This AWS post explains how Cartesia, an AWS Partner, uses State Space Models (SSMs) and AWS infrastructure to deliver sub-100ms text-to-speech (Sonic) and speech-to-text (Ink) at global scale. It details two deployment patterns for Sonic: an Amazon SageMaker AI model package via AWS Marketplace running in-VPC, and self-managed containers on EC2 GPU instances, and describes Cartesia's voice models operating inside Amazon Connect.
Flow AI Fx presents itself as a unified AI creative studio offering four generators—Video, Image, Music, and Voice—for turning prompts or reference images into first drafts. It describes a four-step workflow: choose a starting point, pick a format, explore and refine, then audition and export studio-quality files. The page frames the tools as a starting point for exploration rather than a substitute for the user's artistic decisions.
Angle Health, an AI-native health benefits technology company, piloted ElevenAgents voice agents for provider-service workflows, answering calls in eight languages and working directly with its core administration system. During the pilot, the agents handled about 80% of total provider call volume, with more than 90% resolved end to end without transferring to a person, and the provider team gained back nearly 9,000 minutes a day. Angle Health plans to expand the pilot to its commercial, brokerage and account management segments.
Apple has begun beta testing its AI-powered "Audio Intelligence" features for Apple Watch Series 12 and Apple Watch Ultra 4, initially for developers in the iOS 27.3 update, after announcing them at its September hardware event. The four opt-in features include sound recognition, Shazam-based song recognition, Live Rewind (which shows the last 15 seconds of speech as a text snippet when you double-press the digital crown), and Siri Recaps, which summarize conversations into short notes without saving raw audio or identifying speakers. Apple says processing relies on the S11 chip, the iPhone, and Private Cloud Compute, with snippets deleted after 30 seconds if unsaved and recaps deleted within seven days if not reviewed.
Synthetic voices are built by training neural AI models on large datasets of real human speech, then generating audio from text or cloning a specific speaker from recorded samples. Modern systems sound natural because they replicate prosody and interpret context rather than stitching together individual words or syllables. Voices can also be designed from a text description alone, without any recording.
AI-TRANSLATOR is a synchronous speech and text translator whose translation core emits normalized events to an event bus, which fans them out to any number of independent text sinks (console, WebSocket display page, transcript, Telegram) and voice sinks in parallel. Stage 1 (MVP) is implemented and tested, with Linux as the primary platform, macOS and Windows x86_64 supported for building and running, and a Telegram voice-chat sink built on the libntgcalls media engine. Interfaces may still change before the 1.0 release.
KPN partnered with ElevenLabs to build Mobiel AI, an add-on to KPN EEN MKB mobile subscriptions that offers voicemail transcription, call summaries and transcripts with follow-up actions, and call routing and reachability settings. KPN said ElevenLabs stood out for speech quality and conversational AI in Dutch, and the teams raised verification accuracy from roughly 50% to 90%. Mobiel AI runs on KPN's network and data centers, with voice data staying in the Netherlands and customer data not used to train AI models or shared with third parties.
A Show HN post presents a shared AI computer that users build and explore just by talking. The interface shows a party/audience view with the microphone off, voice-to-full-duplex code, and a Codex caller, noting that both sides of the conversation stay in place even when they overlap. It states it was built with Full-Duplex Code.
Hume introduced an audio-native decision model that takes audio, a question and possible answers and returns a probability for each answer, aimed at evaluating speech and sound in voice applications. Hume says the model agrees with the human crowd 70.7% of the time versus 56.5% for a single human rater, and scores an average 0.82 across 10 leading benchmarks at a median latency of 0.23 s per screen. It is offered via a Decision Model API as a decision layer in voice-agent systems alongside speech recognition and generation.
A Show HN post links to a personal AI portfolio site featuring RAG, a browser-based text-to-speech highlight engine, and a developer area. The submission has 1 point and no comments on Hacker News.
