Search “Speech”

Block releases speech model Dessa

Block's internal AI research lab released the full-duplex speech-to-speech model Dessa on October 23. According to Unite.AI, the model was trained entirely from scratch on human conversational speech, can listen and speak simultaneously, and natively perceives emotion, sarcasm, and laughter; its inference is faster than real time and it can run locally on a MacBook Pro M4 Max or higher configuration, with the team completing training on a single NVIDIA GB200 NVL72. Block has released demo recordings and a public waitlist, and has opened early access to a small number of AI safety researchers. Other reports note that the model does not rely on a text intermediary and supports full-duplex conversation with the ability to interrupt and interject at any time, and that training consumes hundreds of times fewer tokens than large language models, but the author points out that this is currently only a small-scale research preview from Block's internal lab, the demos were selected by the company, and questions remain about how pure audio models perform on complex multi-step logical reasoning.

View story:Block releases full-duplex speech model Dessa

Voice researcher suggests families set up a safe word

Kimberly Hyun, an assistant professor at the University of Cincinnati, suggests that family members agree on a safe word known only to each other; if you receive a call from an unknown number claiming a family member is in trouble, you should hang up immediately and call that person back to verify. Fortune reported on the case of Deborah Del Mastro, a mother in Martinez, California: in May she received a call from someone posing as a Mexican drug cartel, and the caller played an AI-cloned recording of her daughter crying and screaming. Following the instructions, she wired $5,400 to Mexico from multiple locations within five hours, only to discover afterward that her daughter had been at work the whole time. The report also noted that McAfee tests showed that a voice sample of just three seconds is enough to generate a clone that matches 85% of the original voice; a survey by Gallup and the Stop Scams Alliance estimated that in 2025 about 15 million American adults were scammed, with losses of about $68 billion, of which 12% involved AI or deepfakes.

Voice researcher suggests families establish a safety code word

Kimberly Hyun, an assistant professor at the University of Cincinnati, suggests that family members agree on a safe word known only to each other; if you receive a call from an unknown number claiming a family member is in trouble, you should hang up immediately and call that person back to verify. Fortune reported on the case of Deborah Del Mastro, a mother in Martinez, California: in May she received a call from someone posing as a Mexican drug cartel, and the caller played an AI-cloned recording of her daughter crying and screaming. Following the instructions, she wired $5,400 to Mexico from multiple locations within five hours, only to discover afterward that her daughter had been at work the whole time. The report also noted that McAfee tests showed that a voice sample of just three seconds is enough to generate a clone that matches 85% of the original voice; a survey by Gallup and the Stop Scams Alliance estimated that in 2025 about 15 million American adults were scammed, with losses of about $68 billion, of which 12% involved AI or deepfakes.

View story:Voice researcher suggests families set up a safe word

Amazon Polly upgrades generative engine, adds 14 voices

Amazon Polly completed a generative engine upgrade on October 23, 2026. According to information released by AWS Machine Learning, this upgrade adds 14 voices, covers 9 locales, and for the first time officially offers Hindi and Finnish in the generative engine. After the upgrade, Polly offers a total of 57 generative voices, covering 11 languages and 25 locales, with service scope extended to AWS GovCloud (US-West), covering a total of 11 AWS Regions. This is an in-place upgrade; existing voice names and API calls remain unchanged, pricing remains unchanged, and users do not need to modify code. There are currently no subsequent development facts.

Sierra launches voice agent entity detail benchmark τ-Entity

On October 22, 2026, Sierra released the τ-Entity benchmark, designed to test whether voice agents can accurately capture, recognize, and correct errors in entity details such as names, phone numbers, addresses, and email addresses. The benchmark includes 200 tasks, covers ten categories of entities, and features a six-language leaderboard. According to the leaderboard, GPT Live 1 achieved a 60.3% accuracy rate in regular calls, but only 34.0% of tasks were passed under all three conditions: regular, strong noise, and strong speech. Gemini 3.8 Live and Grok Voice Think Fast 2.0 had all-condition pass rates of 19.5% and 24.0%, respectively. The paper also found that prescribing a "ask—spell/repeat—confirm correction—submit" workflow can increase the all-condition success rate by 14 to 31 percentage points, at the cost of 21 to 28 extra seconds per call. There are currently no follow-up developments.

