度以及用户是否面对摄像头，都会影响视觉信息的有效性。
此次讯飞进一步引入全景摄像头与环形麦克风阵列，使系统能够在 360 度空间内处理多人声音。根据披露数据，在多人同时聊天时，系统串扰率可以降至 1.7%以下；最高 5 人同时交互情况下，实时分辨率达到 87%。
在机器人方向，讯飞此次公开的“星火具身多模交互技术”则继续向前走了一步。其能力包括复杂空间动态降噪、多目标人的交互意图识别、语音与动作统一规划和执行，以及由视觉信息触发的主动交互。
这意味着 AI 需要解决的问题不再只是“听见谁说话”，还需要理解不同人在空间里的位置、状态和意图，并进一步决定自己的动作。对于服务机器人而言，这种能力的重要性并不亚于纯粹的大语言模型推理能力。
从一台设备回答，到整个空间“接着聊”
模型能力向物理空间延伸后，设备之间的上下文割裂也开始成为问题。
目前，一个家庭里的电视、手机、音箱、投影仪等设备通常拥有各自独立的交互系统。用户在客厅电视上启动一个任务，换到卧室另一台设备后，系统通常不知道前面发生过什么。
讯飞的解决办法是通过统一云端 Agent 管理上下文与任务状态，让多个设备共享同一段交互过程，实现自动换端、换端继续聊天以及不同设备之间的任务协同。
这意味着语音助手的主体可能逐渐从“一台硬件”转向“整个空间”。用户不再需要主动判断应该对哪台设备讲话，系统可以根据距离、任务类型和设备能力，决定由谁接收指令、由谁进行回答以及任务最后在哪个设备完成。
这与端云模型背后的逻辑实际上是一致的：AI 不再绑定固定终端，而是根据当前任务动态决定应该调用哪种模型、哪台设备和哪项能力。
面向海外硬件市场，多语种也是讯飞此次重点强化的一项能力。
目前讯飞开放平台已经支持 40 种以上语言，其中 30 种语言能够实现免切换语音交互。所谓免切换，即用户无需事先选择英语、法语或其他语言，系统可以直接判断当前使用的语言并完成识别和响应。讯飞开放平台目前还提供独立的“多语种识别大模型”，官方产品页面显示，其可支持 37 种多语种音频免切识别。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
LLM is entering the real world from chat frames, and voice could be one of the most important entry points.
On 22 October, Kodak iFlytek focused on the new generation of interactive technology. Of these, the official publish speech base LLM SparkAudio 1.0 further extends voice recognition, simultaneous transmission and interaction to complex audio understanding, emotional recognition, real-time voice separation, and environmental acoustic understanding; at the same time, the new generation of end clouds is integrated into the LLMZiFlytek fire X2.6* and the side model Spark-X2.6-4B is synchronized.
Data published by the iFlytek Open Platform show that in 2026, the number of hardware installed, activated and human interaction of the LLM related to the platform and individual devices increased rapidly, with the number of voice LLM related calls increasing by more than 20 times.
This means that the requirements of smart hardware for LLM are no longer “dialogueable”. As models enter robotics, cars, intelligent homes and public service equipment, AI needs to face the reality of multiple simultaneous speeches, complex environmental noise, different sources of sound, continuous context and multiple equipment synergies.
SparkAudio 1.0: From "hear a word" to "understand what's going on."
One of the core models of publish this time is Starfire Voice Base LLM SparkAudio 1.0. Unlike the traditional voice-based system, which mainly revolves around automatic voice recognition (ASR) and speech synthesis (TTS), SparkAudio 1.0 attempts to integrate more sound information into a unified understanding system.
According to information published by Big Z66Z, SparkAudio 1.0 not only supports voice recognition, simultaneous and voice interaction, but also handles audio understanding, emotional recognition, real-time voice separation and environmental acoustic understanding.
This is also an important capability boundary between voice LLM and traditional voice recognition models. In the past, a full-fledged voice-interaction system used to be cascaded: first, acoustic algorithms were used to reduce noise, then voice recognition, conversion of voice into text, to natural language understanding or large language models, and finally to voice synthesis output. Each layer of this structure has a clear mandate, but there is also a natural information loss. The voice, emotions, environmental sound, spatial location, etc. of the talking person can easily disappear when the voice is converted into a pure text.
This time, iFlytek emphasized the direction of further harmonizing parts of the previously decentralized capacity so that the model not only addresses language content but also understands the information carried by the sound itself. This is particularly important in a context of multi-person dialogue. The system needs to know who is talking now, whether users are interrupting, where different voices come from, and who should respond.