Retake is a tool that transcribes a talking-head video (MP4, MOV or WebM) so it can be edited like a document: cutting a word cuts the clip, and rewriting a line re-speaks it in the same voice with lips re-synced. Users select words and press Delete to cut, or select words and start typing to rewrite. Claude marks ums, repeats and false starts for cleanup.
Avartha announced Ava Zero, a voice agent built entirely on open models that ranks #1 on Sierra's τ³-Voice benchmark with 86.0% task completion across six languages, ahead of OpenAI's GPT Live 1 (79.9%), xAI's Grok (71.4%), and Google's Gemini (68.5%). The Georgia Tech spinout says it co-designed its inference engine and agent harness as one system, supporting up to an order of magnitude more concurrent conversations per GPU than vLLM and SGLang, and prices Ava at $2.50 per hour of conversation. Its benchmark submission combines Qwen3-ASR, Fish Audio Drama 3, GLM-5.3, and DeepSeek 4.1 Flash.
JJ coders Group published the terms, privacy, DMCA and AI-ethics framework for Suna AI (suna.jjcoders.com), an open 5B flow-matching DiT for 48 KHz music synthesis whose foundation weights are released as jjcoders/suna-v1. The license grants a worldwide, non-exclusive, perpetual, royalty-free right to use, modify and distribute synthesized instrumental audio for personal and commercial projects, with the service provided "as is" and liability capped at fees paid in the prior 12 months or $100, whichever is less. The documents also cover ephemeral GPU-RAM-only processing of uploaded vocals, GDPR/CCPA rights, a DMCA agent at [email protected], and prohibitions on unauthorized vocal clones under the Tennessee ELVIS Act, BIPA and COPPA.
This article compares warm transfers, where the receiving party gets the caller's context before connecting, with cold (blind) transfers, which pass no advance context, and describes a middle "tepid" transfer that dials the destination without a conference or live introduction. It recommends warm transfers for complex, sensitive or high-value calls and cold transfers for simple, high-volume routing, and details how ElevenAgents' transfer_to_number tool supports Conference, Blind and SIP REFER transfer types, including client_message and agent_message parameters and no-answer fallbacks.
The author describes MonsterHouse, a Halloween interactive projection built on the NVIDIA Nemotron Voice Agent with Nemotron Speech ASR, Nemotron 3.5 Bolt E4B 4GB, Magpie TTS Zeroshot and Audio2Face-3D, running on a single workstation with two RTX PRO Blackwell GPUs. The grumpy pirate-ship garage door listens to trick-or-treaters, replies through a LoRA-tuned Bolt E4B model, speaks in a zero-shot Magpie voice with seven moods, and animates its mouth in Unreal Engine 5 via Audio2Face-3D, with optional Nemotron 3 Nano Omni costume descriptions plus YOLO and ByteTrack visitor tracking. The MonsterHouse code is not released, but the post outlines a build path using the open models and reports median end-of-speech-to-audio of 1.70 seconds with Bolt Q4 and TTS, or 0.72 seconds for cached responses.
Starbucks CEO Brian Niccol said on an earnings call that the company will explore experiences that recognize participating customers as they arrive, present personalized digital menus, and make payment more effortless. Starting in 2027, Starbucks will also begin testing background AI models at the drive-thru to help baristas input orders more accurately and efficiently, and it plans to integrate AI into the app's search and discovery features. Niccol said it's still "early" and that the company will "test, learn, and scale what works."
TELUS Digital built an outbound voice agent on ElevenAgents that calls new home internet customers within their first 90 days to explain their first bill, introduce self-serve tools, and prompt bundling discounts. Among customers who received a call, 30-day cancellations were less than half the rate of the broader new-customer cohort, with calls rated 8.5 out of 10 on average and C$55,000 in lifetime revenue retained per 1,000 customers reached. The agent was live in about six weeks from initial vendor engagement, and five additional use cases are now in active development.