Sierra launches voice Agent entity detail benchmark 𝜏-Entity

On October 22, 2026, Sierra released the τ-Entity benchmark, designed to test whether voice agents can accurately capture, recognize, and correct errors in entity details such as names, phone numbers, addresses, and email addresses. The benchmark includes 200 tasks, covers ten categories of entities, and features a six-language leaderboard. According to the leaderboard, GPT Live 1 achieved a 60.3% accuracy rate in regular calls, but only 34.0% of tasks were passed under all three conditions: regular, strong noise, and strong speech. Gemini 3.8 Live and Grok Voice Think Fast 2.0 had all-condition pass rates of 19.5% and 24.0%, respectively. The paper also found that prescribing a "ask—spell/repeat—confirm correction—submit" workflow can increase the all-condition success rate by 14 to 31 percentage points, at the cost of 21 to 28 extra seconds per call. There are currently no follow-up developments.

View story:Sierra launches voice agent entity detail benchmark τ-Entity

MiniMax showcases M3.1 Flash and SGLang Omni voice agent combination

MiniMax showcased a solution combining M3.1 Flash with SGLang Omni for interactive voice agents. According to its introduction, M3.1 Flash provides fast, concise response capabilities, while SGLang Omni supplements streaming output, dialogue state, and interruption handling, enabling the voice agent to be interrupted, corrected, and redirected in real time. A developer replaced the local LLM in the original Qwen3-ASR→Qwen3 LLM→Qwen3-TTS pipeline with the M3.1 Flash API, keeping the rest of the pipeline unchanged, and measured fast responses with concise, direct answers. Currently, this combination is still in the showcase and developer trial stage, with no further progress disclosed.

MiniMax showcases M3.1 Flash and SGLang Omni voice agent combination

MiniMax showcased a solution combining M3.1 Flash with SGLang Omni for interactive voice agents. According to its introduction, M3.1 Flash provides fast, concise response capabilities, while SGLang Omni supplements streaming output, dialogue state, and interruption handling, enabling the voice agent to be interrupted, corrected, and redirected in real time. A developer replaced the local LLM in the original Qwen3-ASR→Qwen3 LLM→Qwen3-TTS pipeline with the M3.1 Flash API, keeping the rest of the pipeline unchanged, and measured fast responses with concise, direct answers. Currently, this combination is still in the showcase and developer trial stage, with no further progress disclosed.

View story:MiniMax showcases M3.1 Flash and SGLang Omni voice agent combination

HeyGen releases speech model Voice Turbo

HeyGen releases speech model HeyGen Voice Turbo. Artificial Analysis reported on October 22, 2026 that the model ranked first in the controlled speech TTS arena with an Elo of 1,202, surpassing the previous generation HeyGen Voice's 1,196 and Qwen-Audio-3.1-TTS-Plus's 1,185. The arena evaluated using the same set of 8 cloned voices (4 American, 4 British), and HeyGen Voice Turbo ranked first among 46 models and ranked first in the Assistants and Entertainment categories. The report also stated its price is 4 USD per million characters, the lowest on the list, 7.5 times cheaper than the previous generation HeyGen Voice's 30 USD per million characters. There are currently no subsequent development facts.

HeyGen releases voice model Voice Turbo

HeyGen releases speech model HeyGen Voice Turbo. Artificial Analysis reported on October 22, 2026 that the model ranked first in the controlled speech TTS arena with an Elo of 1,202, surpassing the previous generation HeyGen Voice's 1,196 and Qwen-Audio-3.1-TTS-Plus's 1,185. The arena evaluated using the same set of 8 cloned voices (4 American, 4 British), and HeyGen Voice Turbo ranked first among 46 models and ranked first in the Assistants and Entertainment categories. The report also stated its price is 4 USD per million characters, the lowest on the list, 7.5 times cheaper than the previous generation HeyGen Voice's 30 USD per million characters. There are currently no subsequent development facts.

View story:HeyGen releases speech model Voice Turbo

English translation is being prepared.

English translation is being prepared.

View story:English translation is being prepared.

