Perplexity has upgraded the voice mode of its Mac app, switching its underlying model to OpenAI's gpt-live-1 voice model, which can handle multiple Computer sessions simultaneously within a single voice session, launch new tasks, and create and manage Projects. Perplexity also recently launched the ability to run AI directly on Mac locally, supporting offline mode and privacy-focused AI workflows.
Matched fieldsTitle、Summary、Body
AnySpeechRecognizer supports engine switching on top of the Apple SpeechAnalyzer API, with options including NVIDIA Parakeet TDT 0.6B v3 (supporting 25 European languages), Moondream post-trained Parakeet Ultra, the English version Parakeet Unified, and Alibaba Qwen3-ASR (0.6B/1.7B, covering 30 languages and 22 Chinese dialects). These models have been converted to Core ML and published on the Hugging Face Hub; Parakeet takes about 20–40 seconds to transcribe 1 hour of audio, while Qwen3-ASR is slower and requires several GB of memory. The package also adds a speaker diarization module, SpeakerDiarizer, which the Apple framework lacks; it is based on NVIDIA Streaming Sortformer and can track up to 4 speakers.
AnySpeechRecognizer supports engine switching on top of the Apple SpeechAnalyzer API, with options including NVIDIA Parakeet TDT 0.6B v3 (supporting 25 European languages), Moondream post-trained Parakeet Ultra, the English version Parakeet Unified, and Alibaba Qwen3-ASR (0.6B/1.7B, covering 30 languages and 22 Chinese dialects). These models have been converted to Core ML and published on the Hugging Face Hub; Parakeet takes about 20–40 seconds to transcribe 1 hour of audio, while Qwen3-ASR is
Matched fieldsTitle、Body
Cartesia has launched the Managed Agents end-to-end voice agent platform, targeting enterprise call scenarios. On Artificial Analysis's Speech Agent Arena, it ranks first among 30 agents with a preference Elo of 1115. On ServiceNow's EVA-Bench, the platform achieved an accuracy of 87.4% and an experience score of 90.7%, both the highest. The platform co-designs the voice model with the harness, offering capabilities such as 44 languages, tool calling, knowledge bases, and outbound calling batches. The company says it has already supported tens of millions of conversations in finance, retail, hospitality, and healthcare scenarios.
Cartesia has launched the Managed Agents end-to-end voice agent platform, targeting enterprise call scenarios. On Artificial Analysis's Speech Agent Arena, it ranks first among 30 agents with a preference Elo of 1115. On ServiceNow's EVA-Bench, the platform achieved an accuracy of 87.4% and an experience score of 90.7%, both the highest. The platform co-designs the voice model with the harness, offering capabilities such as 44 languages, tool calling, knowledge bases, and outbound calling batche
Matched fieldsTitle、Summary、Body
ElevenLabs has released a user guide explaining turn-taking in conversational AI: the model must switch among three actions—replying, holding the floor, and listening to silence—relying on adaptive timing, VAD, and semantic analysis to determine whether the user has finished speaking, rather than depending on a fixed silence timer. When Eleven v4 Turbo is paired with ElevenAgents, the median latency from request to audible speech is 142 ms. Engineering teams can measure conversational fluency using response latency and interruption rate.
ElevenLabs has released a user guide explaining turn-taking in conversational AI: the model must switch among three actions—replying, holding the floor, and listening to silence—relying on adaptive timing, VAD, and semantic analysis to determine whether the user has finished speaking, rather than depending on a fixed silence timer. When Eleven v4 Turbo is paired with ElevenAgents, the median latency from request to audible speech is 142 ms. Engineering teams can measure conversational fluency us
Matched fieldsTitle、Summary、Body
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight large speech model that supports 40 languages and up to 1 hour of audio per pass, and can select output formats such as verbatim, written, speaker-labeled, contextual, and translation via text instructions. The model has approximately 34B total parameters and about 4.3B activated parameters, adopts a hybrid Mamba-Transformer MoE architecture, can run in BF16 on a single H100 80GB, and the NVFP4 version at about 21GB can run on an RTX 5090. NVIDIA self-reports an average WER of 4.71% across 9 public English Open ASR test sets, the best result among open-weight models, though commercial APIs still lead; its shortcomings are that it only supports offline operation, and the Chinese far-field meeting cpCER is 34.95%.