Dodaiso (formerly Personal Jarvis) is an open-source, Apache 2.0 agent OS for desktop that runs coding agents, browser and apps by voice or chat on your own machine, with no account or paid tier and support for hosted, subscription or local models. It offers an Agentic IDE with multiple coding agents in live panes addressed by call signs, plus agents, memory, plugins/MCP, a marketplace and approval checks on system-changing actions. It runs on Windows, macOS, Linux or a headless server, installable via a one-line script or desktop installers.
Edge0 has released Audio8-TTS-Realistic, a 0.6B-parameter open-weight text-to-speech model for zero-shot voice cloning from one reference clip and transcript, supporting 23 languages under Apache 2.0. It uses a dual-autoregressive design on Qwen2.5 0.5B and a three-stage training pipeline whose final GRPO stage is guided by a learned realism critic. Edge0 reports 4.86% average error on CV3 across 9 languages, ahead of Qwen3-TTS-1.7B (7.73%) and VoxCPM2 (5.35%), though larger models like Higgs TTS 3 still lead and Thai WER is near 49%.
OTO AI is a desktop AI meeting assistant that listens to live conversations, detects questions, and generates real-time answers using an AI provider of your choice, supporting OpenAI, Anthropic Claude, Google Gemini and DeepSeek via your own API key. It works across Google Meet, Microsoft Teams, Zoom, Webex and other platforms, can use pre-loaded TXT, Markdown, DOCX and PDF documents as context, and can analyse your screen with compatible multimodal models. It is sold as a one-time purchase rather than a subscription, with users paying the AI provider directly for API usage.
This ElevenLabs guide explains turn taking in conversational AI as the rules that determine when a voice agent should reply, hold the floor, or stay silent, relying on adaptive timing, voice activity detection, and semantic analysis rather than strict silence timers. It describes how ElevenAgents detects speech, silence, and intent, handles interruptions and backchannels, and lets teams tune settings like turn eagerness and silence timeouts to avoid premature endpointing, dead air, and false triggers. It also covers measuring quality through interruption rate and response latency, citing a 142 ms median time from request to audible speech with Eleven v4 Turbo and a 200 ms average human turn gap from Journal of Cognition research.
Simple AI introduces Sentiment in Tango, its conversational awareness model, which it says makes Simple the first voice agent on the market to process sentiment directly from audio, scoring confusion, frustration, and happiness from 0 to 3 every 80ms. Tango produces these scores in the same pass over the same audio it already uses for end-of-turn and interruption detection, with no second pipeline or added latency, and works on compressed 8kHz telephony audio. The company describes uses during calls (clarifying, changing course, matching tone) and after calls (reviewing where conversations turned and spotting patterns across many calls).
Rivian is licensing elements of its AI voice assistant technology to Volkswagen Group as part of their $5.8 billion joint venture, according to Rivian's third-quarter investor presentation, though licensing fees and payment terms were not disclosed. The assistant, which rolled out to Rivian vehicles in September 2026, could debut in Volkswagen's low-cost ID EVERY1, slated to be the first VW vehicle to use Rivian's software and electrical architecture under the joint venture established in 2024. Rivian described the expansion as "a proof point for the extensibility of the platform we've built in-house."
ElevenLabs launched SFX v3, its newest sound effects model, which handles prompts describing multiple sounds more reliably, generates music and vocals more convincingly, and produces cleaner, more natural audio than SFX v2 in benchmark testing. It is available now in ElevenCreative (Flows, the Sound Effects tab, Studio, the Music Editor, and the mobile app) and through ElevenAPI for developers.
Granola, known for its meeting notetaker, launched a dictation feature that works on Mac: users press the "Fn" key and speak, and the tool cleans up and formats the speech and posts the final text across all apps. The company said the feature will land on iOS and Windows soon. The move comes months after Wispr, known for its dictation tool, launched a meeting notetaker.
ROAL is an AI phone host for restaurants that builds itself from a website or Google Maps link in about two minutes, learning the menu and hours and then taking pickup and delivery orders, answering FAQs, and sending kitchen tickets. It charges $0.90 per completed order with no monthly, per-minute, or setup fees, and hands catering, complaints, and manager requests to staff.
