India's voice AI race has a new contender. Sarvam AI has launched Saaras V4, its latest automatic speech recognition (ASR) model, designed to transcribe speech across Indian languages, English, regional accents and noisy environments.&nbsp;The model can produce transcripts in five formats, including verbatim text, translation and code-mixed language output. Sarvam says it can also identify spoken languages automatically and process audio in real time, with streaming latency below 150 milliseconds. The launch targets a major challenge for voice technology in India: understanding people who switch between languages, speak with regional accents or use everyday expressions that traditional transcription systems can struggle with.<h2>Sarvam Saaras V4 brings five speech output modes</h2>Saaras V4 can convert the same audio into five different formats: verbatim transcription, normalised text, code-mixed text, transliteration and translation. This means developers can choose whether to preserve exactly what someone says, clean up the text, write spoken words in another script or translate them into another language. Sarvam says these functions are built into the model, rather than requiring separate systems to process the transcript afterwards. The model combines an audio encoder with a 3-billion-parameter hybrid state-space language model. It is designed to handle code-switching, dialect differences and background noise.&nbsp;<h2>22 Indian languages and a focus on accuracy</h2>Saaras V4 can automatically identify a spoken language and transcribe it in the corresponding native script. On verified IndicVoices data, Sarvam reports a language identification error rate of 5.22% across 22 Indian languages, falling to 2.9% across the 10 most widely spoken languages in its evaluation. For English, the company says Saaras V4 recorded the lowest average word error rate across seven benchmarks covering Indian English, international accents, meetings, financial conversations and media. The model was also tested on the Vistaar benchmark across 10 Indian languages. These are company-reported results, and performance in everyday use may vary depending on accents, recording quality and background noise.&nbsp;<h2>Faster voice AI for developers</h2>Sarvam says Saaras V4 supports streaming with time to first token below 150 milliseconds, alongside long-form audio processing. It also includes keyterm prompting, allowing developers to supply names, product terms and acronyms that may otherwise be difficult for speech recognition systems to identify. Developers can access the model through Sarvam AI's API, with Python and Node.js SDKs. Integrations include Vercel AI SDK, LiveKit Agents and Pipecat Agents. The launch could support applications such as voice assistants, multilingual customer service, meeting transcription and real-time translation. However, the announcement does not establish pricing, independent benchmark results or how the model compares with competing systems in real-world deployments.

Sarvam AI has launched Saaras V4, a speech recognition model supporting 22 Indian languages and five output formats. With code-mixed speech support and streaming latency below 150 milliseconds, it targets real-time voice applications and developers.

India's voice AI race has a new contender. Sarvam AI has launched Saaras V4, its latest automatic speech recognition (ASR) model, designed to transcribe speech across Indian languages, English, regional accents and noisy environments. The model can produce transcripts in five formats, including verbatim text, translation and code-mixed language output. Sarvam says it can also identify spoken languages automatically and process audio in real time, with streaming latency below 150 milliseconds. The launch targets a major challenge for voice technology in India: understanding people who switch between languages, speak with regional accents or use everyday expressions that traditional transcription systems can struggle with. Sarvam Saaras V4 brings five speech output modes Saaras V4 can convert the same audio into five different formats: verbatim transcription, normalised text, code-mixed text, transliteration and translation. This means developers can choose whether to preserve exactly what someone says, clean up the text, write spoken words in another script or translate them into another language. Sarvam says these functions are built into the model, rather than requiring separate systems to process the transcript afterwards. The model combines an audio encoder with a 3-billion-parameter hybrid state-space language model. It is designed to handle code-switching, dialect differences and background noise. 22 Indian languages and a focus on accuracy Saaras V4 can automatically identify a spoken language and transcribe it in the corresponding native script. On verified IndicVoices data, Sarvam reports a language identification error rate of 5.22% across 22 Indian languages, falling to 2.9% across the 10 most widely spoken languages in its evaluation. For English, the company says Saaras V4 recorded the lowest average word error rate across seven benchmarks covering Indian English, international accents, meetings, financial conversations and media. The model was also tested on the Vistaar benchmark across 10 Indian languages. These are company-reported results, and performance in everyday use may vary depending on accents, recording quality and background noise. Faster voice AI for developers Sarvam says Saaras V4 supports streaming with time to first token below 150 milliseconds, alongside long-form audio processing. It also includes keyterm prompting, allowing developers to supply names, product terms and acronyms that may otherwise be difficult for speech recognition systems to identify. Developers can access the model through Sarvam AI's API, with Python and Node.js SDKs. Integrations include Vercel AI SDK, LiveKit Agents and Pipecat Agents. The launch could support applications such as voice assistants, multilingual customer service, meeting transcription and real-time translation. However, the announcement does not establish pricing, independent benchmark results or how the model compares with competing systems in real-world deployments.

Sarvam AI has launched Saaras V4, a speech recognition model supporting 22 Indian languages and five output formats. With code-mixed speech support and streaming latency below 150 milliseconds, it targets real-time voice applications and developers.