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight large speech model that supports 40 languages and up to 1 hour of audio per pass, and can select output formats such as verbatim, written, speaker-labeled, contextual, and translation via text instructions. The model has approximately 34B total parameters and about 4.3B activated parameters, adopts a hybrid Mamba-Transformer MoE architecture, can run in BF16 on a single H100 80GB, and the NVFP4 version at about 21GB can run on an RTX 50
Matched fieldsTitle、Summary、Body
AssemblyAI has released its first text-to-speech model, AssemblyAI-TTS-1, and officially opened it up, targeting real-time telephony voice agents, priced at $15 per million characters, with no annual contracts and no concurrency tiers. The model is offered only as a hosted API in us-east-1 and eu-north-1, supports HTTP, WebSocket and OpenAI-compatible interfaces, and comes with more than 4,000 voices, instant voice cloning and 6 languages. The vendor's own testing shows 93.7% entity accuracy and a median time to first byte of 47 ms, but Cekura's independent benchmark measured 159 ms streaming latency and 81.5% reading accuracy, and at launch it does not yet support SSML, speaking-rate control or emotion tags.
AssemblyAI has released its first text-to-speech model, AssemblyAI-TTS-1, and officially opened it up, targeting real-time telephony voice agents, priced at $15 per million characters, with no annual contracts and no concurrency tiers. The model is offered only as a hosted API in us-east-1 and eu-north-1, supports HTTP, WebSocket and OpenAI-compatible interfaces, and comes with more than 4,000 voices, instant voice cloning and 6 languages. The vendor's own testing shows 93.7% entity accuracy and
Matched fieldsTitle、Summary、Body
Block's AI research lab has released Dessa, a full-duplex speech model, opening a waitlist in the form of a research preview. It is trained only on human voice conversation audio, without using text or pretrained models. The model can run locally on a MacBook Pro M4 Max or higher configuration, faster than real time; training used one NVIDIA GB200 NVL72 (72 GPUs). No benchmark scores, parameter count, weights, paper, or API have been published so far.
Matched fieldsTitle、Summary、Body
Databricks argues that the value of a voice agent lies not in being able to converse, but in combining context, ontology, and governed data to understand callers in real time and execute rules. Its prototype uses LiveKit to carry calls, Deepgram for speech-to-text and default voice, and ElevenLabs for expressive voice, and can hand calls over to a human. The agent runs where customer data resides, connecting to order, loyalty, and CRM systems through read-only queries, and uses the DIVA blueprint as the starting point for the next scenario.
Databricks argues that the value of a voice agent lies not in being able to converse, but in combining context, ontology, and governed data to understand callers in real time and execute rules. Its prototype uses LiveKit to carry calls, Deepgram for speech-to-text and default voice, and ElevenLabs for expressive voice, and can hand calls over to a human. The agent runs where customer data resides, connecting to order, loyalty, and CRM systems through read-only queries, and uses the DIVA blueprin
Matched fieldsTitle、Summary、Body
Alex is a fully DIY AI voice assistant that uses only an ESP32-D0WD-V3 development board without PSRAM, together with an INMP441 microphone, a MAX98357A amplifier, and an SSD1306 OLED, to achieve a real-time closed loop of voice capture, Whisper transcription, GPT-generated responses, and TTS streaming playback. The project places the OLED animation on an independent FreeRTOS task on the second core, preventing the 1-to-3-second HTTP calls from freezing the display each time. During debugging, issues such as insufficient contiguous memory for the TLS handshake, an incorrect SSD1306_EXTERNALVCC power supply mode, and excessively fast TTS speech rate were also resolved.
Alex is a fully DIY AI voice assistant that uses only an ESP32-D0WD-V3 development board without PSRAM, together with an INMP441 microphone, a MAX98357A amplifier, and an SSD1306 OLED, to achieve a real-time closed loop of voice capture, Whisper transcription, GPT-generated responses, and TTS streaming playback. The project places the OLED animation on an independent FreeRTOS task on the second core, preventing the 1-to-3-second HTTP calls from freezing the display each time. During debugging, i
Matched fieldsTitle、Summary、Body
Anthropic has upgraded the Claude mobile app. The voice mode on the iPhone version now supports Apple CarPlay, allowing users to ask Claude questions directly while driving. Claude has added 24 new languages across web, desktop, iOS and Android, bringing total support to 33 languages, while voice mode responds faster and recognizes speech more accurately in noisy environments. The iOS and Android versions also support logging in to multiple accounts simultaneously and switching between them directly.