Amazon has added more than 100 new shopping-related capabilities to Alexa ahead of the holiday season, including subscription and list management, editing delivery instructions, resolving payment issues, and answering questions about Prime subscriptions. Users can create and edit lists, change subscription delivery frequency (e.g., every three months instead of monthly), and update addresses or drop-off details. Amazon said use of agentic features has grown 8x since the start of the year, and over 400 million consumers have used Alexa for shopping in the last 12 months.
Deepgram announced it is acquiring Daily, the maker and maintainer of the open-source voice AI framework Pipecat, with the deal amount undisclosed and Pipecat to remain open-source. Daily, which has raised over $62 million from investors including Tiger Global and Nat Friedman, was in the same 2016 Y Combinator batch as Deepgram, which reached unicorn status with a $130 million Series C earlier this year. Deepgram is also launching the Real-Time Context Standard (RCS), an open protocol for voice agents co-developed with the Pipecat team.
Granite Switch 4.1.1 is a backwards-compatible patch release to Granite Switch 4.1 that adds an updated Granite Speech ASR model covering multilingual speech-to-text across 10 languages with native punctuation and capitalization, plus an updated Decision adapter in granitelib-core with improved performance on the Jev Decision Index test set. Apart from training the new ASR model and Decision adapter, nothing was recomputed and the base model is untouched. The team says it is working on tutorials and an Adapter SDK so developers and third-party providers can build their own adapters.
The author announces two Swift packages: AnySpeechRecognizer, which mirrors Apple's SpeechAnalyzer API while allowing alternative engines such as NVIDIA's Parakeet and Alibaba's Qwen3-ASR running locally on Core ML, plus speaker diarization via NVIDIA's Streaming Sortformer; and AnyDecisionModel, which answers yes-or-no, multiple-choice, and rating-scale questions with typed results and probabilities from backends including MLX, Core AI, llama.cpp, Jev, and Ollama. Both can download open models from the Hugging Face Hub and run them on device, require Swift 6.2 and macOS 26 or iOS 26 or later, and are published on GitHub.
AssemblyAI released AssemblyAI-TTS-1, its first text-to-speech model, now generally available as a hosted API in US and EU regions, priced at $15 per 1M characters with unlimited concurrency and no contract. AssemblyAI's self-reported tests show 93.7% entity accuracy and 47 ms median time to first byte, while Cekura's independent benchmark measured 159 ms median streaming latency and 81.5% reading accuracy. At launch it supports six languages, 4,000+ voices and instant voice cloning, but lacks SSML, speed control, pronunciation dictionaries and emotion tags.
Block announced on October 23, 2026 a research preview of Dessa, a full-duplex speech-to-speech model trained only on human conversation audio, with no text and no pretrained models. Block says it trained Dessa on one NVIDIA GB200 NVL72 rack and that it runs locally on a MacBook Pro M4 Max or better, faster than real time. No benchmark scores, parameter count, weights, paper, API or price have been published, and access is via a waitlist.
NVIDIA released Nemotron 3.5 Transcribe, an open-weights speech LLM with about 34B total and 4.3B active parameters that handles 40 languages and audio up to 1 hour, with text instructions selecting verbatim, speaker-attributed, contextual or translated output. NVIDIA reports 4.71% mean WER on 9 public English Open ASR sets, the best open-weights score, while commercial APIs still lead, and notes weak Mandarin meeting accuracy (34.95% cpCER on AliMeeting) and offline-only operation. Weights ship in BF16, FP8, NVFP4 and GGUF under the OpenMDW-1.1 license, with runtimes including vLLM 0.32.0, Transformers and SGLang-Omni 0.2.0.
This guide explains how banks use conversational AI—NLP-based chat and voice agents that understand natural language, pull account context, and escalate to humans when needed—for 24/7 support, onboarding, disputes, and fraud alerts. It cites CFPB findings on early bank chatbots, McKinsey figures on reduced handling time and call volume, and ElevenLabs customer examples from Customers Bank and Klarna, while detailing security standards such as SOC 2 Type II, ISO 27001, PCI DSS, and Zero Retention Mode.