SparkAudio 1.0 thus begins to expand the understanding of the whole sound scene from the traditional ability to "speak to text." Currently, the capabilities of the model have entered products such as iFlytek input method 16.0, iFlytek hearing, and iFlytek smart office book.
Starfire X2.6-4B: Move more tasks locally
In addition to SparkAudio 1.0, which is dedicated to the sound world, another important model of iFlytek this time is iFlytek fire X2.6. Of these, clouds use fire X2.6 to provide a more complex task handling capability, while the end side is assigned to Spark-X2.6-4B for local tasks.
This 4B lateral model continues the millions of Token long context structures of the previous Spark-X2.5 series, while further enhancing long text understanding, tool call and AI agent mission capabilities through new post-training. Open model information indicates that Spark-X2.6-4B has approximately 4.1 billion parameters, with the maximum context reaching about 1 million Token.
Its predecessor was Spark-X2.5-4B and Spark-X2.5-1.7B in September of this year. The two models also have primary support of up to 1 million Token context, of which 1.7B further reduces the computing resources required for end-side deployment.
According to data published by iFlytek, some of the end-side programmes are now able to reach approximately 70 per cent of cloud capacity. In the past, end-side models were constrained by the size, memory and computation of parameters, with more tasks such as awakening, fixed instructions and simple questions and answers. But as 4B and even smaller models begin to have millions of context, tool call and Agent capabilities, some of the work that had to be done on the cloud has started to stay directly in the facility.
iFlytek states that simple, high frequency tasks involving sensitive data can be performed directly by the end side; more complex tasks call for cloud fire X2.6, while some tasks are carried out jointly.
President Liu Qingfeng of Z66ZK: About 60 per cent of future applications can be implemented locally, 30 per cent can be resolved through cloud-end collaborative solutions, and about 10 per cent of the remaining high-complex tasks are entrusted to cloud LLM.
iFlytek is giving the acoustic space to the model.
Improving language modelling alone will not solve the problem of interaction in the real world. One person speaks to AI in his office, and five people talk at the same time in the lobby, which is a completely different issue for the system.
Thus, in addition to SparkAudio 1.0, iFlytek has published new multimodular and multimode interactive technologies that further integrate acoustic source positioning, multi-person interaction, audio-visual information and the space environment.
Previously, iFlytek had tried to combine the visual information on the movement of the lips with the microphone array to solve the problem of the cocktail parties in a multi-temporal environment. However, such programmes also have limitations. For example, the effectiveness of visual information is affected by different light, angles and whether users face the camera.
This time iFlytek further introduced panoramic cameras and circular microphone arrays, enabling the system to handle multiple voices in 360 degrees. According to the disclosed data, when multiple people talk at the same time, the system string disturbance rate can be reduced to less than 1.7 per cent; up to 5 people interact at the same time, with resolution reaching 87 per cent in real time.
In the robotic direction, iFlytek's open “Modular Interactivity Technologies for Starfire” continues to move forward. Its capabilities include complex space dynamic noise loss, multi-target interactive intent recognition, unified voice and action planning and implementation, and active interaction triggered by visual information.
This means that AI will need to address the issue no longer simply “who hears whom”, but also to understand the position, status and intentions of different people in space and to further determine their own actions. For service robots, this capability is no less important than pure large-language model reasoning.
From one device to the entire space, "Contra."
The fragmentation of context between equipment also began to become a problem as model capabilities were extended to physical space.
Currently, television, mobile phones, soundboxes, projectors, etc. in a family often have separate interactive systems. The system usually does not know what happened in front of it when the user started a mission on the living room television and switched to another device in the bedroom.
The solution for iFlytek is to share the same interactive process with multiple devices through the unified cloud management of context and mission status, to achieve automatic re-ends, to continue chatting and task synergy between different devices.
This means that the subject of the voice assistant may gradually move from “one hardware” to “the whole space”. The user is no longer required to make a proactive judgement as to which device should be spoken, and the system can determine who will receive the instructions, who will answer them and which equipment will be completed at the end of the mission, depending on distance, type of mission and equipment capability.
This is in fact consistent with the logic behind the end cloud model: AI no longer binds fixed terminals, but rather determines which model, which equipment and which capability should be mobilized in the light of current mission dynamics.
For overseas hardware markets, multilingualism is also an enhanced capacity of iFlytek.
Currently, the iFlytek Open Platform has supported more than 40 languages, of which 30 have been able to achieve off-put voice interaction. The switch is that users do not have to choose English, French or other languages in advance, and the system can directly judge the language currently used and complete identification and response. The iFlytek Open Platform also currently provides a stand-alone Multilingual Identification LLM, which is shown on the official product page to support 37 multilingual audio-verbal recognition.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.