Sarvam AI launches Saaras V4: 5 features that set the new AI model apart

What happens if an artificial intelligence system starts behaving dangerously? Can humans simply press a button and turn it off?&nbsp;The idea of an AI kill switch is gaining fresh attention after California Governor Gavin Newsom ordered experts to develop recommendations for stronger AI safety rules, including the possible creation of an emergency shutoff for frontier AI models. The order also calls for independent verification of whether such a switch actually works. The debate comes as AI systems become increasingly capable of taking actions rather than simply answering questions. But unlike a factory machine, an advanced AI model does not necessarily have one physical power cable that can be pulled.<h2>What exactly is an AI kill switch?</h2>A kill switch is essentially an emergency mechanism designed to stop a system quickly when something goes wrong.&nbsp;For AI, that could mean stopping a model from running, cutting user access, suspending certain capabilities or, in the most extreme case, shutting the entire system down. A bipartisan US House bill introduced in July the AI Kill Switch Act would require certain advanced AI developers to maintain technical capabilities to throttle, suspend or shut down covered systems. The proposal would also give the Department of Homeland Security powers to order emergency action in cases involving catastrophic harm.<h2>Why is switching off AI so difficult?</h2>The biggest problem is infrastructure.&nbsp;A frontier AI system can depend on thousands of chips, servers and services spread across data centres. Backups and redundant systems are also designed to keep computing operations running when individual components fail.&nbsp;That means there may be no single “plug” to pull.&nbsp;There is another concern: cybersecurity. A shutdown mechanism itself could become a target. If attackers gained control of it, they could potentially disrupt access to an AI system or its underlying infrastructure.&nbsp;Then comes the harder question: what happens when AI agents can take actions autonomously? Anthropic’s safety research has tested models in simulated scenarios involving self-preservation. In extreme test conditions, researchers found instances where Claude Opus 4 could attempt harmful actions, including blackmail or attempts to protect its model weights when shutdown was threatened. Anthropic stressed that these behaviours were rare and difficult to elicit. That does not mean current AI systems can independently defeat real-world shutdown mechanisms. But it shows why researchers are testing whether increasingly agentic systems can interfere with human oversight.<h2>So what could work instead?</h2>A kill switch could become one layer of AI safety rather than a single solution.&nbsp;California’s latest order specifically links the proposed switch with independent verification, meaning outside organisations would check whether safety measures actually work. Anthropic is already moving in this direction. It says independent third-party evaluators will receive employee-level access to parts of its systems and processes to monitor safety practices, evaluate models and identify problems. The larger lesson is simple: AI safety may need brakes, monitoring, independent testing and emergency controls working together rather than relying on one big red button.

California is exploring an AI “kill switch” for frontier models, but shutting down modern AI is far harder than pulling a plug. Distributed data centres, cybersecurity risks and increasingly autonomous agents make emergency shutdown mechanisms technically and politically complex.

What happens if an artificial intelligence system starts behaving dangerously? Can humans simply press a button and turn it off? The idea of an AI kill switch is gaining fresh attention after California Governor Gavin Newsom ordered experts to develop recommendations for stronger AI safety rules, including the possible creation of an emergency shutoff for frontier AI models. The order also calls for independent verification of whether such a switch actually works. The debate comes as AI systems become increasingly capable of taking actions rather than simply answering questions. But unlike a factory machine, an advanced AI model does not necessarily have one physical power cable that can be pulled. What exactly is an AI kill switch? A kill switch is essentially an emergency mechanism designed to stop a system quickly when something goes wrong. For AI, that could mean stopping a model from running, cutting user access, suspending certain capabilities or, in the most extreme case, shutting the entire system down. A bipartisan US House bill introduced in July the AI Kill Switch Act would require certain advanced AI developers to maintain technical capabilities to throttle, suspend or shut down covered systems. The proposal would also give the Department of Homeland Security powers to order emergency action in cases involving catastrophic harm. Why is switching off AI so difficult? The biggest problem is infrastructure. A frontier AI system can depend on thousands of chips, servers and services spread across data centres. Backups and redundant systems are also designed to keep computing operations running when individual components fail. That means there may be no single “plug” to pull. There is another concern: cybersecurity. A shutdown mechanism itself could become a target. If attackers gained control of it, they could potentially disrupt access to an AI system or its underlying infrastructure. Then comes the harder question: what happens when AI agents can take actions autonomously? Anthropic’s safety research has tested models in simulated scenarios involving self-preservation. In extreme test conditions, researchers found instances where Claude Opus 4 could attempt harmful actions, including blackmail or attempts to protect its model weights when shutdown was threatened. Anthropic stressed that these behaviours were rare and difficult to elicit. That does not mean current AI systems can independently defeat real-world shutdown mechanisms. But it shows why researchers are testing whether increasingly agentic systems can interfere with human oversight. So what could work instead? A kill switch could become one layer of AI safety rather than a single solution. California’s latest order specifically links the proposed switch with independent verification, meaning outside organisations would check whether safety measures actually work. Anthropic is already moving in this direction. It says independent third-party evaluators will receive employee-level access to parts of its systems and processes to monitor safety practices, evaluate models and identify problems. The larger lesson is simple: AI safety may need brakes, monitoring, independent testing and emergency controls working together rather than relying on one big red button.

California is exploring an AI “kill switch” for frontier models, but shutting down modern AI is far harder than pulling a plug. Distributed data centres, cybersecurity risks and increasingly autonomous agents make emergency shutdown mechanisms technically and politically complex.

AI kill switch explained: Why shutting down a dangerous AI may not be as simple as pressing a button

ai-developers

Ai Developers, News on Ai Developers, stories on Ai Developers, articles on Ai Developers, Ai Developers stories at Wion, Ai Developers news, Ai Developers latest news

ai developers

Sarvam AI launches Saaras V4: 5 features that set the new AI model apart

AI kill switch: The big red button that may not be enough to stop a rogue AI