ElevenLabs introduced Cue, a forced alignment model that places every character in an audio file on a precise timeline, powering its Scribe speech-to-text family and available through a Forced Alignment API. ElevenLabs says Cue was more accurate than every aligner it tested across all 29 benchmark languages, handles recordings of any length in a single pass, and aligns an hour of audio in under a second. It also aligned songs in English, French, Spanish and German in one pass, placing 92.7% of words within 0.3 seconds of the label.
Baseten optimized Qwen3-ASR 1.7B, an Apache 2.0 open-weight speech recognition model supporting over 30 languages and 22 Chinese dialects with a 5.4% average WER on the Open ASR Leaderboard, for batch transcription on a single RTX PRO 6000. The optimized stack transcribes 7x more audio per second than out-of-the-box vLLM at high concurrency, reaching 2,645x real-time factor on 5-minute clips and 3,128x on hour-long files, with median latency for 5-minute clips at 64 concurrent requests falling from 50 seconds to 7.3 seconds. Baseten reports a cost of about $0.0013 per audio hour, 8x lower than the same model on vanilla vLLM and roughly 277x less than popular closed-weight providers.
Databricks argues that voice agents are now practical because synthetic voices sound natural and can respond within roughly 200 milliseconds, but that most enterprise AI investment has focused inward on employee assistants rather than customer conversations. It describes a prototype in which LiveKit carries calls, Deepgram handles speech-to-text, ElevenLabs adds voices, and Databricks supplies the "brain" that looks up caller context, applies business rules, and records each decision, built on context, ontology, and governance. The post outlines three pillars—grounding answers in business data, fast and slow AI decision-making including the ai_decide capability, and a learning loop from recorded calls—and points to the open-source DIVA blueprint on GitHub and a suggested 90-day pilot.
Nvidia released Nemotron 3.5 Transcribe, an open speech model that follows text instructions to transcribe 40 languages, translate 31 to or from English, identify speakers in 13, and answer questions about recordings up to an hour long with up to eight speakers. The 30-billion-parameter model activates about 4 billion parameters per step, runs on a single Nvidia GPU, and ranks first among open-weight models on Hugging Face's English Open ASR Leaderboard with a 5.33 percent average word error rate. Nvidia also released the Granary v2 dataset of about 844,000 hours of speech in 48 languages plus dataset-building tools, and reported that targeted fine-tuning on recordings of Indian doctors and medical students cut error rates from 30.8 to 11.8 percent in about 14 GPU hours on one GB300.
Nvidia released Nemotron 3.5 Transcribe, a free, open-weight model that transcribes audio in up to 40 languages and translates 31 languages to and from English, alongside version 2 of its open source Granary speech dataset. Nvidia claims it handles recordings up to an hour long, distinguishes up to eight speakers, and beats rivals on benchmarks including Hugging Face's Open ASR Leaderboard; the 30-billion-parameter model runs on a single chip but works with recorded audio, not livestreams. The release targets developers building business meeting and medical transcription apps and joins other recent voice model launches from Mistral, ElevenLabs and Salesforce.
NVIDIA released Nemotron 3.5 Transcribe, an open-weight, promptable model that handles transcription in 40 languages, translation of 31 to or from English, speaker attribution in 13 languages, and audio question answering in one checkpoint, running on a single NVIDIA GPU. It ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English at 5.33% average WER and posts the lowest diarization error rate on David AI's 21-language leaderboard. The release also includes Granary v2, an open data curation pipeline covering 463,062 hours across 48 languages, and NeMo Speech fine-tuning recipes that cut held-out WER by 45.8% relative to the best prompt in a medical case study.
Why it matters: It documents concrete benchmark standings, an open data pipeline, and a reproducible fine-tuning recipe for teams evaluating or adapting open speech models.