Anthropic has upgraded the Claude mobile app. The voice mode on the iPhone version now supports Apple CarPlay, allowing users to ask Claude questions directly while driving. Claude has added 24 new languages across web, desktop, iOS and Android, bringing total support to 33 languages, while voice mode responds faster and recognizes speech more accurately in noisy environments. The iOS and Android versions also support logging in to multiple accounts simultaneously and switching between them dire
Matched fieldsTitle、Summary、Body
DingTalk A1 Doukou Doctor Edition combines A1's voice recording and AI summary capabilities with Doukou's medical evidence capabilities. After doctors finish organizing case discussions or lecture content, they can also search literature, guidelines, and expert consensus around unresolved medical questions. A gynecology department director at a tertiary hospital in Hangzhou reported that on surgery days she can first record special situations and judgments, then organize them after surgery, while for complex outpatient visits she uses Doukou Doctor to view literature and guideline links. At this stage, it is more like a portable professional assistant that can record, distill, and search literature, rather than a reconstruction of hospital IT systems.
DingTalk A1 Doukou Doctor Edition combines A1's voice recording and AI summary capabilities with Doukou's medical evidence capabilities. After doctors finish organizing case discussions or lecture content, they can also search literature, guidelines, and expert consensus around unresolved medical questions. A gynecology department director at a tertiary hospital in Hangzhou reported that on surgery days she can first record special situations and judgments, then organize them after surgery, whil
Matched fieldsTitle、Summary、Body
Anthropic has upgraded the Claude mobile app. The iPhone version's voice mode now supports Apple CarPlay, allowing users to ask Claude questions directly while driving. This update also lets the iOS and Android versions support simultaneous login to multiple accounts with direct switching. Claude has added 24 new languages, bringing total support to 33 languages, and voice mode now responds faster and recognizes speech more accurately in noisy environments.
Matched fieldsTitle、Summary、Body
At its inaugural India Summit, ElevenLabs signed an MoU with the Government of Karnataka, India, to explore the use of voice AI for citizen services, voice security, and voice restoration. The collaboration will begin with pilots in grievance handling, public hotlines, and welfare applications, covering Kannada and the 13 other official Indian languages already supported by the platform. ElevenLabs will also deploy a security framework covering deepfake detection, watermarking, authorization, and voice cloning restrictions, and will provide eligible residents with free lifetime voice reconstruction licenses through the One Million Voices program.
At its inaugural India Summit, ElevenLabs signed an MoU with the Government of Karnataka, India, to explore the use of voice AI for citizen services, voice security, and voice restoration. The collaboration will begin with pilots in grievance handling, public hotlines, and welfare applications, covering Kannada and the 13 other official Indian languages already supported by the platform. ElevenLabs will also deploy a security framework covering deepfake detection, watermarking, authorization, an
Matched fieldsTitle、Summary、Body
ElevenLabs announced a partnership with the Spanish ALS association adELA to give ALS patients diagnosed in Spain free access to its voice restoration and preservation technology. This is the first project of the ElevenLabs Impact Program in Spain, and Spain is the first country to offer this technology through a nationwide patient association. Patients whose speech has not yet noticeably changed can record new samples, while those whose speech has changed can upload old recordings to generate their own digital voice. It is estimated that there are approximately 4,000–4,500 ALS patients in Spain, many of whom will eventually lose the ability to speak.
ElevenLabs announced a partnership with the Spanish ALS association adELA to give ALS patients diagnosed in Spain free access to its voice restoration and preservation technology. This is the first project of the ElevenLabs Impact Program in Spain, and Spain is the first country to offer this technology through a nationwide patient association. Patients whose speech has not yet noticeably changed can record new samples, while those whose speech has changed can upload old recordings to generate t
Matched fieldsTitle、Summary、Body
Cartesia Managed Agents achieved an Elo of 1115 on the Artificial Analysis Speech Agent Arena, 98 points higher than the previous Cartesia Line (1017), with task success rate rising from 77.5% to 91.7%. The platform is Cartesia's fully managed voice agent solution, running in a cascade around a developer-selected LLM with its self-developed Ink speech-to-text and Sonic text-to-speech models; this evaluation used GPT-5.5 with the Skylar and Daniel voices. On agent tasks requiring tool calls, the Elo was 1112.
Cartesia Managed Agents achieved an Elo of 1115 on the Artificial Analysis Speech Agent Arena, 98 points higher than the previous Cartesia Line (1017), with task success rate rising from 77.5% to 91.7%. The platform is Cartesia's fully managed voice agent solution, running in a cascade around a developer-selected LLM with its self-developed Ink speech-to-text and Sonic text-to-speech models; this evaluation used GPT-5.5 with the Skylar and Daniel voices. On agent tasks requiring tool calls, the
Matched fieldsTitle、Summary、Body
Three studies by ElevenLabs of shoppers in the US, UK and other countries show that when shoppers were asked to actually use the ElevenLabs voice Agent first before answering, nearly half chose "speak and listen to the reply," about 30% chose "speak and read text on screen," and voice ranked first across all six shopping task categories. Among respondents who had not tried it, preference for typing led voice by more than 4:1. The more serious the problem, the less willing shoppers were to hand it over to the Agent; in the erroneous charge scenario, only 34% chose the Agent, and if the Agent failed once, 61% demanded transfer to a human.