Ethan Mollick suggests voice interfaces could end or worsen open-plan offices

Ethan Mollick put forward a judgment: voice interfaces in AI applications may finally be able to end the "monster" that is the open-plan office, or they may simply make it worse. Citing his own research, he pointed out that multiple experiments show that background noise in open-plan offices reduces the ability to do analytical work, increases software bugs, and weakens the ability to find bugs. This judgment was published on October 21, 2026, and there are currently no subsequent developments.

Ethan Mollick suggests voice interfaces may end or worsen open-plan offices

Ethan Mollick put forward a judgment: voice interfaces in AI applications may finally be able to end the "monster" that is the open-plan office, or they may simply make it worse. Citing his own research, he pointed out that multiple experiments show that background noise in open-plan offices reduces the ability to do analytical work, increases software bugs, and weakens the ability to find bugs. This judgment was published on October 21, 2026, and there are currently no subsequent developments.

View story:Ethan Mollick suggests voice interfaces could end or worsen open-plan offices

ElevenLabs launches real-time speech translation model Eleven Realtime Translation

ElevenLabs launches the real-time speech translation model Eleven Realtime Translation, currently available in alpha form to ElevenAPI enterprise customers. The model uses the low-latency speech model Eleven v4 Turbo by default, supports more than 90 languages and automatic language detection, keeps the same voice consistent across languages, and can be selected from a library of more than 17,500 voices or a self-built voice clone. Teams can adjust the translation startup speed, trading off between high-quality translation and low latency for complex calls. Deutsche Telekom has already built new call features based on ElevenLabs in the Magenta AI Call Assistant.

ElevenLabs launches real-time speech translation model Eleven Realtime Translation

ElevenLabs launches the real-time speech translation model Eleven Realtime Translation, currently available in alpha form to ElevenAPI enterprise customers. The model uses the low-latency speech model Eleven v4 Turbo by default, supports more than 90 languages and automatic language detection, keeps the same voice consistent across languages, and can be selected from a library of more than 17,500 voices or a self-built voice clone. Teams can adjust the translation startup speed, trading off between high-quality translation and low latency for complex calls. Deutsche Telekom has already built new call features based on ElevenLabs in the Magenta AI Call Assistant.

View story:ElevenLabs launches real-time speech translation model Eleven Realtime Translation

ElevenLabs releases the Eleven Realtime v1 speech-to-speech model

On October 20, 2026, ElevenLabs released the Eleven Realtime series of speech-to-speech models, comprising v1 Mini for low-cost natural conversation and v1 Max for complex tasks. According to the ElevenLabs blog, both support more than 90 languages and can access its over 17,000 voices; an independent evaluation by Artificial Analysis shows that v1 Max answered 98.8% of Big Bench Audio speech reasoning questions correctly and began speaking in an average of 0.57 seconds, while v1 Mini scored 75.9% and 0.62 seconds, respectively. The API is compatible with the OpenAI Realtime API, so existing integrations need only change the endpoint and model name, and models can also be selected directly for deployment in ElevenAgents. On October 22, Zhidongxi reported that the two models currently support only WebSocket connections; in its testing, an assistant built with Architect could modify Agent settings by voice, but the Chinese accent was noticeable, responses tended to be long, and latency was relatively high.

English translation is being prepared.

English translation is being prepared.

View story:English translation is being prepared.

ElevenLabs releases speech-to-speech model Eleven Realtime v1

On October 20, 2026, ElevenLabs released the Eleven Realtime series of speech-to-speech models, comprising v1 Mini for low-cost natural conversation and v1 Max for complex tasks. According to the ElevenLabs blog, both support more than 90 languages and can access its over 17,000 voices; an independent evaluation by Artificial Analysis shows that v1 Max answered 98.8% of Big Bench Audio speech reasoning questions correctly and began speaking in an average of 0.57 seconds, while v1 Mini scored 75.9% and 0.62 seconds, respectively. The API is compatible with the OpenAI Realtime API, so existing integrations need only change the endpoint and model name, and models can also be selected directly for deployment in ElevenAgents. On October 22, Zhidongxi reported that the two models currently support only WebSocket connections; in its testing, an assistant built with Architect could modify Agent settings by voice, but the Chinese accent was noticeable, responses tended to be long, and latency was relatively high.

View story:ElevenLabs releases the Eleven Realtime v1 speech-to-speech model