Three studies by ElevenLabs of shoppers in the US, UK and other countries show that when shoppers were asked to actually use the ElevenLabs voice Agent first before answering, nearly half chose "speak and listen to the reply," about 30% chose "speak and read text on screen," and voice ranked first across all six shopping task categories. Among respondents who had not tried it, preference for typing led voice by more than 4:1. The more serious the problem, the less willing shoppers were to hand i
Matched fieldsTitle、Summary、Body
ElevenLabs has released a practical guide on how to build an AI voice assistant by chaining speech-to-text, an LLM, and text-to-speech into a real-time pipeline. The guide recommends keeping round-trip responses within 500 ms and pairing this with VAD, while also connecting to internal knowledge bases and business tools, and putting safety measures such as verbal confirmation and handoff to a human agent in place. Developers can build directly with ElevenAPI, or configure and launch on the ElevenAgents managed platform.
ElevenLabs has released a practical guide on how to build an AI voice assistant by chaining speech-to-text, an LLM, and text-to-speech into a real-time pipeline. The guide recommends keeping round-trip responses within 500 ms and pairing this with VAD, while also connecting to internal knowledge bases and business tools, and putting safety measures such as verbal confirmation and handoff to a human agent in place. Developers can build directly with ElevenAPI, or configure and launch on the Eleve
Matched fieldsTitle、Summary、Body
Vapi has introduced a traffic splitting feature that can allocate inbound and outbound call traffic to different versions of voice agents at ratios such as 0.1%, 1%, or 5%, before a full release once validated. The feature supports maintaining version stickiness by caller identity, and the official recommendation is to start with a two-way split, changing only one variable at a time. Previously, releasing a new version would immediately switch 100% of traffic over, so teams did not dare to iterate frequently.
Vapi has introduced a traffic splitting feature that can allocate inbound and outbound call traffic to different versions of voice agents at ratios such as 0.1%, 1%, or 5%, before a full release once validated. The feature supports maintaining version stickiness by caller identity, and the official recommendation is to start with a two-way split, changing only one variable at a time. Previously, releasing a new version would immediately switch 100% of traffic over, so teams did not dare to itera
Matched fieldsTitle、Summary、Body
Text-to-speech (TTS) works in two stages: the front end first normalizes the text and converts it into tokens and vector representations, and then the back end uses a neural model to generate acoustic representations, which are output as speech waveforms through a vocoder. In the article, ElevenLabs uses "Mr. Ríos paid $11 on 10/06" as an example to demonstrate the complete pipeline from normalization, tokenization, and prosody to the acoustic model and vocoder, and mentions that pronunciation dictionaries and phoneme annotations can be used to fine-tune pronunciation. Enterprises can connect their business applications to high-quality speech generation through the TTS API.
Text-to-speech (TTS) works in two stages: the front end first normalizes the text and converts it into tokens and vector representations, and then the back end uses a neural model to generate acoustic representations, which are output as speech waveforms through a vocoder. In the article, ElevenLabs uses "Mr. Ríos paid $11 on 10/06" as an example to demonstrate the complete pipeline from normalization, tokenization, and prosody to the acoustic model and vocoder, and mentions that pronunciation d
Matched fieldsTitle、Summary、Body
Google has begun rolling out Gemini to users in Google Maps on Apple CarPlay. A Gemini star icon appears on the right side of the navigation interface; tapping it lets users issue multi-step, open-ended commands by voice. Users can ask Gemini to recommend scenic spots, check how crowded a place is and suggest quieter times, specify routes that avoid toll roads, and report congestion and accidents while driving. The feature launches first in the United States and will expand globally in the coming weeks.
Google has begun rolling out Gemini to users in Google Maps on Apple CarPlay. A Gemini star icon appears on the right side of the navigation interface; tapping it lets users issue multi-step, open-ended commands by voice. Users can ask Gemini to recommend scenic spots, check how crowded a place is and suggest quieter times, specify routes that avoid toll roads, and report congestion and accidents while driving. The feature launches first in the United States and will expand globally in the comin
Matched fieldsTitle、Summary